Back to articles

AI models

Gemini 3.5 Flash: Benchmarks, Cost, and Guide

Gemini 3.5 Flash beats flagship AI models on agent tasks at 10x speed and one-third the cost. Explore benchmarks, pricing, and usage guide.

Gemini 3.5 FlashAI BenchmarksGoogle AILLM GuideAI Cost
O-mega Team
O-mega Team
••

49 min read

Since this was written, OpenRouter's model feed lists Gemini 4 Hydrogen in place of Gemini 2.5 Flash, Gemini 3.5 Flash, Gemini 3.6 Flash, Gemini 3.7 Flash and Gemini 3.8 Flash; DeepSeek V4 Pro 0813 in place of DeepSeek V4-Pro; DeepSeek V4.1 Flash 1017 in place of DeepSeek V4.1 Flash; GPT-5.4 Nano in place of GPT-4.1 Nano; 2 more models too.

The complete technical breakdown of Google's agent-tier workhorse: every benchmark score, API price, speed metric, and head-to-head comparison with Claude Opus 4.7, GPT-5.5, and every major competitor.

Gemini 3.5 Flash outputs tokens at 289 tokens per second, roughly 10x faster than Claude Opus 4.6, at one-third the price. That single data point captures why this model matters. Announced at Google I/O 2026 on May 19, it is not the most capable AI in existence (that title belongs to the flagships). But it was the first speed-tier model to surpass its own company's flagship on the benchmarks that matter most for real work: coding, agentic task completion, and tool use - Google DeepMind. Google has since shipped three more models on the same tier and a new flagship above it, and on 8 October 2026 it deprecated 3.5 Flash, routing every API request for gemini-3.5-flash to Gemini 3.6 Flash and dropping it from its model list, so the live question for anyone deploying today is which Flash, not whether Flash.

This was not a minor iteration. Gemini 3.5 Flash surpasses Gemini 3.1 Pro on Terminal-Bench 2.1 (coding), GDPval-AA (real-world agentic evaluation), MCP Atlas (scaled tool use), and Finance Agent v2. At launch it led the entire field (including Claude Opus 4.7 and GPT-5.5) on MCP Atlas, Toolathlon, Finance Agent v2, CharXiv Reasoning, and MMMU-Pro. And it does this while costing $1.50 per million input tokens, compared to $5.00 for Claude Opus 4.6 and $5.00 for GPT-5.5. Those two prices have not moved, although neither model is its lab's flagship any longer (Claude Opus 5.5 lists at $4 in and $20 out). What moved is the Flash tier itself, where Gemini 3.8 Flash now lists at $0.75 in and $3.75 out.

The catch? It still trails on pure reasoning benchmarks. Gemini 3.1 Pro beats it on ARC-AGI-2 (by 5 points), Humanity's Last Exam (by 4.2 points), and long-context retrieval. The flagships from Anthropic and OpenAI remain better at the hardest intellectual tasks. But if your use case is "an AI agent that needs to get things done fast and cheaply," Gemini 3.5 Flash is the model that made the Flash tier the default choice.

This guide provides the complete technical picture: every benchmark score, every pricing detail, speed comparisons, and head-to-head analysis against every major competitor. If you are choosing a model for production deployment, building agents, or working out where 3.5 Flash now sits against its own successors, this is the reference. Every price and benchmark figure below was re-verified against provider documentation on 30 October 2026.

The release timing is not accidental. Google launched 3.5 Flash the same day it announced Gemini Spark (a 24/7 personal AI agent), Information Agents in Search (background web monitoring), and Universal Cart (intelligent shopping). All of these products need a fast, cheap, capable model to run at scale. Flash is the infrastructure that makes Google's entire agentic product lineup viable. Understanding Flash's capabilities and limitations is therefore essential for understanding where the entire AI industry is heading.

