Best AI Agents for Slack 2026: 7 Ranked on Real Cost

Claude Tag, Slackbot, ChatGPT, Notion, Gemini, Dust and Perplexity Teammates compared on what the work really costs, what each agent can finish and how safely it runs in your channels

On November 4, 2026, Perplexity launched Teammates, long-running agents with their own computer, and gave them a door into team chat: "Add a teammate to a Slack channel and colleagues can hand it work or ask for updates on a task" - Perplexity. It joined a crowded room. Anthropic put Claude Tag into Slack channels in June, OpenAI connected its workspace agents to Slack in April and its always-on dots in September, Google listed Slack among the channels of its new Gemini agent in October, and Slack's own rebuilt Slackbot has been generally available since January. Every large AI lab now sells an agent that sits in a channel and takes work from whoever tags it.

Demand is real. Slack says Slackbot passed 1 million monthly active users, up 145% in a single quarter - Slack. Salesforce reported that Slack's Model Context Protocol server "surpassed 1 million active users within six weeks of launch" - Salesforce. The question for a team lead is no longer whether an agent belongs in Slack, but which one, and what it will actually cost once people start using it.

That is where the comparisons online break down. The seven agents in this guide are priced in seven incompatible ways: per seat, per channel usage, per action, per credit with no published dollar value, per workspace plan, inside a $200 individual subscription, or per credit that depends on which model answered. A Slackbot "advanced action" costs 200 Flex Credits, which works out to $1.00 at Salesforce's list price, while a simple Notion agent task uses about 5 credits, or $0.05. Anthropic, for its part, says plainly that "there's no published per-task cost guidance" for Claude Tag - Anthropic. Ranking these products on their headline price tells you almost nothing.

This guide ranks them on real cost: the plan you must buy, the meter that runs while the agent works, the model behind that meter, and the work you still have to check and redo. Every product is scored on the same five criteria, a 25-person team is worked through end to end, and every price links to the vendor page it came from. The specialist agents for code and incidents, the independent platforms and the open-source route are covered after the main seven, along with the security failures that make Slack a harder place for agents than it looks.

This guide reflects products, plans and prices as of November 14, 2026. Several of these agents are in beta or preview and their pricing changes often, so check the linked vendor pages before you buy.

Contents

1. What an AI Agent for Slack Actually Is
2. Why Every AI Lab Moved Into Slack in 2026
3. What "Real Cost" Means for a Slack Agent
4. Claude Tag: The Multiplayer Coworker Billed by the Channel
5. Slackbot: The Agent You May Already Be Paying For
6. ChatGPT in Slack: Workspace Agents, Dots and @ChatGPT
7. Notion Custom Agents: The Cheapest Meter for Knowledge Work
8. Gemini Agent: Google's Coworker With Its Own Identity
9. Dust: A Default Agent in Every Channel
10. Perplexity Teammates: The Newest Entrant, Built for Ongoing Work
11. The Real Cost Side by Side: A 25-Person Team Worked Through
12. Coding and Incident Agents You Tag in Slack
13. The Specialists, the Automation Platforms and the Open-Source Route
14. Security: What a Slack Agent Can Leak, and How Each Vendor Contains It
15. Where Slack Agents Work, and Where They Fail
16. How to Roll Out a Slack Agent Without Overpaying
17. What Comes Next for Agents in Slack
18. Conclusion: Which Slack Agent to Choose

The 7 Best AI Agents for Slack, Scored

Before the detail, here is the whole field in one table. The seven products below are general-purpose agents a team can put into Slack today and hand work by message: answering, researching, drafting, updating systems and, for the strongest of them, running multi-hour jobs. Coding and incident agents are scored separately in section 12, because they are judged on different work. Each product is scored from 0 to 10 on five criteria, every cell carries its evidence, and the final column is the weighted average, sorted from highest to lowest.

The weights come from the economics of delegation, which section 3 works through. An agent is worth adding when the value of the work it finishes exceeds what you pay for it plus the time spent checking and fixing its output. Real cost carries the most weight (35%) because it is the factor buyers most often misread: the sticker price is easy to find, while the meter, the model and the prerequisite plans are not. Work it can finish (25%) decides how much of the team's queue the agent can actually take, which is the value side of the same equation. Slack fit (15%) measures whether the agent works the way a channel works, with many people handing it work in threads. Control (15%) caps the cost of mistakes, and availability (10%) decides whether you can buy it today.

Three cautions on reading the table. The scores reflect publicly verifiable facts as of November 14, 2026, not hands-on benchmarking of every product. The gaps in the middle of the table are small: ChatGPT, Notion and Gemini sit within a quarter of a point of each other, and your existing software (Microsoft, Google or Notion, Salesforce or not) should decide between them. And a low score on availability or cost is not a verdict on quality. Perplexity Teammates rank last mostly because they are a research preview sold to individuals, not because the agent is weak.

1. What an AI Agent for Slack Actually Is

Slack has hosted bots for a decade, so it is worth being precise about what changed. A bot reacts to a command or an event: it posts a deploy notification, answers /weather, or files a ticket from a form. An AI assistant answers when you ask, usually in a side panel or a direct message, and its job ends when the answer is written. An AI agent takes a task, decides the steps, uses tools over minutes or hours, and comes back with finished work: a pull request, an updated CRM record, a research memo, a filed set of tickets. The difference is not intelligence alone. It is that the agent carries the work between your message and its reply.

That changes the questions a buyer should ask. With an assistant, the main questions are accuracy and speed. With an agent, three structural questions matter more: who can direct it, whose permissions it acts with, and who pays when it works. The products in this guide answer those questions very differently, and the answers determine both the cost and the risk.

Three shapes of Slack agent

The first shape is the personal agent that lives in Slack. Slackbot is the clearest example: it works for you, sees what you can see, and answers mostly in your own conversation with it. OpenAI's dots also fall here. OpenAI's documentation is explicit that "Only the owner can direct their personal dot" - OpenAI. A personal agent is cheap to govern because it inherits one person's permissions, but it does not take work from the team.

The second shape is the team coworker in a channel. Claude Tag is built for this: Anthropic describes "one Claude that interacts with everyone" in a channel - Anthropic. OpenAI's workspace agents, Dust, Notion Custom Agents, Atlassian Rovo and Glean all offer some version of an agent that a channel shares. This shape fits how Slack is actually used, because requests arrive from many people at once, but it raises harder questions: whose account does the agent use when five people ask it for five things, and who is billed?

The third shape is the specialist agent you tag for one kind of job: Cursor, Devin and GitHub Copilot for code, Traversal or Datadog's Bits for incidents. These are judged on their domain work and are covered in section 12.

Some products sit across shapes. Gemini's agent can work for you and can also spin up "coworker agents" with their own email address and identity, and Perplexity Teammates belong to one subscriber yet take hand-offs from colleagues in the channels they watch. When you evaluate one, ask which shape it is when it answers a message in a shared channel, because that is where its permissions and its bill are decided.

What does not count

Some well-known products touch Slack without being a Slack agent. Anthropic's Claude Duet, a personal agent launched as a research preview on October 21, reads Slack only through a connector, and its support pages note that Slack works "with limits" in Duet - Anthropic. Workflow tools that post AI summaries into a channel are useful, but they do not take work from people in the thread. This guide only ranks agents that a person can hand a task to inside Slack itself. For the personal-agent category, our comparison of Claude Duet and ChatGPT dots goes deeper.

