Codex Pricing 2026: Plans, Limits, Resets and Credits

What Codex really costs in October 2026: every plan, the five-hour and weekly limits, speed multipliers, credits, and the resets everyone is tracking

On the $20 ChatGPT Plus plan, OpenAI's own pricing page estimates you get 5 to 45 GPT-6 Astra messages, 15 to 160 GPT-6.1 Sol messages or 350 to 3,000 GPT-6 Luna messages every five hours in Codex - OpenAI Codex pricing. On every Pro plan there is no five-hour limit at all. Turn on Fast mode and the same work drains your allowance 2.5 times faster. And since September 2025, one independent tracker has logged 43 verified global resets, moments when OpenAI simply refilled everyone's limits at once - codex-reset.com.

That is why "how much does Codex cost?" is a harder question than it looks. The sticker prices are simple: $0 to $500 a month for individuals ($0, $8, $20, $100, $200 or $500), $20 to $25 per seat for teams, or plain API rates per token. The real price is how much agent work each of those plans buys you, and three dates within three weeks change that answer. On September 29 OpenAI restructured Pro into three tiers and halved the allowance of new $200 subscriptions. On October 4 the Codex lead promised a daily improvement or a "full reset" for 28 days. And on October 14, GPT-5.5 leaves Codex entirely.

This guide covers every Codex plan and what it includes, how the usage limits are actually metered, the four different kinds of reset and how each one changes your weekly schedule, what credits cost per message, how Fast and Ultrafast multiply your spend, and how Codex compares with Claude Code, Cursor, GitHub Copilot and Google's plans. Every number comes from OpenAI's live documentation or a primary source checked on October 7, 2026, and where OpenAI does not publish a number (weekly caps, credit pack prices) we say so instead of guessing.

Contents

1. Codex Pricing in October 2026: The Plan Ladder
2. How Codex Usage Limits Work: Five-Hour Windows, Weekly Caps and One Shared Pool
3. Codex Resets Explained: Weekly, Global, Banked and Paid
4. The 28-Day Run: What OpenAI Is Shipping Every Day in October
5. Fast, Ultrafast and Reasoning Effort: The Hidden Multipliers
6. Credits: The Overflow Currency and What a Message Really Costs
7. ChatGPT Pro 100, 200 and 500: What the September 29 Change Means
8. Business, Enterprise and Edu: Seats, Pooled Credits and Spend Controls
9. Using Codex With an API Key: When Pay-per-Token Wins
10. Choosing a Model on a Budget, and the GPT-5.5 Retirement
11. Codex vs Claude Code, Cursor, Copilot and Gemini: Pricing Compared
12. Why Codex Is Priced This Way: A First-Principles View
13. Making Your Codex Limits Last: A Practical Playbook
14. What Comes Next: The Dates That Will Change Your Bill

Codex Plans Scored: Which Tier Buys the Most Work

Before the detail, here is the whole ladder on one table. We scored all nine ways to pay for Codex on four criteria that decide whether a plan is worth its price for real agent work: what the included work costs per dollar, how far you get before a hard stop, which models and speed modes you can reach, and how much control you have over overflow spending. Each cell carries the score and the fact behind it, so you can disagree with a score and still use the data.

The ranking measures plan quality, not which plan you should buy. A solo developer who codes a few evenings a week should not buy the top-ranked plan, and the table makes that visible: the cheapest plans lose points on headroom, not on value. Use the final score to see trade-offs at a glance, then use the sections below to match a plan to your own usage.

How to read the criteria. Cost of work asks what you pay per unit of included agent work, using OpenAI's own plan multipliers and published rates. Headroom asks how much work you can do before something stops you: the five-hour window, the weekly cap, or a fixed message range. Models and speed covers access to GPT-6 Astra, GPT-6.1 Sol, Fast mode, Ultrafast and Codex cloud. Control covers the overflow options (credits, resets, pay-as-you-go), spending caps and admin tooling. Pro 100 and Pro 200 tie at 7.3 and are listed alphabetically.

1. Codex Pricing in October 2026: The Plan Ladder

The first thing to understand about Codex pricing is that Codex is not sold on its own. Since OpenAI folded its coding agent into the ChatGPT subscription, Codex is included in every ChatGPT plan, from Free to Enterprise, and the plan you pay for decides how much of it you get. The pricing page now says it plainly: ChatGPT Work and Codex share usage, with the same pricing, credits and limits - OpenAI Codex pricing. So when you buy "Codex," you are really buying a share of an agentic usage pool that also feeds ChatGPT Work, ChatGPT for Excel and, through Sign in with ChatGPT, a growing list of partner apps.

The second thing is that the dollar amount on the plan card tells you very little. Two Plus subscribers paying the same $20 can get wildly different amounts of work out of it, because Codex meters tokens, not prompts, and a long agentic task with a large repository in context can consume more than a hundred short questions. That is why OpenAI publishes message estimates as ranges (5 to 45 for Astra on Plus) rather than fixed quotas, and why the rest of this guide focuses on the mechanics behind those ranges.

Here is the full ladder as OpenAI lists it on its Codex pricing page and help center on October 7, 2026:

A few lines on that table deserve emphasis. The Business price is $20 per user per month billed annually for two or more users, or $25 billed monthly - OpenAI Codex pricing. Pro is billed monthly only, with no annual option, and comes in exactly three tiers: Pro 100, Pro 200 and Pro 500, where only Pro 500 includes Astra Ultrafast - OpenAI Help Center. And Codex Cloud (tasks that run on OpenAI-managed machines while your laptop sleeps) is available to Plus, every Pro tier, Business, Enterprise, Healthcare and Education, but not to Free or Go - OpenAI Help Center.

What changed in the last six months

Codex pricing in 2026 has been a moving target, and most of the movement happened in a tight window around OpenAI's DevDay on September 29. The $100 Pro tier arrived on April 9 with "5x more Codex usage than Plus," in OpenAI's words as shared by its Codex lead - Thibault Sottiaux on X. GPT-6 Sol and GPT-6 Luna launched on September 22 alongside a 50% API price cut, and DevDay brought GPT-6.1 Sol, Pro 500 with Ultrafast, and a new allowance for Pro 200. Less than a week later the 28-day run began.

The sequence matters because each step changed what a dollar buys, and several of the changes come with deadlines that have not passed yet. If you subscribed to Pro 200 before September 29, for example, your old allowance runs out on October 29. If any of your scripts or saved settings pin GPT-5.5, they break on October 14. The diagram below lays out the timeline, including the dates still ahead.

OpenAI's own keynote is the best primary record of the September changes. The full DevDay recording covers GPT-6.1 Sol, Astra Ultrafast and the new Pro tier, and OpenAI's written recap counts more than 20 major announcements and a "collective 1.2B weekly users" across ChatGPT - OpenAI DevDay 2026 recap.

Why this matters: the plan you picked a month ago may no longer be the right one, because the ratios between tiers changed. How to apply it: start with the scored table above to see which tier matches your workload, check the deadlines on the timeline that apply to you, and then read section 2 to understand what the limits on your current plan actually measure. If you are comparing Codex against Anthropic's agent, our breakdown of Claude Code pricing uses the same structure, so the two guides line up side by side.

2. How Codex Usage Limits Work: Five-Hour Windows, Weekly Caps and One Shared Pool

Codex limits are easiest to understand as two buckets stacked on top of each other. The first is a five-hour window: on Plus and Standard Business seats, it caps how much you can do in any short burst of work. The second is a weekly limit, which OpenAI describes only as "weekly limits may also apply," and which caps your total over seven days - OpenAI Codex pricing. On Pro, the five-hour bucket does not exist: "Pro plans currently have no five-hour limit," and the Codex lead has said OpenAI is "committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want" - Thibault Sottiaux on X.