Contents

  1. What Gemini 3.5 Flash Is (and What It Is Not)
  2. Full Technical Specifications
  3. The Master Benchmark Table: Every Model Compared
  4. The Master Pricing Table: Every Model Compared
  5. Speed Comparison: Tokens Per Second
  6. Where 3.5 Flash Leads the Field
  7. Where 3.5 Flash Falls Short
  8. Gemini 3.5 Flash vs. Previous Gemini Models
  9. Gemini 3.5 Flash vs. Claude Opus 4.6 and 4.7
  10. Gemini 3.5 Flash vs. GPT-5.5 and GPT-4.1
  11. Gemini 3.5 Flash vs. Open-Weight Models
  12. Cost Analysis: What Agents Actually Cost to Run
  13. The Thinking Mode: Dynamic Reasoning Control
  14. When to Use Gemini 3.5 Flash (and When Not To)
  15. Gemini 3.5 Pro: What We Know
  16. First-Principles Analysis: Why Speed Beats Capability
  17. Conclusion

1. What Gemini 3.5 Flash Is (and What It Is Not)

Gemini 3.5 Flash is the first model in Google's 3.5 generation, designed specifically for the agent era. Google's framing is deliberate: this is not a general-purpose model that happens to be fast. It is an agentic model that was optimized from the ground up for the combination of intelligence, speed, and cost efficiency that makes AI agents viable at scale - Google Blog.

The strategic context matters for understanding why Google released Flash before Pro. In previous generations, Google released the flagship (Pro) first and the speed-optimized model (Flash) later. This time, Flash came first. The reason is that Google's entire I/O 2026 product lineup (Gemini Spark, Information Agents, agentic booking in Search, Universal Cart) runs on Flash-class models. Agents make dozens of model calls per user request. If each call is slow or expensive, the agent is impractical. By optimizing Flash for agent economics first, Google ensures its agentic products work at scale from day one.

This is not the model you choose for solving olympiad math problems or writing the next great novel. It is the model you choose when you need an AI system that can reason about a task, call tools, evaluate results, and take actions, all in under a second and at a cost that scales to millions of users. For the full progression of Google's model family and how earlier generations laid the groundwork for 3.5 Flash, see our Gemini 3.1 Pro guide and Flash Lite guide.

The Naming Convention Explained

Google's model naming can be confusing, so here is how to decode it. The first number (3.5) indicates the generation. The word after (Flash or Pro) indicates the tier: Flash is speed-optimized, Pro is capability-optimized. Within a generation, Flash is always cheaper and faster; Pro is always more capable on hard tasks. Across generations, each new Flash model closes the gap with (or surpasses) the previous Pro model. Gemini 3.5 Flash surpassing Gemini 3.1 Pro on agentic benchmarks is a milestone in this pattern.

The previous generation (Gemini 3.0/3.1) established the pattern: Gemini 3 Flash already beat Gemini 2.5 Pro on 18 of 20 benchmarks while being faster and cheaper. Gemini 3.5 Flash continues this trajectory, raising the bar for what "speed tier" means in the industry. A "speed" model that beats flagships on real-world agent tasks is no longer just a compromise option. It is the primary model for the majority of production workloads.


2. Full Technical Specifications

Every specification that matters for production deployment, sourced from Google's official documentation and the model evaluation report - Google AI Developer Docs.

SpecificationValue
API model namegemini-3.5-flash
Input context window1,048,576 tokens (1M)
Maximum output tokens65,536
Output speed~289 tokens/sec (May 2026)
Input pricing$1.50 / 1M tokens (list price until deprecation)
Output pricing$9.00 / 1M tokens (list price until deprecation)
Cached input pricing$0.15 / 1M tokens (90% discount)
Input modalitiesText, image, audio, video
Output modalitiesText
Knowledge cutoffJanuary 2026
Thinking modeDynamic thinking (on by default), with configurable thinking levels
CapabilitiesFunction calling, structured output, search-as-a-tool, code execution
SafetyGoogle Frontier Safety Framework. Strengthened cyber and CBRN safeguards. Interpretability tools for reasoning verification
Artificial Analysis Intelligence Index33 on the earlier index; not scored on the v5.0 index Artificial Analysis now runs, where the model is marked deprecated
AvailabilityAt launch: Gemini app, Google AI Studio, Antigravity 2.0, Gemini API, AI Mode in Google Search. No longer selectable in Google AI Studio
Lifecycle statusDeprecated on 8 October 2026: the endpoint is not shutting down, but API requests to gemini-3.5-flash are routed to Gemini 3.6 Flash