2. Why Every AI Lab Moved Into Slack in 2026

The rush into Slack looks like a fashion, but it follows from what an agent needs. The hardest input for any agent is not intelligence; it is context and the hand-off: what exactly is wanted, by whom, with which constraints, and what has already been tried. In most companies that context already exists as text, in a Slack thread, written by the people who need the work done. An agent that sits where the request is written gets the request and its context in one place, and it can report back where the requester is already looking. Every agent vendor is competing for that moment, because whoever owns the hand-off owns the work that follows it.

That is also why Slack and its owner, Salesforce, have spent 2026 tightening control of the data layer while opening the front door. Three changes matter to anyone buying a Slack agent.

The data door narrowed for unlisted apps. In May 2025 Slack cut the rate limits for commercial apps that are not in its Marketplace: conversations.history and conversations.replies dropped to "1 request per minute" with a maximum of 15 objects per call, while internal apps keep 50 or more requests per minute. Slack's stated reason was that such apps "have the potential to exfiltrate large amounts of sensitive conversational data" - Slack. Its API terms now forbid third-party apps from using Slack data "to train a large language model" or to "bulk export" message data - Slack API terms.

A new, sanctioned door opened instead. On February 17, 2026, Slack made its MCP server and the Real-time Search API generally available, describing the latter as the evolution of its Data Access API, built to "retrieve relevant Slack data without storing customer information on external servers" - Slack changelog. The catch is in the developer docs: "Only apps published in the Slack Marketplace and internal apps can use MCP at this time; unlisted apps are prohibited from using MCP" - Slack developer docs.

Slack positioned itself as the place agents work, not just talk. At Dreamforce in September, Slack said Slack Code, its channels for coding agents, was available now, named Anthropic, Cognition, GitHub, Vercel and OpenAI among the agents already working in code channels, and said "Slackbot in Channels is in early pilot" - Slack. Earlier in the year it reported "300% growth in new custom agents built on Slack since January" - Slack blog. Slackbot itself now acts as an MCP client, calling other agents' tools from inside Slack.

The image above is Slack's own picture of the strategy: Slackbot takes the request, then reaches into Linear, Notion, Vercel and Claude to do the work. For buyers, the practical consequence is simple. An agent's quality in Slack depends on which data door it has. A Marketplace-listed app or an app your own admins build gets real-time search and MCP; an unlisted commercial app may be throttled to one history request a minute. Before you trial any vendor's Slack agent, check that it is listed in the Slack Marketplace or installed as an internal app. Our guide to MCP and A2A explains the protocol layer these agents now share.

The labs have their own reason to be here, and it is commercial. Seat-based chat subscriptions are capped by the number of employees; usage-based agent work is not. An agent that a whole channel can hand work to turns one company's Slack into a stream of billable tasks, which is why Anthropic, OpenAI, Google and Perplexity all built channel agents within six months of each other. That incentive is worth remembering when reading their pricing pages: the meter is where the revenue is.

3. What "Real Cost" Means for a Slack Agent

The sticker price of a Slack agent is the least informative number about it. To see why, start from what you are buying. You are not buying software; you are buying finished work: a question answered with the right source, a bug fixed, a renewal brief written, a ticket routed. The real cost of one finished piece of work has five parts, and vendors publish only some of them.

The first is the entry ticket: the plan or seats you must buy before anyone can use the agent. This includes prerequisites that are easy to miss. Slackbot's full experience requires Slack's Business+ plan or higher, Gemini's Slack app states that "A Gemini Enterprise license is required" and "A paid Slack plan is required" - Slack Marketplace, and Lindy's setup notes say "New Slack installations require a paid Slack plan" - Lindy.

The second is the meter: what each task draws while the agent works. Some vendors meter per action, some per credit, some per token at list price. The third is the model behind the meter, which multiplies everything: the same task can cost fifty times more on a frontier model than on a volume model. The fourth is review time: the minutes a person spends checking the work before it ships. The fifth is failure cost: redoing work the agent got wrong, or cleaning up something it should not have done.

The chart below collects the per-unit prices that vendors actually publish. They are not directly comparable (a Salesforce "action" is not a Notion "task"), but they show the range, and they show how few vendors put a dollar figure on a single unit of agent work at all.

Here is where those numbers come from. Salesforce sells Flex Credits at "$500 USD/Per 100k Credits", and a standard Agentforce action costs 20 credits, or $0.10 - Salesforce. Its November rate card prices a "Slackbot Standard Action" at 40 credits and a "Slackbot Advanced Action" at 200, which at list price is $0.20 and $1.00 - Salesforce rate card. Notion says "simple tasks use about 5 credits, medium tasks about 10, and complex tasks about 25" - Notion, and sells extra credits at $10 per 1,000. Datadog reports that the average cost of one autonomous investigation by its Bits agent fell from $25 at launch to "about $4" - Datadog.

That last figure is the most useful anchor in the chart, because it is a measured cost of a real multi-step agent job rather than a price for a unit. A thorough investigation that reads logs, queries metrics and writes a root-cause summary costs a few dollars. A quick answer costs a few cents. Most Slack agent bills are decided by how many of the tasks fall into the first category.

The model multiplies the meter

For the agents that bill by token, the model choice moves the bill more than any other decision. To make that concrete, this guide uses three task sizes throughout, stated so you can replace them with your own numbers. A quick task (answer a question from a thread, summarize a discussion) is about 15,000 input tokens and 1,000 output tokens. A medium task (research a question across two or three tools and draft a reply or a doc) is about 100,000 input and 5,000 output. A heavy task (reproduce a bug and open a pull request, or assemble a report from several systems over an hour) is about 600,000 billed input tokens and 30,000 output. These are estimates for illustration, not vendor figures; heavy agent loops vary widely, and prompt caching can cut the input share considerably.

At list prices, the spread is enormous. Anthropic's top model, Claude Fable 5.2 (released November 9), costs $8 per million input tokens and $40 per million output tokens. Claude Opus 5.5 costs $4 and $20, Claude Sonnet 5.5 and OpenAI's GPT-6.1 Sol both cost $2 and $10, and Meta's new volume model Muse Ember costs $0.10 and $0.50 for prompts up to 200,000 tokens - Meta.

A heavy task on Fable 5.2 costs about $6 at these sizes; the same task on Muse Ember costs under eight cents. That does not mean the cheap model is the right choice: a heavy task is exactly where a stronger model earns its price, because a failed attempt costs both the tokens and a person's time. Anthropic reports that Fable 5.2 scores 41.3% on the Remote Labor Index, a benchmark of real freelance projects, against 17.9% for Fable 5.1 - Anthropic. If a stronger model finishes twice as many heavy jobs without help, it can be cheaper per finished job even at twice the token price. Our Fable 5.2 versus Opus 5.5 cost analysis works through that trade-off in detail, and the Muse Ember, Haiku 5.5 and GPT-6 Luna comparison covers the volume end.

The practical rule that follows: match the model to the channel, not the company. An intake channel answering policy questions can run on a mid-tier model; an engineering channel where the agent opens pull requests deserves the strongest model you can justify. Claude Tag lets admins change the model per scope, and Perplexity lets you choose a model for each teammate, which makes this rule enforceable rather than aspirational.

Seats versus usage: the structural choice

The deepest difference between these products is not the price level but the shape of the price. Seat-priced agents (Slackbot inside Business+, ChatGPT Business, Gemini Enterprise, Dust) charge for every person who might use the agent, whether or not they do. Usage-priced agents (Claude Tag channel work, Notion credits, Agentforce actions) charge only for work done, but the bill grows with every heavy task.