Neither bucket counts messages. Both count the compute your work consumes, and OpenAI lists what drives it: "Model choice, context, reasoning, tool use, retrieval, and caching all affect usage, so prompt length alone isn't a reliable estimate" - OpenAI Codex pricing. A one-line fix in a small script and a two-hour refactor across a monorepo are both "one message," yet one costs a fraction of a percent of your window and the other can take a large slice of it. That is why the published numbers are ranges, and why a range like 15 to 160 for GPT-6.1 Sol is honest rather than evasive.

The chart makes one pattern obvious: the model you choose moves your allowance far more than anything else you control. Astra's best case (45 messages) is less than a third of GPT-6.1 Sol's best case (160), and GPT-6 Luna's range of 350 to 3,000 would dwarf both if we drew it. Since the five-hour window is the binding constraint on Plus, picking the model is effectively picking your daily capacity.

Everything that draws from the same pool

The shared pool is the part of Codex pricing that surprises people most, because usage leaks in from places that do not look like "coding." OpenAI's help center lists the members: Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents "use a shared allowance and credit pool when those features are available on your plan" - OpenAI Help Center. Voice tasks started in the desktop app, image generation (which uses limits 3 to 5 times faster than a similar turn without images) and partner apps all draw from it too.

Partner apps are the newest and least visible drain. Through Sign in with ChatGPT, Plus and Pro subscribers can let third-party tools such as Devin, Notion, Vercel, Warp, Amp Code and Kilo Code spend their plan, and OpenAI's DevDay recap counts 16 partners - OpenAI DevDay 2026 recap. The consent screen looks like this:

The wording in that dialog is precise: the app may "consume usage from your ChatGPT Plan's limits." It does not get a separate allowance. OpenAI's documentation adds that "connecting an app does not add a new allowance," and that you can cap each app at a percentage of your weekly plan usage under Settings, Usage, App limits, where "an app's limit is a cap, not a separate allowance or a reserved portion of your plan" - OpenAI Sign in with ChatGPT docs. Sharing purchased credits with partners is a separate switch that is off by default - OpenAI Help Center.

When a request comes in, the path it takes through all of this looks like the diagram below. The key branch is the first one: whether you are signed in with a ChatGPT plan or an API key decides whether limits apply at all.

What happens at the limit, and where to check it

OpenAI has made one generous design choice at the edge of the limit: "If you reach your usage limits during an active turn, the agent will be able to continue working on that turn, subject to fair use limits" - OpenAI Codex pricing. A long refactor that crosses the line is not cut off halfway. The next message is what gets blocked. On Pro, the help center adds that "model allowances vary by tier," so a single model can become unavailable at its own limit while others still work - OpenAI Help Center.

You can watch all of this in two places. The usage dashboard at chatgpt.com/codex/settings/usage shows each allowance, your credit balance and reset times, and the /status command shows remaining limits inside an active Codex CLI session. OpenAI's own advice is to "check the dashboard every week or two to understand your pace." Most heavy users check far more often than that, which is part of why an entire ecosystem of reset trackers now exists (section 3).

Why this matters: on Plus, your effective price per task depends almost entirely on model choice and context size, not on the $20. How to apply it: before a long session, check /status, pick the cheapest model that can do the job, and cap partner apps you have connected so they cannot quietly eat your weekly allowance. For a deeper look at why context size dominates agent costs, see our guide on how to cut LLM costs.

3. Codex Resets Explained: Weekly, Global, Banked and Paid

"Reset" means four different things in Codex. OpenAI's help center draws the distinctions carefully, and confusing them is an easy way to lose usage you thought you had. A weekly reset is the automatic refill at the end of your seven-day window. A global reset is OpenAI refilling everyone's limits at once, usually announced on X. A banked reset is a saved, one-time refill you apply yourself. A paid instant reset is a refill you buy.

The reason the distinctions matter is that banked and paid resets officially move your weekly reset date, and community trackers report that your new week also starts at your next request after a global reset. That sounds like a technicality until you apply a banked reset on a Tuesday and discover your usual Friday refill no longer comes. Here is how each one works, according to OpenAI's own documentation.

Banked resets: saved refills with an expiry date

A banked reset is "a one-time Codex usage-limit reset saved to your account until you use it or it expires." Using a full one "refreshes your 5-hour and weekly Codex usage windows and changes your weekly reset date," and OpenAI gives a concrete example: if your week was due to reset on Friday and you apply a banked reset and resume on Tuesday, "your next weekly reset will be around the following Tuesday. You won't also receive a weekly reset on the original Friday" - OpenAI Help Center.

Two details in that article are worth knowing before you hoard banked resets. First, a reset "is consumed only when it successfully refreshes at least one eligible usage window," so applying one with nothing to reset does not waste it. Second, expiry is final: "Once an unused reset expires, it cannot be restored or reissued," and Support "does not provide manual or courtesy resets." Banked resets come from referral promotions and launch events. OpenAI loaded one into every Plus, Pro and Business account when GPT-6 Sol and Luna launched on September 22, according to the codex-reset.com timeline.

Paid instant resets: buying next week early

The paid reset is the newest and least understood option. Plus and Pro personal accounts can "buy an instant reset from Usage settings in ChatGPT Desktop before reaching a limit, or from an in-app offer after reaching the weekly limit." The purchase "immediately restores both 5-hour and weekly usage," but the key phrase is the next one: it "pulls your normal weekly allowance forward rather than adding a separate usage entitlement" - OpenAI Help Center.

In practice that means you are not buying extra usage, you are buying time. Your new weekly period starts with your first Codex or Work request after the reset, and your next automatic reset is seven days after that. The price appears only at checkout, availability "may vary by account" and billing country, it is not offered on Free, Go, Business, Enterprise or Edu, and "generally, we do not offer refunds." The banner that offers it only appears after you exhaust the weekly limit, not the five-hour one.

These mechanics make a few rules worth following:

• Never buy early in the week. A paid reset restarts your clock, so buying one with plenty of allowance left wastes what remains.
• Compare it with credits first. Credits add usage without moving your reset date (section 6).
• Apply banked resets late. Use them when the weekly bucket is nearly empty and the automatic refill is still days away.
• Check /status afterwards. OpenAI notes the dashboard "may take a short time to update."

The common thread is that every reset other than the automatic one trades your schedule for capacity. That trade is good when you are blocked mid-project with days to go, and bad when you are simply impatient. Because OpenAI also hands out global resets frequently, an instant purchase made the evening before a surprise global refill is money spent on nothing, which is exactly why so many people now watch for global resets before paying. The practical order of operations is therefore: spend your included allowance, then any banked reset, then credits for small overruns, and only then a paid reset when you are blocked for days.

Global resets and the trackers that watch for them

Global resets are where Codex pricing turns into a spectator sport. The independent codex-reset.com timeline counts 43 verified global resets and 11 banked-reset drops since September 2025, each linked to the announcement. They happen for very different reasons: compensation after outages and slowdowns, a rewritten billing system ("We rewrote the underlying system to track and bill usage in Codex and we have reset usage limits in the process"), model launches, and user milestones. When Codex passed 25 million active users on August 31, the Codex lead reset every paid subscription and signed off with "See you soon for more news from The Reset Company" - Thibault Sottiaux on X.

The milestones also tell you why capacity keeps shifting. Every jump in demand puts pressure on the limits OpenAI can afford to offer, and the jumps have been steep.