The 1M token context window is the same as the previous Flash generation and as Gemini 3.1 Pro's window. For most agent use cases, 1M tokens is more than sufficient (it is roughly 750,000 words). The 65K max output is generous and supports generating complete code files, long documents, and detailed analysis in a single call.

The cached input pricing at $0.15/1M tokens (90% cheaper than standard input pricing) deserves special attention. For agent systems that repeatedly reference the same context (system prompts, tool definitions, reference documents), caching dramatically reduces costs. A system prompt that costs $1.50 per call at standard pricing costs $0.15 with caching. Over millions of agent interactions, this is the difference between viable and unviable economics. We analyzed these economics in depth in our guide to the true cost of LLM inference.

Safety and Interpretability

Google developed 3.5 Flash under its Frontier Safety Framework, with strengthened safeguards against cyber and CBRN (chemical, biological, radiological, nuclear) misuse. The model includes advanced reasoning checks that evaluate the safety of responses before generation, which contributes to what Google describes as being "less likely to generate harmful content or incorrectly refuse valid queries." This dual improvement (fewer harmful outputs AND fewer false refusals) is practically important for agent systems where incorrect refusals break automated workflows.

The model also ships with interpretability tools that allow developers to inspect the model's inner reasoning process. For production deployments where you need to understand why an agent took a specific action (compliance requirements, debugging unexpected behavior, auditing decision-making), interpretability is not a research curiosity. It is an operational requirement.

Modality Support

Gemini 3.5 Flash accepts text, image, audio, and video as inputs but produces only text as output. For agents that need to generate images or audio, you would pair Flash with specialized models (Gemini Omni for video, Nano Banana for images, Lyria 3 Pro for music). The multimodal input capability is where Flash excels: its MMMU-Pro score (84.0%, highest recorded at launch) means it understands images, charts, screenshots, and visual content at a level that matched or exceeded every other model then available. For our coverage of Google's image generation model, see the Nano Banana 2 guide.


3. The Master Benchmark Table: Every Model Compared

This is the comprehensive comparison table. Every number is sourced from official model cards, evaluation reports, or Artificial Analysis. Where a score is not publicly available, it is marked with a dash. Scores are percentages unless noted otherwise.

BenchmarkGemini 3.5 FlashGemini 3.1 ProClaude Opus 4.6Claude Opus 4.7Claude Sonnet 4.6GPT-5.5GPT-4.1Llama 4 MaverickDeepSeek V4-ProGrok 4.3
Terminal-Bench 2.1 (coding)76.270.3---82.7 (v2.0)----
GDPval-AA (agentic, Elo)1,6561,314---~1,769----
MCP Atlas (tool use)83.678.2< 83.6< 83.6-< 83.6----
Finance Agent v257.943.0< 57.9< 57.9-< 57.9----
MMMU-Pro (multimodal)84.0-< 84.0< 84.0-< 84.0----
CharXiv Reasoning84.2-< 84.2< 84.2-< 84.2----
SWE-Bench Verified--80.887.679.688.7--80.6-
SWE-Bench Pro55.1---------
GPQA Diamond~90.4-91.3-74.1--69.890.1-
MMLU-Pro78.3------80.587.5-
ARC-AGI-272.177.1---85.0----
Humanity's Last Exam40.244.4--------
MRCR v2 (128k)77.384.9--------
MMMU82.5---------
HumanEval~92---------

Sources: Google DeepMind Model Evaluation, Anthropic Claude 4, OpenAI GPT-5.5, Llama 4, DeepSeek V4, Artificial Analysis