Neither is cheaper in general. Seat pricing wins when usage per person is high and predictable; usage pricing wins when a few people use the agent a lot and most use it rarely, which is how most Slack channels behave in the first months. Seat pricing also hides a trap specific to Slack: a channel is multiplayer by nature, so an agent priced per seat either has to be bought for everyone in the channel or has to be restricted to some of them, which defeats the point of putting it in a shared channel. Anthropic's design choice here is explicit: "Adding Claude to Slack doesn't add a per-seat charge" - Anthropic. Runbear, a smaller vendor, markets the same idea with the phrase "No seat tax" - Runbear. Section 11 puts numbers on both shapes for a real team size.

4. Claude Tag: The Multiplayer Coworker Billed by the Channel

Claude Tag is Anthropic's agent for Slack channels, launched on June 23, 2026 "in beta for Claude Enterprise and Team customers" - Anthropic. It ranks first here because it is the only product in the field designed from the start around how a Slack channel actually works: many people, many requests, one shared thread history, and work that has to stay visible to everyone who cares about it.

How it works in a channel

Anyone in a channel where Claude has been added can mention @Claude and hand it a task. Anthropic's documentation lists the kind of work it expects: "reproduce a bug and open a pull request, turn a decision thread into a doc, assemble the state of a project." Claude replies in the thread, posts a checklist as it works, and "the whole exchange stays visible to the channel" - Anthropic docs. Each conversation runs in an ephemeral sandbox in Anthropic's cloud that is created when the conversation starts and discarded when it goes idle, so a task can clone a repository, run tests and open a pull request without touching anyone's laptop.

Two design choices separate it from a chatbot with a Slack connector. First, "there's one Claude that interacts with everyone" in a channel, and with ambient behavior switched on it will "follow up on threads or tasks that have gone quiet without being resolved" - Anthropic. Second, administrators can view "a log of everything that @Claude has done, along with who requested each task", which matters when an agent is taking work from forty people at once.

The video Anthropic released at launch shows the basic loop: a teammate tags Claude in a thread, Claude posts its plan, works, and returns with the result in the same place.

What Claude can reach is configured centrally rather than per person. Admins connect tools with credentials that belong to Claude itself: the setup guide tells them to create a mailbox such as claude@yourcompany.example.com, invite it into each tool "the way you'd add a new teammate", and give it "the narrowest role that covers the work, read-only where the tool offers it" - Anthropic setup guide. Those connections can then be scoped to all of Slack, one workspace or one channel, and the model can be changed per scope as well. Members can also let Claude use their own claude.ai connectors for their own requests.

What it really costs

Claude Tag has the most buyer-friendly pricing shape in this guide, and the least predictable price level. The shape: "Adding Claude to Slack doesn't add a per-seat charge" and "Slack users don't each need a Claude seat to work with Claude in channels." Channel and thread work draws from an organization usage balance under a monthly spend limit that an Owner sets, while one-to-one direct messages bill to the sender's own Claude seat - Anthropic docs. Reading a channel, deciding whether to reply, and short replies from what Claude already knows are not billed at all - Anthropic billing docs. Spend is counted at list price, and Claude posts a heads-up in the channel at 75% and 90% of the limit.

The entry ticket is a Claude Team or Enterprise plan. Team seats cost $25 per seat per month billed monthly ($20 billed annually) for teams of 2 to 150, and Enterprise costs "US$20/seat/month, billed annually, 20 seat minimum" plus "usage at API rates" - Anthropic pricing. A Team plan must buy usage credits before anything runs: "on a Team plan nothing runs until credits are loaded." The setup flow preselects a $2,500 monthly spend limit, with presets from $500 to $5,000, unlimited, or a custom amount up to $1,000,000 - Anthropic setup guide. A launch credit of $2,500 per Team organization and $25,000 per Enterprise organization expired on October 1, 2026 - Claude Help Center.

The unpredictable part is the price of a task. Anthropic says "there's no published per-task cost guidance" and recommends running a pilot with a spend limit and watching the per-channel breakdown. Using this guide's illustrative task sizes, a quick answer on Opus 5.5 costs about eight cents and a heavy coding task about $3, so a channel's bill depends almost entirely on how many heavy tasks it hands over. The analytics page helps here: it splits channel spend into Engaged, Proactive, Scheduled and Monitoring work, so you can see whether ambient follow-ups are driving cost and switch off "Respond automatically" in channels where they are not worth it.

Where it is strongest

Claude Tag is strongest where a channel's work ends in a change to a system: engineering channels (bugs to pull requests), data channels (questions to warehouse queries and charts), and operations channels (alerts to investigations). Anthropic's own claim is striking but internal and unaudited: "Today, 65% of our product team's code is created by our internal version of Claude Tag" - Anthropic. Cat Wu, who leads the product, described the design goal to Fortune: "Claude Tag is built to be interactive and multiplayer" - Fortune.

Where it falls short

It is still a public beta, and it is not sold on individual plans: Free, Pro, Max and Ultra users cannot buy it. Organizations with Zero Data Retention, customer-managed encryption or the HIPAA configuration for local Claude Code and Cowork are excluded, because "Claude Tag stores channel memory and session transcripts, which ZDR doesn't permit" - Anthropic setup guide. The shared-connection model also has a sharp edge, covered in section 14: whatever an admin-connected account can read or write is available to every member of the channels where that connection applies. And teams that used the earlier Claude app in Slack should know it stopped answering in channels of paired workspaces on October 5, 2026; the older Claude Code in Slack integration now answers only Pro, Max and Ultra accounts in workspaces not connected to Claude Tag - Claude Code docs.

Verdict: the best agent for a team that wants one shared coworker in its channels and is willing to manage a usage budget. Start with a modest spend limit, one model per channel, and a read-only first connection.

5. Slackbot: The Agent You May Already Be Paying For

The cheapest Slack agent for many teams is the one they already own. Salesforce relaunched Slackbot as "your personal agent for work" and made it "generally available for Business+ and Enterprise+ customers" on January 13, 2026 - Salesforce. By Dreamforce in September it had passed one million monthly active users.

What it does

Slackbot's advantage is access. It searches everything you are allowed to see in Slack, with Slack's own permissions, and it reaches beyond Slack: Slack's help center lists that it can "View your Google or Outlook calendar", "Search and take actions in Salesforce", and run deep research using "connected MCP servers" - Slack Help Center. It can create canvases, slides and documents, run scheduled tasks, and use skills that a workspace defines.

The screenshot shows the pattern Slackbot is built for: a salesperson asks for a renewal brief, Slackbot searches the workspace and the web, and the result lands as a canvas the team can edit. That is valuable work, and it is personal work. Slackbot answers the person who asked; it is not yet a coworker that a channel shares. Slack says "Slackbot in Channels is in early pilot" - Slack, and the Dreamforce keynote below is the best view of where Slack is taking it.

What it really costs

On Slack's US price list, Pro costs $8.75 and Business+ costs $18 per user per month billed monthly, with annual billing saving "approximately 16%"; Enterprise+ is sold through sales - Salesforce. Business+ includes Slackbot, which the price list calls "your personal AI agent for work". Slack retired its separate AI add-on in 2025: "The Slack AI add-on is no longer available to purchase on the Slack website" - Slack Help Center.