Going from 3 million weekly users on April 7 ("up from two million a little under a month ago") to 25 million by the end of August is roughly an eightfold increase in five months - Thibault Sottiaux on X. Each spike in demand forces OpenAI to choose between slower responses, tighter limits or more compute, and global resets are the release valve it uses when new capacity comes online or when a bug drained people's limits unfairly.

That unpredictability has created a small industry. At least six websites now exist purely to answer "did Codex reset today?", and they work by watching OpenAI's announcements rather than your account. Each one counts differently, from 43 to 61 resets depending on what it treats as "verified," so the differences between them are worth knowing before you trust one:

• willreset.com gives odds of a reset in the next 24 and 48 hours and a typical gap of about three days.
• codexresets.com answers "Did Codex Reset Today?" and sends email alerts.
• resetbeacon.com checks six sources every five minutes for the next shared reset.
• codexreset.org combines X signals into a 24 and 48 hour estimate.
• aiidelist.com keeps a reset history with average and median intervals.

None of these sites can see your account, so none of them can tell you when your own weekly window refills; they only predict the shared, announced resets. For your personal numbers, a second category of tool has grown on GitHub: menu-bar apps and command-line meters that read the same data /status shows and alert you as you approach a limit. The healthy way to use both is to treat a predicted global reset as a pleasant surprise, never as capacity you have budgeted for, because when a user petitioned to keep the natural weekly reset on schedule, the Codex lead's reply (archived by codexreset.org) was "There is no schedule, only resets."

When the automatic reset does not happen

Weekly resets are anchored per account, not to a universal day. The trackers note that after a reset, "your next seven-day window starts when you first use Codex again," which matches OpenAI's description of paid and banked resets but is not stated as a rule for every case. Occasionally the anchor misbehaves. On October 6, a developer reported on OpenAI's forum that a weekly reset "scheduled to reset on October 5, 2026" never happened, usage kept falling to 29%, and the next reset moved to "October 10, 2026 at 7:21 AM JST" - OpenAI Developer Community.

If that happens to you, the help center's instruction is to contact Support with "the limit or credit balance you're questioning, the reset time shown, a screenshot, the Codex client and model, and when it happened, including your time zone." Support cannot grant courtesy resets, but it can investigate a reset that was "counted incorrectly or access failed to return after a reset" - OpenAI Help Center.

Why this matters: resets are now a real part of the price, worth anywhere from a few hours to a full week of usage each time. How to apply it: never plan a deadline around a global reset, always use your automatic and banked resets before paying for one, and take a screenshot of /status before and after any reset so you have evidence if the numbers look wrong.

4. The 28-Day Run: What OpenAI Is Shipping Every Day in October

On Sunday, October 4, 2026, OpenAI's Codex lead Thibault Sottiaux posted a promise that turned Codex pricing into a daily event: "Over the next 28 days, each day we'll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset. Let the improvements begin" - Thibault Sottiaux on X. The post quoted his message from earlier that day: "Only things being worked on are simplifications, more efficiency for more usage, groundbreaking features or new models." It passed 7.5 million views within three days.

Day 1 was October 5, so the run ends around October 31 or November 1. The structure is unusual for a pricing change because the fallback option is itself a pricing event. On a day without a qualifying improvement, everyone's limits refill. In other words, OpenAI has committed to delivering value every day for four weeks, and it has named a full reset as the unit of value it will pay out when it cannot ship something better. The community tracking thread on OpenAI's forum keeps a running list of each day's result - OpenAI Developer Community.

Day 1: a free speed upgrade, with a catch worth understanding

The Day 1 announcement read: "We have optimized the default speed to be ~50% faster across GPT-6 Astra and GPT-6.1 Sol through the subscription across all our products and partners using Sign in With ChatGPT (including OpenCode, Pi, Amp, Devin, ...). No changes needed on your end" - Thibault Sottiaux on X. The announcement covered subscription usage and said nothing about API speed.

The catch is subtle. Speed and allowance are different things. In principle, a turn that runs 50% faster still processes the same tokens, so it consumes the same share of your limit. If your bottleneck was the wall clock (you wait on the agent all afternoon), Day 1 gave you more done per hour. If your bottleneck was the five-hour or weekly allowance, it changed nothing. What it did change is the case for Fast mode: before October 5, the only way to get a quicker Astra was to pay 2.5 times the allowance per turn; now part of that speed comes free by default.

Run-to-run variance is the other reason to read speed claims carefully. One forum member who benchmarks models on their own coding tasks posted two runs of GPT-6.1 Sol at High effort on the same task: about 34 minutes and 45.2 million tokens (roughly $5.33 at API rates) during US working hours on October 5, and about 10 minutes and 6.4 million tokens (roughly $1.11) the next morning "whilst US sleeps" - OpenAI Developer Community. That is one user's informal test, not a controlled benchmark, but it shows how much the same task can cost from one run to the next, and why OpenAI quotes allowances as ranges rather than fixed counts.

Day 2: auto-review became free, and then a reset anyway

Day 2's headline was that Auto-review (OpenAI's "Approve for me" mode) no longer counts against usage: "We have made Auto-review free for all users signed in through a ChatGPT account... It allows you to run long tasks while having a second agent review all actions taken by the primary agent" - Thibault Sottiaux on X. In his roundup he put a number on it: auto-review "can be between 2-10% of plan when used." The help center now confirms that "auto-review safety checks are free and don't count toward your plan's usage limits" - OpenAI Help Center.

This OpenAI screenshot of the Codex CLI shows the mode in context, with "approve for me" in the session header beneath the model and reasoning setting:

Auto-review works by routing approval requests that would normally pause for you to a separate reviewer agent. According to OpenAI's documentation, it evaluates requests that cross the sandbox boundary, such as:

• Escalated shell commands that ask for permissions beyond the sandbox
• Blocked network requests under the current policy
• File edits outside the allowed writable folders
• Tool calls from MCP servers or apps that require approval
• Computer Use visits to a new website or domain

The reviewer is designed to block data exfiltration, credential probing, broad security weakening and destructive actions, and its policy is published in the open-source Codex repository - OpenAI auto-review docs. OpenAI is also candid that "it is not a deterministic security guarantee" and should complement good sandbox design. The pricing consequence is straightforward: if you already used auto-review, you just got 2 to 10% of your plan back, and if you did not, the main cost argument against it is gone. You can enable it under Settings, Permissions, Auto-review, or in config.toml:

# Route sandbox-boundary approvals to the reviewer agent instead of pausing for you
approval_policy = "on-request"
approvals_reviewer = "auto_review"

Then, a few hours later, Day 2 ended in a reset as well. After a public vote, Sottiaux posted: "We shipped four things that were deemed good to great and some math proofs, but the vote is clear and the community demands a reset. I did calibrate it and it seems that the game is rigged in reset's favor, but such are the rules at the moment. Therefore ... the reset has been processed" - Thibault Sottiaux on X. A follow-up promised "we won't unship the improvements."

What the 28 days tell you about Codex pricing

The timing is not subtle. The pledge arrived five days after OpenAI halved the allowance for new Pro 200 subscriptions, a change that drew hundreds of comments on Hacker News (section 7). It also arrived two days after a global reset that OpenAI issued because GPT-6.1 Sol had a "massive load spike in the first two days" - Thibault Sottiaux on X. Read together, the run looks like a deliberate effort to rebuild goodwill while capacity catches up with demand.

Why this matters: for the rest of October, the effective price of Codex is lower than the plan pages suggest, because each day adds either a feature or a refill. How to apply it: turn on auto-review now that it is free, do not buy a paid instant reset during the run (a free one may arrive the next day), and do not budget November on October's generosity. When the run ends, the limits revert to whatever the plan pages say, which is the number to plan around.