The table reveals three critical insights. First, Gemini 3.5 Flash leads the entire field on agentic benchmarks (MCP Atlas, Finance Agent v2, CharXiv, MMMU-Pro), beating even the expensive flagships. Second, the flagships (Claude Opus 4.7, GPT-5.5) still lead on traditional software engineering benchmarks (SWE-Bench Verified) and pure reasoning (ARC-AGI-2). Third, there are significant gaps in publicly available benchmark data across competitors, which makes direct comparison difficult for some categories. For a broader analysis of model benchmarks across the full AI landscape, see our AI model benchmarks and pricing guide.

How to Read This Table

A few important caveats about benchmark comparisons. Different models are evaluated on different benchmark suites, which is why so many cells are dashes. Google's evaluation of 3.5 Flash prioritizes agentic benchmarks (Terminal-Bench, GDPval-AA, MCP Atlas, Finance Agent v2) over traditional academic benchmarks (MMLU, MATH). Anthropic's evaluations prioritize SWE-Bench and GPQA. OpenAI publishes broad benchmark suites but often uses different variants or scoring methods. This means cross-provider comparisons on the same benchmark are the most reliable, while assuming equivalence across different benchmarks is risky.

The benchmark that Google deliberately chose to highlight (MCP Atlas for tool use, Finance Agent for financial tasks, MMMU-Pro for multimodal) are the benchmarks where Flash excels. This is not deception. It is strategic communication. Every provider highlights their model's strengths. The responsible approach for businesses is to evaluate models on the benchmarks most relevant to their specific use case, not the ones the provider advertises. If you are building a financial AI agent, Finance Agent v2 is the relevant benchmark and Flash leads. If you are building a code refactoring tool, SWE-Bench is relevant and Claude Opus 4.7 leads.

The full model evaluation report is publicly available at Google DeepMind's website - Gemini 3.5 Flash Model Evaluation PDF. This 30+ page document provides detailed methodology, per-benchmark breakdowns, and comparison data that goes beyond what Google published in blog posts. For serious production evaluation, read the full report.


4. The Master Pricing Table: Every Model Compared

This table covers every major model available via API, verified against provider pricing pages on 30 October 2026. Prices are per million tokens - Gemini API Pricing, Anthropic Pricing, OpenAI Pricing.

ModelProviderInput $/1MOutput $/1MCached Input $/1MContext WindowSpeed (tok/s)
Gemini 3.5 Flash (deprecated)Google$1.50$9.00$0.151M~289
Gemini 4 ArgonGoogle$2$10$0.101M-
Gemini 3.8 FlashGoogle$0.75$3.75$0.0751M~281
Gemini 3.7 Flash (deprecated)Google$0.75$3.75$0.0751M-
Gemini 3.1 ProGoogle$2.00$12.00$0.201M~122
Gemini 3 FlashGoogle$0.50$3.00$0.051M-
Claude Opus 5.5Anthropic$4$20$0.201M-
Claude Opus 5Anthropic$5.00$25.00$0.501M-
Claude Opus 4.6Anthropic$5.00$25.00$0.501M~25
Claude Opus 4.7Anthropic$5.00$25.00$0.501M-
Claude Sonnet 5.5Anthropic$2$10$0.101M-
Claude Sonnet 5Anthropic$2.00$10.00$0.201M-
Claude Sonnet 4.6Anthropic$3.00$15.00$0.301M~65
Claude Haiku 5.5 (prompts up to 100K)Anthropic$0.10$0.50$0.011M-
Claude Haiku 4.5Anthropic$1.00$5.00$0.10200K~93
GPT-6 AstraOpenAI$10$50$1--
GPT-6.1 SolOpenAI$2$10$0.101.05M-
GPT-6.1 TerraOpenAI$1$5$0.051.05M-
GPT-6 LunaOpenAI$0.10$0.50$0.011.05M-
GPT-5.5OpenAI$5$30$0.50-~63
GPT-5.5 ProOpenAI$30$180---
GPT-4.1OpenAI$2$8$0.501.05M-
GPT-4.1 NanoOpenAI$0.10$0.40$0.0251.05M-
Grok 4.7xAI$2$6$0.50500K-
Grok 4.6xAI$2.00$6.00$0.50500K-
Grok 4.3 (high)xAI$1.25$2.50$0.201M~98
DeepSeek V4-ProDeepSeek$1.32$3.96$0.0441M-
DeepSeek V4.1 FlashDeepSeek$0.30$1.20$0.0061M-
Llama 4 MaverickMetaOpen-weightOpen-weight---
Mistral Large 3Mistral$0.50$1.50-262K-