The limits are the part to read carefully. "Each member of a Business+ workspace can send 15 messages to Slackbot per week", with "fifteen automatic task runs per day"; Enterprise+ members "are not subject to weekly or lifetime message limits"; Free and Pro workspaces only get Slackbot "during a trial or preview" - Slack Help Center. Beyond those limits, usage is bought with Salesforce Flex Credits: 40 credits per standard action and 200 per advanced action, which is $0.20 and $1.00 at the $500 per 100,000 list price. Using Flex Credits for Slackbot requires a Salesforce Enterprise or Unlimited edition with a Digital Wallet, and Salesforce customers on Pro Suite or higher get Slackbot on any Slack plan once Slack is connected to their Salesforce org.

That makes Slackbot's real cost unusually easy to reason about. For a team already on Business+, the marginal cost of the first 15 messages per person per week is zero. For a team on Pro, the real cost is the upgrade: $9.25 per user per month at monthly prices, for every member of the workspace, whether they use Slackbot or not.

Agentforce in Slack is the heavier sibling for teams that build their own agents on Salesforce. It is "Available on all Slack plans with a Salesforce Agentforce license", but agents for Slack must be built as the Agentforce Employee Agent type - Slack Help Center. Agentforce is priced at $2 per conversation, through Flex Credits, or through per-user licenses such as the $5 Agentforce User License and the $50 Agentforce Coworker - Salesforce. Our guide to Salesforce's named Agentforce agents covers that side.

Where it falls short

Slackbot is weaker at long, multi-step execution than the agents above and below it. It does not run code in a sandbox or open pull requests, and its deep research is bounded (Slack documents runs of up to ten minutes). Fifteen messages a week is plenty for occasional questions and too few for someone who wants an agent to carry real work. Its security record also deserves a note: when researchers at PromptArmor showed in 2024 that a planted message in a public channel could make Slack AI leak data from private channels, Slack's first response was "This is intended behavior" - PromptArmor. Slack has since rebuilt its AI stack, but the episode shows why permissioned search is not the same thing as injection-proof search.

Verdict: if you are already on Business+ or Enterprise+, switch it on before buying anything else and measure what it cannot do. If you are on Pro, compare the per-member upgrade against the usage-priced agents in section 11 before upgrading for Slackbot alone.

6. ChatGPT in Slack: Workspace Agents, Dots and @ChatGPT

OpenAI reaches Slack through three products that overlap, and understanding which is which is half the work of buying it.

Workspace agents, launched on April 22, 2026 for ChatGPT Business, Enterprise, Edu and Teachers plans, are agents a company builds and shares. VentureBeat described the Slack version precisely: "When that same agent is deployed to a Slack channel, it is a Codex instance listening for mentions and threading its work back in" - VentureBeat. VentureBeat quoted Rippling's Ankur Bhatt describing an agent that "posts deal briefs directly into the team's Slack room", and reported that write actions require approval by default. Our guide to OpenAI workspace agents covers how to build and govern one end to end.

Dots, launched at DevDay on September 29, are always-on agents with their own cloud computer, powered by GPT-6 Astra, which you can reach "in Slack and Teams" - SiliconANGLE. A dot is personal: you can message it directly or mention it in a channel, but "Adding it to a channel alone doesn't start monitoring" - OpenAI. OpenAI's launch video shows the always-on model in action.

@ChatGPT is the third route and the one engineers will use most. The old @Codex app is being folded into it: "If you previously used @Codex, start new requests with @ChatGPT after your admin enables the ChatGPT app" - OpenAI. Repository work is handed to a Codex cloud task that "runs in a separate task using your account", it works in public and private channels, and "Slack Connect channels shared with another organization aren't supported" - OpenAI docs.

What it really costs

ChatGPT Business costs "$20/ user / month" billed annually for two or more users, or $25 billed monthly - OpenAI. Agent work beyond plan allowances is paid in credits from a pool shared by Codex, ChatGPT Work and workspace agents, and here the transparency stops: "Credit purchase prices and applicable discounts depend on your plan or agreement." The same page does publish credit rates per model: GPT-6.1 Sol costs 50 credits per million input tokens and 250 per million output, and GPT-6 Astra costs 250 and 1,250.

One inference makes those numbers usable. Every rate on that table is exactly 25 credits per dollar of the model's API list price (GPT-6.1 Sol's API price is $2 and $10 per million tokens), and an August promotion valued 2,500 workspace credits at $100 - IT Brief. If your agreement buys credits near four cents each, agent work in ChatGPT costs roughly what the same tokens cost on OpenAI's API. That is this guide's arithmetic, not an OpenAI price, so confirm your own rate in Settings before budgeting.

Dots cost more to reach. Each Pro plan ($100, $200 or $500 a month) and each Business Premium seat includes one dot, and Pro dots are not offered in the European Economic Area, Switzerland or the UK - SiliconANGLE. Premium seats cost $125 per user per month, or $100 on annual billing, and provide "five times more usage than Standard seats" - IT Brief. For Enterprise workspaces, dots are switched on by default from November 18, 2026 unless the workspace opts out - OpenAI.

The structural cost issue is that ChatGPT's Slack app connects each person's own account. OpenAI's docs say "Workspace setup and your personal account connection are separate", and Codex work runs under the requester's account - OpenAI. In practice, everyone who hands ChatGPT work from Slack needs a seat in your ChatGPT workspace, which is why it scores lowest on cost among the top seven despite competitive token rates.

Where it is strongest, and where it falls short

The strengths are capability and breadth. Dots run on GPT-6 Astra with a full cloud computer, workspace agents can be built by non-engineers, and @ChatGPT gives engineering channels Codex's cloud tasks. OpenAI's safety design for dots is among the most explicit in the field: background research "is read-only", and its security notes warn that "an email asking a dot to forward a private file does not authorize the dot to send it" - OpenAI.

The weaknesses are fragmentation and opacity. Three overlapping products, a credit price you cannot see before you sign, and personal dots that do not take work from a channel make ChatGPT harder to roll out as a shared channel coworker than Claude Tag. Teams already standardized on ChatGPT Business will find it the path of least resistance; teams choosing fresh should price the seats for everyone in the channel first.

Verdict: the strongest choice for organizations already paying for ChatGPT Business or Enterprise, especially engineering-heavy ones. Not the cheapest way to put a shared agent in front of a whole team.

7. Notion Custom Agents: The Cheapest Meter for Knowledge Work

Notion Custom Agents launched on February 24, 2026 and have used Notion credits since May 4: "Starting May 4, 2026, they'll use Notion credits, available as an add-on for Business and Enterprise plans" - Notion. They are the most transparent meter in this guide, and for teams whose knowledge already lives in Notion they are the cheapest way to get a Slack agent that answers from it.

How it works in Slack

A Custom Agent can be triggered from Slack by a new message, an emoji reaction or an @mention, can include thread replies, and "can read and write to Slack through a dedicated integration" once a Slack admin connects Notion's Slack AI connector - Notion Help Center. The typical jobs are triage and answers: watch an intake channel, classify each request, answer what the docs already answer, file the rest into a Notion database, and post the link back in the thread. Notion's demo shows an agent answering questions in Slack from a team's documentation.

There is one inconsistency to check before you rely on private channels. Notion's help page says triggers work in public and private channels, while the product page says "Read and post messages in public channels" and "Private channels coming soon" - Notion. Test it in your own workspace.

What it really costs

Notion Business costs $20 per member per month billed yearly or $24 billed monthly, and agent usage beyond the included amount is "Free to try, then $10 per 1,000 monthly Notion credits" - Notion pricing. Each member gets 1,500 Standard credits and 500 Select credits a month, and by default "a Custom Agent uses its owner's Select credits, even when other people use it." With simple tasks at about 5 credits, medium at 10 and complex at 25, a Notion agent task costs between five and twenty-five cents at list price.