5. Fast, Ultrafast and Reasoning Effort: The Hidden Multipliers

The biggest swings in what Codex costs you do not come from the plan. They come from three settings you can change mid-session: speed mode, reasoning effort and model. OpenAI publishes the speed multipliers exactly, which makes them the easiest of the three to reason about. Fast mode "uses included subscription limits at 2.5x the Standard rate," while purchased credits and Enterprise pay-as-you-go usage are billed at 2x. GPT-6 Astra Ultrafast uses included limits at 8x and credits at 6x - OpenAI Speed docs.

OpenAI is careful to add that "these billing multipliers don't describe speed increases." The speed side looks like this: Fast mode speeds up GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna where available (for GPT-5.6 and GPT-5.5, the increase is 1.5x), while Ultrafast "generates tokens up to 8x faster than GPT-6 Astra in Standard mode." That comparison measures token generation, not the time to finish a whole task, which also includes tool calls, test runs and file reads that do not speed up.

The API column comes from OpenAI's API pricing page, which notes that "Priority processing was renamed Fast mode on July 30, 2026" - OpenAI API pricing. Ultrafast has more restrictions than the others: other self-serve plans "don't have access to Ultrafast at launch, even with purchased credits," it is off by default in Enterprise workspaces, and it is not available to workspaces that require inference residency outside the United States.

OpenAI's own short launch video for Ultrafast shows what the top speed tier is for. Its description puts a number on the gap between the modes: in Codex, Ultrafast "runs up to 8x faster than Astra Standard and 4x faster than Astra Fast."

The arithmetic of buying speed

Three quirks in those multipliers are worth noticing. The first is that the included-usage multiplier is steeper than the credit multiplier (2.5x against 2x for Fast, 8x against 6x for Ultrafast). If you are going to run past your allowance anyway, the credit side of Fast is the cheaper side, which nudges heavy users toward buying credits for speed-sensitive work and spending the plan on Standard turns.

The second comes from OpenAI's own speed claims. If Ultrafast is up to 8x faster than Standard and 4x faster than Fast, then Fast is roughly 2x faster than Standard for Astra, while it costs 2.5x the allowance. Ultrafast, by contrast, charges 8x for up to 8x the speed. On a pure speed-per-allowance basis, Fast is the slightly worse deal of the two, which is worth knowing before you leave it on all day.

The third is what Ultrafast does to the top plan. Pro 500 includes 25 times the usage of Plus (section 7). If you ran Astra on Ultrafast for everything, that 25x divided by an 8x multiplier leaves roughly 3x Plus worth of Standard Astra work, delivered much faster. Ultrafast is therefore best understood as a way to buy back your own time on the few tasks where waiting is expensive, not as a default. The same logic applies to Fast on Plus: a five-hour window that holds 100 Standard GPT-6.1 Sol turns holds about 40 in Fast.

You can toggle Fast per session with /fast in the CLI, show it in the footer with /statusline, or make it the default in config.toml:

# Default to GPT-6.1 Sol with Fast mode (uses included limits at 2.5x)
model = "gpt-6.1-sol"
service_tier = "fast"

[features]
fast_mode = true

Reasoning effort is a multiplier too

Reasoning effort does not have a published multiplier, but it moves cost just as surely. Codex offers Light (called low in the CLI), Medium, High and Extra High, plus Max, which "gives the selected model more time to reason about a single task," and Ultra, which "uses subagents to handle separate parts of a complex task in parallel" - OpenAI Codex models docs. OpenAI's own guidance is that "higher reasoning effort can improve results for complex tasks, but it takes longer and uses more tokens," and that "most tasks do not need Max or Ultra."

The recommended starting points are High for Luna and Light for Astra, with GPT-6.1 Sol starting at your client's default. That asymmetry is a useful hint: a capable model at low effort often beats a small model at high effort, and it usually costs less. If you want to understand why parallel subagents multiply spend, our guide to Claude Code subagents walks through the same mechanics on Anthropic's side.

Why this matters: a Plus user who leaves Fast on and reasoning at Extra High can burn through a week of allowance in a day without changing a single prompt. How to apply it: keep Standard speed and default effort as your baseline, switch to Fast only for interactive work where you are actively waiting, and reserve Max, Ultra and Ultrafast for tasks where an hour of your time is worth more than a large slice of the week's allowance.

6. Credits: The Overflow Currency and What a Message Really Costs

Credits are what you spend after your included usage runs out. On Plus and Pro you can buy them from Settings, Usage in ChatGPT or Usage and Billing in the Codex app, and they work across Codex, ChatGPT Work, Word, Excel and PowerPoint. Included usage is always spent first. Credits are "valid for 12 months from purchase," are non-refundable except where law or an approved exception applies, cannot be transferred or resold, and can even go negative if a task that started with a positive balance finishes after concurrent usage drained it - OpenAI Help Center.

Automatic reload is available on some accounts: when your balance falls below a minimum you choose, OpenAI buys enough to return it to your target balance, "up to any maximum monthly spend you set." Credits are also separate from your ChatGPT wallet, the dollar balance that holds gift-card funds, so redeeming a gift card does not add usage credits. A limited group of Free and Go users can buy credits too, but buying them "does not change your Free or Go plan or increase your included usage limit."

What one credit is worth

OpenAI does not publish a single price for personal credit packs on the pages we checked; it says purchase amounts and prices "can vary by account, region, and plan." It does, however, publish two rate cards for business customers, one in credits and one in US dollars, and they line up exactly. GPT-6 Astra costs 250 credits per million input tokens on the credit card and $10 per million on the Enterprise USD card, and every other model follows the same ratio of 25 credits per dollar, which implies roughly $0.04 per credit - OpenAI credit rate card. Treat that as a derived figure, not a quoted price, and check the price shown at checkout for your own account.

The credit rates come from OpenAI's Codex pricing page, which adds that "Codex credit billing has no separate cache-write charge"; the dollar column comes from the Enterprise USD rate card - OpenAI Enterprise USD rate card. Note the cached input column. GPT-6.1 Sol charges just 2.5 credits per million cached tokens, half of GPT-6 Sol's rate, and agentic coding rereads the same repository context constantly, so cache rates matter more than they look.

A worked example: one typical agent turn

To turn those rates into something tangible, take a representative agentic turn: the agent reads 150,000 tokens of context (90% of it cached from earlier in the session) and writes 6,000 tokens of reasoning and code. This small script applies OpenAI's published credit rates to that turn, and you can change the inputs to match your own sessions:

# Credits per 1M tokens: (input, cached input, output), from OpenAI's Codex rate card
RATES = {
    "gpt-6-astra": (250, 25, 1250),
    "gpt-6.1-sol": (50, 2.5, 250),
    "gpt-6-luna": (2.5, 0.25, 12.5),
}

def turn_credits(model, input_tokens, cached_share, output_tokens):
    inp, cached, out = RATES [model]
    fresh = input_tokens * (1 - cached_share)
    reused = input_tokens * cached_share
    return (fresh * inp + reused * cached + output_tokens * out) / 1_000_000

for m in RATES:
    c = turn_credits(m, 150_000, 0.9, 6_000)
    print(f"{m}: {c:.2f} credits, about ${c * 0.04:.3f}")

Extending the same calculation to all six models on the rate card gives the spread below. The same turn costs about 14.6 credits (roughly $0.58) on GPT-6 Astra, 2.6 credits (about $0.10) on GPT-6.1 Sol, and 0.15 credits (well under a cent) on GPT-6 Luna.

The example passes a sanity check against OpenAI's own figure: the pricing page says "a typical GPT-5.6 Sol task may use 2-15 credits," and our turn lands at 5.9. It also shows how lopsided the model choice is. One Astra turn costs about as much as five or six GPT-6.1 Sol turns and nearly 100 Luna turns, which is the credit-side version of the message ranges in section 2.

Credits versus your included allowance

It is tempting to convert a plan into credits and compare. OpenAI explicitly warns against it: "API token prices are separate from subscription usage; don't use them to estimate included tasks," and "credit prices alone don't determine included subscription usage." The allowance is calibrated separately and changes with promotions, resets and capacity. Still, a rough comparison is instructive. At about $0.04 per credit, $20 buys roughly 500 credits, which covers around 190 of our example GPT-6.1 Sol turns. A Plus subscriber can get up to 160 GPT-6.1 Sol messages in a single five-hour window, and several windows per week, so for anyone who uses Codex steadily, the subscription almost certainly delivers more than its price in credit terms. The unknown is the weekly cap, which OpenAI does not publish.

Two promotions and side costs round out the picture. GPT-5.6 Sol is on promotional credit pricing "at least through Nov 21, 2026," but the promotion "applies to eligible usage paid for with purchased credits. Included plan usage, 5-hour and weekly limits are unchanged" - OpenAI credit rate card. And a few activities carry their own rates: desktop voice costs $0.05 per minute of your Codex budget (1.25 credits per minute on credit plans), image generation uses limits 3 to 5 times faster, and standard Codex Cloud environments have "no separate VM charge" at launch.

Why this matters: credits are the only way to keep working past a limit without moving your reset date, and their value depends heavily on the model you spend them on. How to apply it: buy small amounts (they expire after 12 months), set a maximum monthly spend if you turn on automatic reload, and spend credits on GPT-6.1 Sol or Luna rather than Astra unless the task genuinely needs Astra. For a model-by-model view of what agent tasks cost at API rates, see our analysis of GPT-6.1 Sol versus Claude Opus 5.5 cost per task.

7. ChatGPT Pro 100, 200 and 500: What the September 29 Change Means

For heavy Codex users, one of the most consequential pricing changes of 2026 happened at DevDay. OpenAI's help center describes it in two sentences: "ChatGPT Pro now offers Pro 500, a new $500/month plan that includes Astra Ultrafast. Pro 200 is also available for new subscriptions again. New subscriptions that aren't eligible for grandfathering include a lower usage allowance" - OpenAI Help Center. The help article does not say how much lower. The Codex lead did, on the morning of DevDay: "In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan" - Thibault Sottiaux on X.

Later that day he restated it as multiples: "we are changing the relative difference between plans to be Plus = 1X, Pro 100 = 5X, Pro 200 = 10X," adding that existing subscribers "will keep the 20X multiplier for a bit and also receive a lot of additional credits" - Thibault Sottiaux on X. The email existing subscribers received, quoted on Hacker News, filled in the rest: "Starting October 30, 2026, your included usage in ChatGPT Work and Codex will decrease from 20 times to 10 times the ChatGPT Plus allowance," and "the new Pro 500 tier offers our highest usage allowance, with 25 times the usage of Plus" - Hacker News.

The bulk discount is gone

Look at the fourth column. Before September 29, Pro 200 was the only plan that sold Codex usage in bulk at a discount: $10 for each Plus-sized allowance, half the price of buying the same capacity through Plus. After October 29, every tier sells usage at exactly the same rate.

That flat line is the most important fact about Pro in 2026. Moving up a tier no longer makes each unit of agent work cheaper. What it buys instead is no five-hour limit, Pro features in ChatGPT itself, and on Pro 500, Astra Ultrafast. The Codex lead's stated reasoning is that it does not want "to artificially inflate the API list prices to make it look like you are getting a lot," preferring to pass efficiency gains through API price cuts, and it promises that "over time you always get more work done and with an increasing level of quality." The September 22 launch of GPT-6 Sol and Luna "at 50% of their previous price" is the example it points to.

Grandfathering, cancellations and edge cases

If you held an active Pro 200 subscription "at any point from September 22, 2026 through 10 a.m. Pacific Time on September 29, 2026," you keep the previous allowance "through October 29, 2026" while the subscription stays active. The price stays $200. The same rule applies if you cancel and resubscribe on the same account. Keeping the old allowance "does not upgrade your plan or add Ultrafast," and at launch, buying credits on Pro 100 or Pro 200 does not unlock Ultrafast either. Pro is billed monthly only, and you change tiers under Settings, My Plan.

The change was not popular. The Hacker News thread on Pro 500 drew 219 points and 269 comments - Hacker News. A separate post titled "OpenAI halved the allowance of the $200 plan to 10x" carried the subscriber email. One commenter summed up the skeptic's view in a sentence: "Wait... I thought Inference was profitable?" - Hacker News. The 28-day run that started five days later (section 4) is hard to separate from that reaction.

How to choose a tier now comes down to four questions:

• Do you hit the five-hour limit on Plus? If yes, any Pro tier removes it.
• How many Plus-sized weeks do you burn? Size the tier to that number (5, 10 or 25).
• Do you wait on Astra all day? Only then does Pro 500's Ultrafast earn its premium.
• Were you grandfathered? Re-measure your usage before October 29 decides it for you.

Because pricing per unit is flat, the cost of choosing too big a tier is the unused allowance, and the cost of choosing too small is time lost to limits. Since Pro is billed month to month, the practical approach is to start one tier lower than you think you need, watch /status for two weeks, and move up only if you hit the weekly limit. Remember that what you can buy on top (credits, paid instant resets) is the same on every Pro tier, so the tier only needs to cover your typical week, not your worst one.

Why this matters: for heavy users, Pro 200 is effectively half the deal it was in August, and no tier offers a volume discount any more. How to apply it: if you are grandfathered, plan your October around the old allowance and decide on November before October 29; if you are new, pick the tier by weekly usage and treat Pro 500 as a speed product rather than a savings product.

8. Business, Enterprise and Edu: Seats, Pooled Credits and Spend Controls

Team pricing for Codex works differently from personal plans in one important way: the overflow is pooled. A ChatGPT Business workspace costs $20 per user per month billed annually (two or more users) or $25 monthly, and each Standard seat gets the same five-hour estimates as Plus. When a member exhausts their included usage, "eligible activity may draw from the workspace credit pool when flexible pricing is enabled and spend controls permit it" - OpenAI Help Center. Business also adds larger virtual machines for cloud tasks, SAML SSO and MFA, and no training on business data by default.

The controls are where admins need to pay attention, because the defaults are permissive. Only workspace owners can buy credits, which are valid for 12 months. Owners can turn on automatic reload with a minimum balance, a target balance and a monthly recharge limit, and "leaving this field blank allows unlimited automatic reload purchases each month." Owners and admins can set monthly credit limits per seat type (Standard and Premium) with per-user overrides, but "by default, all seats and users have no limits specified."

Enterprise and Edu: no fixed limits, but real bills

Enterprise and Edu pricing is negotiated with sales, and the headline difference is that workspaces "with flexible pricing have no fixed rate limits. Usage scales with credits," while those without flexible pricing "have the same per-seat usage limits as Plus for most features" - OpenAI Codex pricing. Some Enterprise agreements bill in US dollars per token instead of credits, using the Enterprise USD rate card. On top of Business features, Enterprise adds priority request processing, SCIM, EKM, role-based access control, the Compliance API with audit logs, an analytics dashboard and API, data residency controls and Codex Security for connected repositories.

Several powerful options are off by default in Enterprise, which is good for budgets and worth knowing when users ask why a feature is missing. GPT-6.1 Sol stays off "until an administrator enables it," Astra Ultrafast is off until workspace owners enable it for selected users, and cloud access is off for workspaces that have not turned it on. For Ultrafast in particular, OpenAI warns admins to review per-user spend limits first "because the higher usage rates can consume a user's budget faster" - OpenAI usage limits docs.