That design has a useful consequence for Slack: people who trigger the agent from a channel draw on its owner's credits, so the channel does not need a Notion seat for every member to use one agent. It also has a trap: one owner's 500 Select credits cover only about 50 medium tasks, after which the workspace-managed balance has to be switched on.

Where it falls short

Notion agents are excellent at work that ends inside Notion and limited outside it. They are not built to open pull requests, operate a browser for an hour, or act across a dozen external systems. The vendor-published customer stories are positive but unaudited, and the product page quantifies none of them. For a team that runs on Notion, that is fine; for a team that wants one agent to do everything, it is a component, not the whole answer.

Verdict: the best value in this guide for answer-and-triage channels in Notion-based teams. Pair it with a heavier agent for execution work rather than stretching it.

8. Gemini Agent: Google's Coworker With Its Own Identity

Google announced the Gemini agent on October 8, 2026, at its Gemini at Work event, as "a single, universal agent for work" that can be reached "from any device" and "any channel (command line, Google Workspace, Microsoft 365, or Slack)" - Google Cloud. Its distinctive idea is identity. Gemini can act as a "coworker agent" with "a persistent, defined role", and "Coworker agents have dedicated identities including their own @agents.company.com emails", with a Workspace account, calendar and Drive of their own.

The screenshot shows the agent's home: a task box with an Ask for approval control and a list of finished jobs such as a weekly status report. In Slack, the same agent is reached through Google's Slack app. Google's launch video walks through the coworker model.

What it really costs

Gemini Enterprise Business starts "at $21 per seat per month" with a 30-day trial - Google Cloud, and Reworked reported that the agent comes "at no additional charge" for Gemini Enterprise customers - Reworked. The Slack app's listing is explicit that "A Gemini Enterprise license is required" and "A paid Slack plan is required." Google also describes real-time spend caps: "Gemini monitors token usage and sandbox costs, and if a spend cap is triggered, that project's agent pauses." That implies consumption charges for heavy agent work beyond the seat, but Google's announcement does not publish the rates, so budget a pilot before you assume the seat price is the whole bill.

Where it is strongest, and where it falls short

The governance design is the most complete of any agent here. Each agent gets "its own identity, cryptographically attested", "every action Gemini takes is written to an audit trail and attributed to the agent rather than to a person", and all traffic passes through Agent Gateway, which Google calls "an AI network firewall". The agent also routes each job across "our Gemini family of models and Claude models from Anthropic", so model choice is the platform's job rather than yours.

The weaknesses are newness and center of gravity. The agent was announced five weeks ago and there is little independent evidence of how it performs in Slack specifically; TechCrunch's launch coverage named On, Shopify and PayPal as early testers - TechCrunch. And it is built around Google Workspace: Slack is one of its channels, not its home. Our guide to Gemini's coworker agents goes deeper on the identity model.

Verdict: the right pick for a Google Workspace company that uses Slack for chat and wants agents governed like employees. Less compelling if your documents and systems live elsewhere.

9. Dust: A Default Agent in Every Channel

Dust is a platform for building agents on top of many models and connectors, and Slack is one of its main surfaces. It reports "300,000+ Agents deployed, 3,000+ Teams running on Dust" - Dust. Its Slack design is the most aggressive of the seven: an admin can make an agent the default for a channel, after which "the agent set as the channel's default agent responds to messages posted in the channel, not just to @Dust mentions" - Dust docs.

The screenshot captures Dust's typical job: someone asks for a specific piece of internal knowledge, and the agent answers in the thread with the link, a numbered source and a confidence label.

What it really costs

Dust Pro costs €30 per seat per month, or €24 billed yearly, with 8,000 credits per seat per month; Max costs €150, or €120 yearly, with 40,000 credits; Enterprise is custom - Dust pricing. Credits are charged per message according to the model and tools used, and they do not roll over. Dust does not document whether a Slack user without a Dust seat can talk to its agents, so price a seat for everyone you expect to use it until your own trial shows otherwise.

Where it is strongest, and where it falls short

Dust's strength is choice: more than 20 frontier models, agents that call other agents, and connectors to the usual systems. Clay reports that its 20-person go-to-market engineering team "saves around 58 hours each month" with a Dust agent in a dedicated Slack channel - Dust, a vendor-published case study.

Its auto-reply mode shows the trade-off of an always-answering agent. In "all messages" mode, the agent must reply to every message, including thread replies, and the documentation is candid that instructions cannot make it stay silent: an empty response is treated as a failed generation and retried. That is ideal for a help-desk channel and noisy anywhere people also talk to each other.

Verdict: a strong choice for teams that want to build several specialized agents across models and give each channel its own default. Seat pricing makes it a better fit for teams where most members will use it.

10. Perplexity Teammates: The Newest Entrant, Built for Ongoing Work

Perplexity Teammates launched on November 4, 2026 and are the news that prompted this guide. They are agents built for complex work that continues over days: each teammate "runs on its own isolated computer, with its own desktop and browser", runs on schedules, and can start sub-agents in parallel - Perplexity Help Center. In Slack, a teammate watches the channels you choose, and "colleagues can hand it work or ask for updates on a task."

Teammates differ from the other six in what they are for. Claude Tag and Slackbot are built to answer a request and finish it; a teammate is built to own a stream of work over days, such as monitoring a market, running a recurring QA pass or keeping a research brief current, and to report into Slack as it goes. New teammates use Claude Opus 5.5 by default, and you can choose a different model for each one.

What it really costs

Teammates are included in Perplexity Max, which costs "$200/monthly or $2000/annually" - Perplexity Help Center. For now usage is effectively free: "For a promotional period, every Max plan will include free usage for Perplexity teammates", and afterwards "teammates use the credits that are already included with your Max plan" - Perplexity. Perplexity does not publish the size of that monthly allocation or when the promotion ends, and it warns that "Complex projects can use more credits, particularly when agents run in parallel."

For a team, the economics are unusual. Because a teammate belongs to one Max subscriber but takes hand-offs from colleagues in the channels it monitors, one $200 subscription can put a capable agent in front of a whole channel. Digital Trends' headline still captured the general reaction: the new coworkers "don't come cheap" - Digital Trends, and $200 is a lot for an individual even if it is modest for a team.

Where it falls short

The low rank comes from availability and control, not ability. Perplexity's help center says "The initial research preview requires an individual Perplexity Max subscription", and "Enterprise Pro and Enterprise Max are not included by default in this launch." That means no admin console governing which channels a teammate watches or what it may do there: the controls are the owner's. Teammates use the owner's installed plugins and skills with those plugins' existing permissions, which IT Brief also highlighted in its coverage of the launch - IT Brief. The isolation design is good (credentials are encrypted and "Teammates cannot view them"), but a company should not let an individual's subscription become the de facto agent for a shared channel without deciding who is accountable for it.

Verdict: the most interesting long-running agent here for individuals and small teams, and a cheap way to test an always-on teammate in one channel. Wait for the Enterprise version before making it a company standard.

11. The Real Cost Side by Side: A 25-Person Team Worked Through

Prices only become comparable when they are applied to the same team doing the same work. So take one concrete case and run every product through it, with each assumption stated so you can swap in your own.

The team: 25 people, already paying for Slack Pro on monthly billing, with no Claude, ChatGPT, Gemini, Notion, Dust or Perplexity plan yet. The work: 500 agent tasks a month, about 20 per person, split into 300 quick tasks, 150 medium tasks and 50 heavy tasks, at the token sizes defined in section 3. The prices: monthly list prices from the vendor pages cited in sections 4 to 10, with no negotiated discounts.