How much will a team actually spend?

OpenAI published one rare, concrete data point on this. In an analysis of first-time enterprise adopters of ChatGPT Work in Marketing, annualized consumed-credit value per employee came out at a median of about $53 per year, an average of about $309, and a 90th percentile of about $1,154 - OpenAI ChatGPT Work usage and cost. For 100 employees, the average works out to roughly $30,900 a year, or $2,575 a month as a planning figure.

The shape of that distribution is the whole story of agent budgeting. The 90th-percentile user consumes about 22 times the median user, and the average sits nearly six times above the median because a minority of heavy users pulls it up. OpenAI is careful to call these "directional estimates" from "one week of adoption-period usage," and engineering teams running Codex will skew far heavier than marketers drafting documents. But the lesson transfers: per-seat averages hide the fact that a handful of people drive most of the bill.

Workspace admins can see that distribution in Codex's own analytics. OpenAI's documentation shows a "Use cases and tasks" view that breaks token usage and estimated cost down by category:

Three warnings from OpenAI's cost documentation are worth repeating to anyone running a rollout. User usage limits and workspace overage limits "are separate controls." Alerts "provide notice but don't stop spending." And where overage is permitted, "No limit isn't a zero-spend cap." The safe setup is the opposite of the default: set a seat-type credit limit, set a monthly recharge limit, and raise limits deliberately for the people who need them.

Why this matters: on team plans the per-seat price is predictable, but the overflow is not, and the default configuration lets it grow without a ceiling. How to apply it: before inviting a team, set monthly credit caps per seat type, cap automatic reload, run a two-week pilot, and size the budget from the pilot's median and 90th percentile rather than the average. If you are weighing ChatGPT Work against Anthropic's equivalent for the same team, our ChatGPT Work vs Claude Cowork comparison covers the non-coding side.

9. Using Codex With an API Key: When Pay-per-Token Wins

Every Codex user has a second way to pay that skips plans entirely: sign in with an OpenAI API key. The Codex pricing page describes it as "great for automation in shared environments like CI," with Codex available in the CLI, SDK and IDE extension, "model availability follows the API models available to your key," and usage billed at API rates - OpenAI Codex pricing. There are no five-hour windows and no weekly caps, only your API organization's rate limits and whatever budget you set.

The trade-off is that you lose everything built on ChatGPT's cloud: no Codex cloud tasks, no GitHub code review and no Slack integration, and ChatGPT's credit multipliers no longer apply, because API pricing has its own Fast and Ultrafast rates. Even plan subscribers can mix the two: "All users may also run extra local chats using an API key, with usage charged at standard API rates." Signing in from the CLI takes one command, taken from OpenAI's authentication docs:

# Sign the Codex CLI in with an API key (billed at standard API rates)
printenv OPENAI_API_KEY | codex login --with-api-key

# Confirm which authentication method is active
codex login status

OpenAI's guidance is to "use API key authentication for programmatic Codex CLI workflows, such as CI/CD jobs," and not to "expose Codex execution in untrusted or public environments" - OpenAI Codex authentication docs. Enterprise workspaces have a third option, Codex access tokens, for trusted automation that needs workspace entitlements without a browser sign-in.

The API price list for Codex models

These are OpenAI's Standard API prices per million tokens for prompts up to 272K tokens, with longer prompts priced higher. Batch and Flex processing cost half, Fast costs double, and regional data-residency endpoints carry a 10% uplift for models released on or after March 5, 2026 - OpenAI API pricing.

Two details separate API billing from plan credits. The API charges cache writes on the GPT-6 family (for example $12.50 per million on Astra), while Codex credit billing "has no separate cache-write charge." And GPT-5.5 remains on the API after October 14, so API-key users are not forced to migrate on that date.

Break-even: when the API is cheaper

Using the example turn from section 6, one GPT-6.1 Sol turn costs about $0.10 at API rates and one Astra turn about $0.58. A Plus subscription at $20 therefore breaks even against the API after roughly 190 Sol turns or about 34 Astra turns in a month, and it allows up to 160 Sol messages in a single five-hour window. For anyone who uses Codex interactively most weeks, the subscription is almost always cheaper. The API wins in three situations: very light, irregular use; automated jobs in CI where you want exact per-token accounting and no shared allowance to drain; and workloads that need models or settings a plan does not offer. For a broader comparison of API prices across providers, see our cheapest LLM APIs for agents price table.

Why this matters: mixing payment methods is allowed and often optimal, but only if you know which credentials each process uses. How to apply it: keep your interactive work on a plan, put CI and scheduled jobs on an API key with a hard budget, and run codex login status in any script that matters, because a process that silently falls back to the wrong credentials either drains your plan or bills your API account.

10. Choosing a Model on a Budget, and the GPT-5.5 Retirement

Model choice is the largest cost lever you control, and OpenAI's own guidance is unusually direct about which model is for what. GPT-6 Astra is "our most capable model for complex work across code, apps, and research." GPT-6.1 Sol "offers near-Astra performance for complex work at a lower cost than Astra." GPT-6 Luna is "our most efficient model for focused, high-volume tasks, including summarization, extraction, and focused coding" - OpenAI Codex models docs. In ChatGPT, all three are available in Work and Codex but not in regular Chat.

A useful way to think about it is that OpenAI is telling you to make GPT-6.1 Sol your default: "For complex coding and agentic workflows, use GPT-6.1 Sol when available to your account and client." Astra is for "the hardest end-to-end work," and Luna is for tasks "when you know what a good result looks like, such as extraction, classification, transformation, and structured summaries." The desktop app's Power slider wraps these choices into presets (from Luna High through Astra Extra High), with an Advanced view for picking the exact model, effort and speed.

There is one curious gap in the numbers. On credits, Astra costs five times as much as GPT-6.1 Sol per token. But in the Plus five-hour estimates, Astra's best case (45 messages) is only about 3.5 times smaller than 6.1 Sol's (160). One plausible reading is that Astra often finishes a task in fewer tokens, needing fewer retries and less exploration. Another is that the allowance is simply calibrated differently from credits. Either way, it suggests Astra is a little less expensive in practice than its rate card implies, though GPT-6.1 Sol remains the efficient default for most work.

The October 14 retirement checklist

On October 14, 2026, GPT-5.5 "will retire from ChatGPT, ChatGPT Work, and Codex on all plans, including consumer, Business, Enterprise, and Edu plans." The API is not affected. OpenAI's replacements are GPT-6 Sol (gpt-6-sol) on Plus, Pro, Business, Enterprise and Edu, and GPT-6 Luna (gpt-6-luna) on Free and Go - OpenAI Codex models docs. It is the third retirement in two months: gpt-5.4 and gpt-5.4-mini left Codex with ChatGPT sign-in on August 31, and GPT-5.3-Codex-Spark retired on September 14.

Retirements are easy to underestimate because a pinned model rarely lives in one place. A developer who typed gpt-5.5 into the CLI once may also have it in a shared team config, a scheduled nightly task and a script that a colleague wrote months ago. OpenAI's migration guidance names the places a pinned model hides, and every one of them needs checking before October 14: workspace defaults and saved model settings, managed configurations pushed by administrators, custom agents that select a model, scheduled tasks and automations, and scripts that pass a model, such as codex exec --model. Administrators get their own page on workspace model availability, because a managed default that still says GPT-5.5 affects every member at once.

For most people the fix is one line in config.toml, which the desktop app, CLI and IDE extension share. Switching to GPT-6.1 Sol rather than the named GPT-6 Sol replacement is reasonable if it is available to your account, since OpenAI recommends it for complex work and its cached input is cheaper:

# Replace any gpt-5.5 pin before October 14, 2026
model = "gpt-6.1-sol"

Cost-wise, the migration is good news. At credit rates, GPT-5.5 costs 125 credits per million input tokens and 750 per million output; GPT-6.1 Sol costs 50 and 250. In our example turn, the same work drops from about 8.1 credits to 2.6. If GPT-5.5 was your habit because it felt familiar, the retirement is a forced price cut. For how these models rank on agent tasks beyond price, our October 2026 LLM ranking for agents covers the late-September releases, and our guide to AI model routing explains how to send each task to the cheapest model that can handle it.

Why this matters: a pinned GPT-5.5 is both a reliability risk on October 14 and an overpayment today. How to apply it: search your configs and scripts for gpt-5.5 now, default to GPT-6.1 Sol, escalate to Astra only for tasks that fail on Sol, and push repetitive extraction or classification work down to Luna.

11. Codex vs Claude Code, Cursor, Copilot and Gemini: Pricing Compared

Codex does not exist in a vacuum, and the most striking thing about the competition in October 2026 is how similar the pricing shapes have become. Anthropic, Google and OpenAI all use a short refill window under a weekly cap, and nearly every vendor now has an individual tier around $100 to $200. The differences are in the details: how the window resets, what overflow costs, and which models you get. All prices below come from each vendor's own pricing page or documentation as of October 7, 2026.

The Claude Code row comes from Anthropic's pricing page, which says "every plan has usage limits that reset on a rolling five-hour session window, and paid plans add weekly limits on top," with Claude chat and Claude Code drawing from the same pool - Claude pricing. One difference from Codex is worth highlighting: Anthropic's Max weekly limit "resets at a fixed time each week that is assigned to your account," and "your reset day and time stay the same regardless of when you start using Claude" - Claude Help Center. Codex's weekly date, by contrast, moves whenever you apply a banked or paid reset. Claude Code defaults to Claude Opus 5.5 on paid plans, while Claude Fable 5.1 bills to usage credits on Pro and can use up to 50% of weekly limits on Max.

Cursor and GitHub take a different approach: monthly pools rather than rolling windows. Cursor splits usage into a pool for its own models (Grok 4.7 and Composer 2.5) and a pool for third-party models charged at API prices, and its own docs say daily agent users typically spend $60 to $100 a month in total - Cursor pricing docs. GitHub Copilot now bills in AI Credits worth $0.01 each, with Pro at 1,500 credits, Pro+ ($39) at 7,000 and Max at 20,000 a month, reset on the first of the month - GitHub Docs. Copilot Pro+ and Max can also delegate tasks to OpenAI Codex itself, in preview.

Google sells coding capacity as part of its consumer AI plans: Google AI Pro at $19.99 and AI Ultra at $99.99 (5x Pro) or $199.99 (20x Pro), where Gemini app limits refresh "every 5 hours until you reach your weekly limit" - Google AI subscriptions. Its Gemini CLI uses daily request quotas instead, with 1,500 requests a day on AI Pro and 2,000 on AI Ultra - Gemini CLI quotas.

What the comparison actually tells you

At the $20 tier, the choice is mostly about models and ecosystem, because every vendor rations heavy use. At the $100 to $200 tier, the structures diverge. For anyone who subscribes after September 29, Codex's Pro 200 sells usage at exactly the same rate as Plus (10x the usage for 10x the price), while Google's $199.99 AI Ultra tier advertises 20x AI Pro's limits for about 10 times AI Pro's price, and Claude Max offers 5x or 20x Pro's per-session usage. Codex's counterweights are the absence of a five-hour limit on every Pro tier, the frequency of global resets, and the reach of one plan across ChatGPT Work and partner tools.

Reset mechanics may matter more than headline prices for heavy users. Anthropic's fixed weekly anchor is predictable; Codex's moving anchor rewards people who time their resets well and punishes those who do not. Copilot and Cursor's monthly pools are the easiest to budget but give no mid-month relief. For side-by-side capability comparisons rather than prices, see our guides to the best AI coding CLI, Claude Code pricing in June 2026 and the top Cursor alternatives.

There is also a category that sidesteps per-seat coding allowances altogether. Platforms like O-mega build and operate an entire autonomous company (website, app, billing, content and admin) through one conversation, so what you manage is the business you want running rather than an agent's token budget. It is a different trade-off from a coding agent you steer message by message: less hands-on control over each line, in exchange for not having to drive the agent through every session yourself.

Why this matters: the cheapest plan on paper is rarely the cheapest for your work pattern, because reset mechanics and overflow pricing dominate for heavy users. How to apply it: pick the tool by the models and workflow you want first, then compare at the tier you will actually use, and look closely at how each vendor's weekly reset is anchored if your work comes in bursts.

12. Why Codex Is Priced This Way: A First-Principles View

It is easy to read the limits as arbitrary, or as a straightforward attempt to squeeze subscribers. A better way to understand them is to ask what OpenAI is actually selling. It is not messages and not even tokens. It is GPU time: a share of a finite fleet of accelerators, at a given priority, during a given hour of the day. Every mechanism in Codex pricing is a tool for rationing that one scarce input among tens of millions of people whose demand is extremely uneven.

Start with the shape of that demand. OpenAI's own Work data shows a 90th-percentile user consuming about 22 times the median (section 8). A flat subscription is therefore an insurance pool: light users pay more than they cost, heavy users less, and the pool only works if the heaviest users cannot take unlimited amounts. The weekly cap is the pool's protection against the tail. Demand is also bursty in time, concentrated in working hours, and a single user running parallel agents can consume an afternoon's worth of capacity in minutes. The five-hour window smooths those bursts. Pro removes the window but keeps the weekly cap, which is exactly what you would expect if OpenAI is less worried about any one Pro user's bursts than about their weekly total.

Speed modes follow the same logic. Generating tokens faster generally means serving a request with less batching and more dedicated hardware, so a faster answer costs more GPU time per token. That is why Ultrafast is priced almost exactly in proportion to its speed: up to 8x faster token generation for 8x the included usage. OpenAI is charging for the machine time, not the words. Day 1 of the 28-day run, which made the default speed about 50% faster at no extra charge, is best read as an efficiency gain being passed through rather than a change in that underlying logic.

Resets are the most interesting lever, because they look irrational until you consider idle capacity. Compute bought for peak demand sits partly unused at other times, and when a new cluster comes online ("A lot of compute is online for this increase," as the Codex lead put it on September 29), the marginal cost of letting everyone use more is low. A global reset converts spare capacity into goodwill and publicity at little cost. The 28-day run formalizes this, and the Codex lead's own aside that the vote "seems" to be "rigged in reset's favor" acknowledges that users have learned to value refills above features.

Finally, the shared pool and Sign in with ChatGPT are a distribution strategy more than a pricing one. By making one subscription the wallet for Codex, ChatGPT Work, Excel, Word and 16 partner tools, OpenAI turns the ChatGPT plan into the default way to pay for AI work in other people's products. That makes the subscription harder to cancel, and it means competitors are not only competing with Codex, but with the account their users already have.

The counter-arguments deserve weight. Rationing by unpublished weekly caps transfers risk to users, who cannot budget against a number they cannot see. The September 29 change shows that the allowance behind a price can halve with about a month's notice, which is a real cost for anyone who planned a project around it. And the convergence of Anthropic, Google and OpenAI on nearly identical structures suggests these are industry-wide responses to the same compute constraint, not one company's choice. The honest conclusion is that Codex is cheap for steady, moderate users, unpredictable for heavy ones, and likely to stay that way until compute supply catches up with agentic demand. Our analysis of the true cost of LLM inference goes deeper into those supply economics.