Here is what each product costs that team, and what it can and cannot do for the money.

The chart shows the same numbers split into the two parts of every bill: what you pay before anyone uses the agent, and what the work itself draws. Dust is left out because it prices in euros.

What the numbers say

The cheapest bills are cheap because the agent does less. Notion at about $62 and Slackbot at about $231 cannot do the 50 heavy tasks at all, so a team that needs them buys a second tool. Measured per task they can actually finish, Slackbot costs about $0.51 for each of the 450 lighter tasks; Claude Tag on Opus 5.5 costs about $0.60 per task across all 500, heavy work included.

Seat pricing sets a high floor; usage pricing sets a steep slope. ChatGPT and Gemini charge $525 to $625 before the first task, and the cost per extra task is low. Claude Tag charges $50 before the first task, and every heavy task adds about $3 on Opus 5.5 or $6 on Fable 5.2. Solve for the crossover and the result is instructive: with Claude Tag on Opus 5.5 at an average of about $0.50 a task and ChatGPT at about $0.25 a task in credits, the two bills only meet at roughly 2,300 tasks a month, or about 92 per person. Below that, the usage-priced agent is cheaper even on a model with twice the token price. Against Slackbot's flat $231, Claude Tag on Opus becomes more expensive above roughly 360 tasks a month, but only for the kind of task Slackbot can do.

The model is the biggest lever inside a usage-priced bill. Moving Claude Tag's channels from Opus 5.5 to Fable 5.2 nearly doubles the usage line; moving quick-answer channels to Sonnet 5.5 halves it. Most teams should not run every channel on the same model. Our guide to model routing for agents and the one on prompt caching explain the two techniques that cut token bills most.

Annual billing changes the seat-priced products more than the usage-priced ones. Slack saves about 16% on annual terms, ChatGPT Business drops from $25 to $20, Claude Team from $25 to $20, and Notion Business from $24 to $20. Usage lines do not move.

The cost no vendor puts on its pricing page

Every number above is small next to the time people spend checking the agent's work. Suppose, as an illustration, that reviewing a quick task takes 3 minutes, a medium task 10 minutes and a heavy task 30. For the team above, that is 3,900 minutes, or 65 hours a month. At a loaded cost of $60 an hour, review costs about $3,900 a month, more than five times the price of the most expensive agent in the table.

That is the most important number in this guide, and it changes how to read the ranking. An agent that costs $300 more a month but halves review time is the cheaper agent. This is why the scores weight the quality of finished work almost as heavily as price, why sources and confidence labels in an agent's reply (Dust's Slack replies show both) are worth real money, and why the first metric to track in a pilot is not spend but review minutes per finished task. Our analysis of why most agent pilots never scale shows how often this hidden cost is what kills them.

12. Coding and Incident Agents You Tag in Slack

Engineering teams were the first to hand Slack threads to agents, and the specialist agents for code and incidents are judged on different work than the generalists above: a merged pull request or a correct root cause, not a good answer. They also bill differently, mostly through the coding platform's own plan.

Slack made the category official on August 20, 2026 with Slack Code, channels where developers and agents build together, launched with Claude, Devin, GitHub Copilot and Vercel's agent, with ChatGPT "available soon" - Slack. GitHub's screenshot shows what that looks like: a code channel per task, the agent's messages in the thread, and the diff in a side panel.

In the image, a developer asks @GitHub for a feature, a colleague redirects the design in the thread, and the agent reports "Fixed and pushed to PR #7" with the changed files open beside the conversation. The table compares the main coding agents you can call this way.

GitHub Copilot has the clearest meter. Its plans page lists Pro at $10 with $15 of monthly AI credits, Pro+ at $39 with $70, and Max at $100 with $200, and states that "1 AI credit = $0.01 USD" - GitHub plans. The new Slack experience, in public preview since August, is available "to organizations on GitHub Copilot Business and GitHub Copilot Enterprise plans" and draws on "existing Copilot cloud agent budgets" - GitHub changelog. In a thread, Copilot "will capture the entire thread as context", and Slack guests cannot start or steer a session. At GitHub Universe on October 28, GitHub went further: "GitHub Enterprise customers can use all Copilot features without buying a separate subscription", paying for the AI credits they use - GitHub blog. For GitHub Enterprise shops, that makes @GitHub in Slack close to free to switch on and metered entirely by usage.

Cursor starts a cloud agent when you type @Cursor with a prompt, and its agents "read the whole thread for context" - Cursor. Setup requires turning on usage-based pricing, and cloud agents are included only from the Premium plan at $30 a month; third-party model usage "is not included in the monthly prices shown and must be purchased separately" - Cursor pricing. Cursor also carries the most vendor risk in this table. SpaceX completed its $60 billion acquisition of Cursor on August 14, and OpenAI told Cursor its "proposed date is Nov. 12, 2026" for ending access to its models, saying "we cannot be confident that SpaceX will use our technology within our terms of service" - CNBC. As of November 14 we found no official confirmation of what happened on that date, and Cursor's model documentation still listed OpenAI's GPT-5.6 models. Cursor's chief executive said OpenAI models served "about 5% of Cursor user traffic", so the practical impact is limited, but the episode is a reminder that a Slack coding agent is only as stable as its model supply. Our explainer on the SpaceX and Cursor deal has the background.

Devin, from Cognition, was built around Slack from its first release. You tag @Devin in any channel, it replies in the thread, and "code channels" give each session its own Slack channel - Cognition. Pricing runs from Pro at $20 to Teams at "$80/month for team plan + $40/mo per full dev seat", with overage "consumed at API pricing" - Devin. Cognition's August video shows the Slack workflow.

GitHub also shipped Copilot Fusion at Universe, which routes and orchestrates several models on one task and is billed "for each model it uses, at each model's standard rate" - GitHub changelog. For Slack users, the significance is that the model question is moving inside the agent: you hand the task over in a thread, and the platform decides how many models it is worth spending on. Our comparison of Claude Code, Codex and Cursor covers the coding agents themselves in depth.

Incident agents: the Slack channel as the war room

Incidents are the purest Slack use case, because the incident channel already holds the alert, the people and the timeline. Traversal raised a $120 million Series B at a $1 billion valuation on November 3, with customers including Nubank, Harvey, DigitalOcean, PepsiCo and American Express, and more than 150,000 investigations last quarter - The Next Web. In Slack, @Traversal "starts or joins the incident channel the moment something breaks, or investigates the moment you mention it" - Traversal; pricing is by request.

Datadog's Bits agent is the best-documented on cost. You reply to an alert in Slack with "@Datadog Investigate this alert" and it uses the whole thread as context - Datadog docs. An investigation uses about 4 AI credits, sold at $1.30 each on demand or from $500 for 500 a month on annual terms - Datadog pricing, and Datadog reports average investigation time falling from 11 minutes to about 2 - Datadog blog. incident.io charges $19 per user per month on Team ($15 annual) and $25 on Pro, with its AI investigations sold as an add-on - incident.io pricing. It also makes a strong safety promise: "The only change Investigations can make to your systems is a pull request you review and merge yourself" - incident.io. PagerDuty's Paige agent recommends remediation and executes it only "if approved", metered in AI Actions - PagerDuty.

For incident work, the right comparison is not agent against agent but agent against the cost of downtime. A $4 to $5 investigation that shortens a high-severity incident by ten minutes is the cheapest line item in the whole response.