Why this matters: once you see limits as GPU rationing, the rules stop looking arbitrary and become predictable. How to apply it: work with the grain of the system (cheap models for routine work, speed only when time is scarce, heavy jobs outside peak hours where possible), and treat any unpublished limit as something that can move.

13. Making Your Codex Limits Last: A Practical Playbook

OpenAI publishes its own advice for stretching an allowance, and it is better than most third-party tips because it targets the real cost drivers: context and output. Every token the agent reads on every turn counts, so the biggest savings come from shrinking what it must read, not from writing shorter prompts.

It helps to picture what a single agent turn actually contains. Alongside your request, the model rereads the conversation so far, the instruction files for the project, the definitions of every connected tool, and the output of every command it ran. Caching makes repeated context cheaper, but not free, and a bloated context compounds across every turn of a long session. With that picture in mind, the five habits on OpenAI's pricing page make sense as a single strategy - OpenAI Codex pricing:

• Control prompt size. Be precise, and remove unnecessary context.
• Limit source material. Point the agent only at relevant files.
• Match output to the need. Define audience, format and length.
• Slim down AGENTS.md. Nest files so each folder loads only its own rules.
• Limit MCP servers. Each one adds context to every message.

The last two are the ones most people miss. An AGENTS.md file at the root of a large repository is injected into context on every turn, so a 3,000-word file is a tax on every message you send. OpenAI recommends nesting instruction files so that each part of the codebase loads only the instructions relevant to it. Likewise, every connected MCP server adds tool definitions to the context whether you use them or not, so disabling servers you are not using in a session is free capacity. Our guide to writing loops for AI coding agents covers how to structure instructions for long agent runs without bloating every turn.

Our additions to OpenAI's list

Beyond context hygiene, a few settings and habits make a measurable difference to how far a plan goes. These come from the pricing mechanics covered above rather than from guesswork, and most of them take a minute to set once and then work in the background for every session afterwards.

The logic behind them is to attack the multipliers before the base. A slimmer prompt might save a tenth of a turn's cost; choosing a model five times cheaper, or switching off a mode that charges two and a half times the allowance, saves far more with no effort per task. So start with the biggest multipliers and finish with the habit that protects you from surprises.

First, default to GPT-6.1 Sol and escalate to Astra only when Sol fails, because model choice is worth up to 5x on its own. Second, turn on auto-review: it has been free since October 6 and, by OpenAI's own estimate, used to cost 2 to 10% of a plan. Third, keep Fast off by default, since it charges 2.5x your allowance per turn, and switch it on only while you are actively waiting. Fourth, cap partner apps under Settings, Usage, App limits, because their share of your plan is the one you only discover when your weekly bar empties unexpectedly. Finally, check /status before long tasks, so big jobs start with room to finish. Checking /status matters because OpenAI lets an active turn finish past the limit "subject to fair use limits," but blocks the next one, so a long refactor started with 5% left may stall partway through a multi-step plan.

If you run Astra daily, creator Eric Michaud's breakdown of why GPT-6 Astra drains Codex's weekly limit so quickly is one of the most-watched takes on the problem since Astra reached paid plans. It is a third-party view rather than official guidance, so weigh it against the mechanics above:

Failure modes to avoid

The costly mistakes with Codex pricing are rarely about the plan itself. They come from timing and configuration, and they repeat often enough in OpenAI's own help articles and community threads to be predictable. OpenAI's help center has separate articles just for paid resets, banked resets and missing credits, which tells you where people actually get hurt.

Almost every one of these mistakes is invisible until the day it bites: a configuration that worked for months, a purchase that looked sensible at checkout, a balance that seemed safe. Knowing them in advance is the cheapest protection there is, and each takes only a minute to rule out.

The five that cost the most are easy to name. A stale model pin breaks on October 14 for anything still set to gpt-5.5. A mistimed instant reset restarts your week and throws away whatever allowance was left. The grandfathering cliff halves old Pro 200 allowances after October 29. Hoarded credits expire 12 months after purchase with no refund. And uncapped team overflow is the default on Business, where seats have no credit limits until an admin sets them.

Most of these share a root cause: Codex pricing has several clocks (five-hour, weekly, monthly billing, 12-month credit expiry, retirement and grandfathering dates), and they do not line up. Writing the dates that apply to you in one place, and checking them when you plan a big project, prevents most surprises. For teams running agents unattended for hours, our guide to long-running coding agents covers checkpointing so a limit hit mid-task does not lose work.

Why this matters: two users on the same plan can get several times more work out of it depending on configuration alone. How to apply it: do a one-time setup pass this week (default model, auto-review on, Fast off, partner caps, slim AGENTS.md), then re-check /status habits for two weeks before deciding whether you need a bigger plan.

14. What Comes Next: The Dates That Will Change Your Bill

Codex pricing will keep moving, but several changes are already scheduled, and they are worth putting in a calendar. The near-term dates come straight from OpenAI's documentation and announcements; the further-out items are commitments OpenAI has made without dates. Together they suggest the price of a Codex subscription in November will look different from October's, in both directions.

Two open-ended items could change costs again. OpenAI says Ultrafast support for GPT-6.1 Sol is "coming later", which would give Pro 500 and Enterprise users a cheaper fast option than Astra Ultrafast. And on October 5 OpenAI announced that "over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union," in response to the EU AI Act - OpenAI. That is not a price change, but it is a policy change for Codex users in the EU; OpenAI's own tests showed detection falling from about 92% to 66% when 10% of words were replaced with synonyms - The Next Web.

The longer-term direction is clearer than any single date. Every vendor is moving toward a hybrid of flat subscriptions for typical use and metered credits for the tail, and OpenAI's Codex lead has said where he expects it to end: "Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent" - Thibault Sottiaux on X. In other words, the subscription's value is meant to come from API price cuts passed through, not from ever-bigger allowances. If that holds, the best Codex deals will come from model efficiency (the same task in fewer tokens) rather than from plan changes, which is another reason to keep your default model current.

Why this matters: several of these dates silently change what you pay or what works. How to apply it: put October 14, October 29 and the end of the 28-day run in your calendar now, re-check the Codex pricing page after each, and revisit your plan choice in the first week of November when the temporary generosity of October is gone.

Conclusion: How to Pick the Right Codex Plan

Codex pricing in October 2026 comes down to a simple rule hidden under a complicated surface: every subscription tier now sells agent work at the same rate, about $20 per Plus-sized allowance, and the tiers differ in headroom, speed and features rather than unit price. Model choice, speed mode and context size move your real cost far more than the plan does, and resets add free capacity at unpredictable moments. The decision tree below captures the choice for most readers.

Put more simply: start on Plus, default to GPT-6.1 Sol, turn on auto-review, and keep Fast off. Move to Pro 100 if the five-hour window blocks you more than once a week, to Pro 200 if you burn through Pro 100's weekly allowance, and to Pro 500 only if Astra's speed is worth $300 a month more than Pro 200 to you. Put CI and scheduled jobs on an API key, and if you run a team, buy Business seats with credit caps set before the first invite.

The habit that matters most is the cheapest one: check /status and your usage dashboard regularly, because Codex's limits, resets and promotions change faster than any static plan page. The rest of October is unusually generous thanks to the 28-day run, so use it to measure your real usage, and choose your November plan from that data rather than from the price list.

This guide reflects OpenAI Codex pricing, plans and usage limits as of October 7, 2026. Plans, allowances, promotions and model availability change frequently, sometimes daily during OpenAI's 28-day run, so verify current details on OpenAI's pricing page and your own usage dashboard before purchasing.