13. The Specialists, the Automation Platforms and the Open-Source Route

Beyond the seven ranked agents, a long tail of products puts agents into Slack for one function or one platform. Several are excellent at their job; they are not ranked above because they are either tied to a platform you must already own or because they do not publish enough to score their real cost.

Atlassian Rovo is the strongest of these for teams on Jira and Confluence. It is included with Standard, Premium and Enterprise cloud plans, with 25, 70 or 150 credits per user per month for Jira and Confluence, and overage at "$0.01 per credit over your allowance ($10.00 per 1,000 credits)" - Atlassian. In Slack it answers DMs and mentions, can be triggered by keywords, emojis or every new message, and can turn a discussion into Jira issues or Confluence pages. Its access model is worth copying: "By default, Rovo replies privately to you", and "Public replies use Rovo's own access, never yours" - Atlassian Support. Atlassian also published one of the few concrete Slack results we found, from its customer SpotOn: "Rovo answered 95% of questions in the channel, and our clean deflection rate increased from 40% to 53%" - Atlassian blog.

Glean brings enterprise search into Slack. "Glean in Slack" is included in its per-user seat fee, model usage beyond that is billed at provider API rates, and its seat price itself is not published - Glean. One design limit matters for channel planning: "Each Slack channel can have only one agent at a time" - Glean docs.

Lindy sells a "Slack-native teammate in threads + mentions" at $29.99 per user per month with 3,000 credits per seat, and anyone who mentions Lindy in Slack takes a billed seat after a seven-day trial - Lindy. Runbear takes the opposite approach, with a $99 Team plan including 20,000 credits and "No seat tax" - Runbear. Zapier Agents and n8n suit teams that think in workflows rather than coworkers: Zapier's Agents Pro costs $33.33 a month on annual billing for 1,500 activities - Zapier, and n8n's cloud plans start at €20 a month for 2,500 executions, with a self-hosted community edition - n8n. Our comparison of n8n, Zapier and Make for agents covers the trade-offs. In IT service, Moveworks now belongs to ServiceNow, which closed the acquisition on December 15, 2025 and counts "approximately 5.5 million employee users" - Moveworks.

Building your own

The open-source route is real, and it has become cheaper to run as volume models fall in price. OpenClaw, the self-hosted agent with one of the largest followings on GitHub, connects to Slack through Socket Mode or HTTP - OpenClaw docs, and Nous Research's Hermes Agent reaches Slack along with Telegram, Discord, WhatsApp and Signal "from a single gateway process", with the suggestion to "Run it on a $5 VPS" - GitHub. For teams that want to write their own, Slack's Bolt framework includes an AI app sample that "supports integration with the Claude Agent SDK and OpenAI Agents SDK" - Slack.

Two constraints from section 2 shape that choice. A home-built agent installed as an internal app keeps generous API limits and can use Slack's MCP server; the same agent sold to other companies as an unlisted app cannot. And the software is free but the safety is yours to build: our comparison of Hermes Agent and OpenClaw and the guide to building an agent on seven paths explain what that costs in practice.

Finally, a Slack agent is the right tool when the work is a stream of requests from a team. When the goal is instead to hand off a whole business function (the website, the billing, the content, the back office) rather than individual tasks, the category changes: platforms such as O-mega build and run an autonomous company from one conversation, which is a different purchase with a different cost model from a coworker in a channel.

14. Security: What a Slack Agent Can Leak, and How Each Vendor Contains It

Slack is a harder environment for agents than it looks, for a structural reason. An agent in a channel reads text written by many people, some of whom it should not obey; it holds credentials to other systems; and it writes its output into an app that fetches and previews links automatically. Each of those is an attack surface on its own. Together they make the classic agent risk, indirect prompt injection, unusually easy to attempt.

Planted instructions are the first risk. In 2024, PromptArmor showed that a message planted in a public Slack channel could cause Slack AI to leak secrets from private channels through a crafted link, and reported that Slack's first answer was "This is intended behavior" - PromptArmor. Vendors now document the risk directly. Anthropic's Claude Code in Slack docs warn that "Claude may follow directions from other messages in the context, so users should make sure to only use Claude in trusted Slack conversations" - Anthropic. Claude Tag treats other people's posts in a thread that uses someone's personal connectors "as information for the task rather than as instructions" - Anthropic security docs. OpenAI's dots documentation makes the same point about email.

Link previews turn a reply into an exfiltration channel. In February 2026, The Register reported PromptArmor research showing that an agent can be tricked into writing an attacker's URL with secrets in it, which the messaging app's link preview then fetches, so "data exfiltration can occur immediately upon the AI agent responding" - The Register. Slack was named among the apps with previews enabled by default or in some configurations, a Slack setup using Cursor's Slack bot was listed as at risk, and the Claude app in Slack as a safer configuration. No click is needed; the preview does the leaking.

Shared connections widen every request. When an admin connects a tool to a channel agent, everyone in that channel can ask the agent to use it. Anthropic states this bluntly for Claude Tag: "whatever the connected account can read or write is available to every member of those channels" - Anthropic security docs. Rovo's answer is to reply privately by default and to use "Rovo's own access, never yours" in public. Glean's public reply option restricts answers to "only broadly shared content" - Glean docs. The general rule: the least-privileged account in a channel should define what the agent can touch there.

Tokens are a supply-chain risk. In 2025, the breach of the AI chatbot vendor Salesloft Drift exposed tokens for services "including Slack, Google Workspace, Amazon S3, Microsoft Azure, and OpenAI", and Salesforce blocked Drift from integrating with Slack - Krebs on Security. Every agent you install holds a token like that.

Platform rules help. Slack requires Marketplace AI apps to follow "a zero copy and zero LLM training policy", and "Guests can't use AI apps or agents" - Slack Help Center. ChatGPT's Slack flow does not support Slack Connect channels shared with another organization, and OpenAI advises against using dots in them.

The vendors that score highest on control in the ranking share four habits, and you can demand all four from any product you trial. They give the agent its own identity in each tool (Claude Tag's service accounts, Gemini's attested agent identities) so its actions appear in audit logs under its name. They keep credentials out of the model's sandbox (Claude Tag says the credentials you provision "aren't placed in that sandbox" - Anthropic). They default to read-only for background work (dots' proactive research is read-only). And they require approval for writes that leave the company (workspace agents by default, Gemini's "Ask for approval" setting, PagerDuty's "if approved"). Our guides to prompt injection defense and non-human identity go deeper on both controls.

15. Where Slack Agents Work, and Where They Fail

The evidence on where Slack agents earn their cost is still thin and mostly vendor-published, but it points in a consistent direction: agents succeed where a channel already has a repeated request shape and a clear definition of done.

Intake and triage channels are the clearest wins. SpotOn's sales-support channel, where Rovo answered 95% of questions and lifted clean deflection from 40% to 53%, is the archetype: many similar questions, a known body of answers, and a human who only needs to see the leftovers. Knowledge lookup is the second: Clay's report of about 58 hours saved a month with a Dust agent in one channel is the same pattern applied to internal know-how. Recurring briefs are the third, as in Rippling's deal briefs posted into a sales room. Incidents are the fourth, where Datadog's numbers (about $4 and about 2 minutes per investigation) show what a narrow, well-instrumented job looks like. Engineering is the most ambitious case, and Anthropic's claim that its internal version of Claude Tag creates 65% of its product team's code shows where the strongest teams are heading, with the caveat that it is the vendor describing itself.

The failures follow the same logic in reverse. Agents struggle where ownership is ambiguous: when anyone in a channel can hand the agent work, nobody may feel responsible for checking it, and an always-answering agent like Dust's auto-reply mode can turn a conversation channel into noise. They struggle where context is missing: in a 2025 report on Cursor's forum, a user found the agent received "only the 'please fix' and nothing else from the Slack thread" - Cursor forum, a reminder to test what each agent actually reads. They struggle against limits: 15 Slackbot messages a week is a ceiling a heavy user hits by Wednesday. And they create cost surprises where proactive features run unattended, which is why Claude Tag splits proactive and scheduled spend from engaged spend.

The wider market data is sobering. Menlo Ventures found in December 2025 that "Only 16% of enterprise and 27% of startup deployments qualify as true agents" - Menlo Ventures, meaning most "agent" projects were still assistants or fixed workflows. A Slack agent is a good way to cross that line cheaply, precisely because it can start with one channel and one job.

16. How to Roll Out a Slack Agent Without Overpaying

The vendors' own advice converges on a pilot, and Anthropic's is the most direct: run one "with a spend limit set and watch the per-channel breakdown" - Anthropic. That generalizes into a rollout that works for any of the seven agents.

1. Pick one channel and one job. Choose a channel with a repeated request shape (an intake queue, a bug channel, a sales-support channel) and write down what "done" means for a task there.
2. Choose the shape before the vendor. If the job is personal (briefs for one person, a personal inbox), a personal agent such as Slackbot or a dot fits. If the job comes from many people, buy a channel agent, and prefer one that does not charge a seat for everyone in the channel.
3. Set the budget in the product, not in a spreadsheet. Use spend limits, per-channel limits and alerts where they exist (Claude Tag, Gemini, Agentforce through Flex Credits), and pick the model per channel.
4. Give the agent its own accounts, read-only first. Create a service identity, connect the narrowest role each tool offers, and add write access only after the agent has earned it in that channel.
5. Write channel instructions. Tell the agent what it should and should not pick up, where to file what it cannot finish, and how to show its sources.
6. Measure four numbers for 30 days: tasks finished, review minutes per task, spend per channel, and tasks redone. The cost per finished, accepted task is the only number that compares products fairly.
7. Expand by channel, not by company. Add the next channel only when the first one is cheaper per accepted task than doing the work by hand.

Two practical notes keep pilots honest. First, run the pilot after any launch promotion ends, or at least read the list-price column: Claude Tag's analytics show list-price spend even when a promotional credit covers it, and Perplexity's promotion makes teammate usage free for now. Second, include the people who review the agent's work in the pilot from day one; review time is the cost that decides the outcome, and it is invisible in every vendor dashboard.

17. What Comes Next for Agents in Slack

Four shifts are already visible in the product changelogs and announcements, and each changes how to buy.

Slack is becoming an agent platform with its own rules. Slack has announced that its assistant_view will be "deprecated in February 2027" in favor of an Agent Sessions API, added an Agents and Tools tab, and enabled multiple agents to stack in one conversation - Slack changelog. It also tightened the Marketplace, which now requires at least ten active workspace installs for a submission - Slack Marketplace changelog. Expect the agents with Marketplace status and deep Slack integration to pull further ahead of everything else.

Slackbot will compete with the labs inside Slack. Once "Slackbot in Channels" leaves its pilot, the default agent in every Business+ workspace will take team requests, not just personal ones. That will put pressure on seat-priced third-party agents first, because Slackbot will already be paid for.

Enterprise defaults are flipping on. ChatGPT dots become available by default in Enterprise workspaces from November 18 unless admins opt out, Perplexity already tells Enterprise users to ask their admin for Teammates access, and Google is pitching coworker agents with their own accounts. Admins should decide channel policy for agents before their users decide it for them.

Token prices fall while work per task rises. Fable 5.2 launched with a 20% price cut from Fable 5.1's $10 and $50 per million tokens - Implicator, and Meta's Muse Ember prices capable volume work at ten cents per million input tokens. Yet Google's own framing is the right one: "While per-token prices have dropped 98% since 2024, enterprise AI volume has exploded" - Google Cloud. Agents that work for hours spend more tokens per task than chatbots ever did. Cheaper models will not shrink Slack agent bills on their own; routing each channel to the right model will. The Cursor episode adds a second lesson: prefer agents that let you change the model (Claude Tag per scope, Perplexity per teammate, Gemini and Copilot Fusion by routing, Dust across 20 or more), because model access can change for reasons that have nothing to do with you. Our November ranking of models for agents tracks which ones are worth routing to.

18. Conclusion: Which Slack Agent to Choose

There is no single best AI agent for Slack, but there is a best first move for most teams, and the ranking makes it fairly clear.

If you are already on Slack Business+ or Enterprise+, switch on Slackbot first. It costs nothing extra within its limits, it sees everything you can see, and the tasks it cannot do will tell you exactly what a second agent needs to handle.

If you want one shared coworker that a whole channel can hand real work to, Claude Tag is the strongest product today and the only one priced the way channels work, with no seat for every member. Start with a spend limit you are comfortable losing, a strong model in engineering channels and a cheaper one elsewhere, and read-only connections.

If your company runs on ChatGPT Business or Enterprise, use workspace agents and @ChatGPT in Slack rather than adding another vendor, and budget the credits from day one. If your knowledge lives in Notion, Notion Custom Agents are the cheapest meter in this guide for answer-and-triage channels. If you are a Google Workspace company, Gemini's coworker agents bring the best governance model, at the cost of being five weeks old. If you want to build many specialized agents across models, Dust gives each channel its own default. And if you want to test an always-on agent that owns ongoing work, Perplexity Teammates are the most interesting new entrant, best tried in one channel by one owner until an Enterprise version arrives.

Whichever you choose, remember what "real cost" means. The plan price is the smallest number in the bill, the meter and the model are the next, and the largest is the time your people spend checking the work. The best AI agent for your Slack is the one whose finished work your team accepts fastest, and the only way to know which one that is, is to measure it in one channel before you buy it for all of them.

What this article rests on

Every model figure above is read from its record, re-read from OpenRouter's model feed every 6 hours, so the figures on this page are current even where the prose around them is older. Prices are OpenRouter's listed price per million tokens: for a lab's closed models that is the lab's own list price, and for an open-weight model it is the price of the host OpenRouter routes to by default, which can differ from the lab's own API and from other hosts.

• Claude Fable 5.2 (Anthropic): $8 in and $40 out per million tokens, cached input $0.20, a 1M token window, listed 9 November 2026.
• Claude Opus 5.5 (Anthropic): $4 in and $20 out per million tokens, cached input $0.20, a 1M token window, listed 22 September 2026.
• Claude Sonnet 5.5 (Anthropic): $2 in and $10 out per million tokens, cached input $0.10, a 1M token window, listed 28 September 2026.
• Muse Ember (Meta): $0.10 in and $0.50 out per million tokens, cached input $0.004, a 1M token window, listed 9 November 2026.

In this topic: Agents at work

• Best Personal AI Agent 2026: Claude Duet vs ChatGPT Dots
• Google Gemini Agent: AI Coworkers Get Their Own Inbox
• Top 10 Manus Alternatives: Pricing & Verdicts (Oct 2026)
• AI Agents for Small Business: 10 Jobs to Hand Off (2026)
• AI Employees 2026: Copilot Autopilot vs Dots vs Agentforce
• Best Enterprise AI Agent Platforms 2026, Ranked
• Salesforce Agentforce: 7 Named AI Agents in 2026
• Hermes Agent vs OpenClaw 2026: Security, Cost, Verdict

Which AI agents can do work for your business? Every guide on it