The pricing reveals the core trade-off. Gemini 3.5 Flash at $1.50/$9.00 is 3.3x cheaper on input and 2.8x cheaper on output than Claude Opus 4.6 ($5/$25). Compared to GPT-5.5, it is 3.3x cheaper on input and 3.3x cheaper on output. But it is 3x more expensive than its predecessor Gemini 3 Flash ($0.50/$3.00), reflecting the substantial capability improvement. For budget-constrained deployments, Claude Haiku 5.5 ($0.10/$0.50 for prompts up to 100,000 tokens), Grok 4.3 ($1.25/$2.50), and DeepSeek V4.1 Flash ( $0.30/$1.20 at peak hours, half that off-peak) offer even lower list prices, and Haiku 5.5 now undercuts every Gemini Flash model on both input and output.

The awkward row in that table is 3.5 Flash's own. Google left its price at $1.50/$9.00 while listing every later Flash model at $0.75/$3.75, an introductory rate that runs through 31 December 2026 before stepping up to $1.50/$7.50, so the legacy model cost double the models that replaced it, and its cached input at $0.15/1M was double their $0.075. Google then settled the question itself: since 8 October 2026, 3.5 Flash is deprecated, gone from the pricing page, and every API request for it is served by Gemini 3.6 Flash. A deployment still calling gemini-3.5-flash is already running on its successor, and the remaining migration step is changing the model string.


5. Speed Comparison: Tokens Per Second

Speed is where Gemini 3.5 Flash dominated the field at launch, with no close competitor among frontier models - Artificial Analysis.

At ~289 tokens per second, Gemini 3.5 Flash is:

  • 11.6x faster than Claude Opus 4.6 (~25 tok/s)
  • 4.6x faster than GPT-5.5 (~63 tok/s)
  • 4.4x faster than Claude Sonnet 4.6 (~65 tok/s)
  • 3.1x faster than Claude Haiku 4.5 (~93 tok/s)
  • 2.9x faster than Grok 4.3 (~98 tok/s)
  • 2.4x faster than Gemini 3.1 Pro (~122 tok/s)

For agent systems, speed is not just about user experience. It directly affects the total time to complete a multi-step task. An agent that makes 15 model calls to complete a task takes 15 seconds at Gemini 3.5 Flash speed (assuming 1-second calls at 289 tok/s). The same 15 calls take 150 seconds with Claude Opus 4.6. That is the difference between a responsive, usable agent and one that makes users wait minutes for results.

Speed also affects time-to-first-token (TTFT), which is the latency before the model starts generating output. For interactive applications where users are waiting for a response, TTFT matters more than throughput. GPT-5.5, for instance, has a measured TTFT of 85.57 seconds on its most capable mode (xhigh), which means users wait over a minute before seeing any output begin. Flash's TTFT is a fraction of this, making it suitable for real-time interactive experiences.

The compounding effect of speed across agent workflows deserves emphasis. A Gemini Spark agent that checks your email, identifies action items, drafts responses, and updates your task list might make 8-12 model calls for a single morning brief. At Flash speed, this entire workflow completes in under 10 seconds. At Opus speed, it takes over a minute. When this workflow runs for millions of Spark users every morning, the aggregate compute savings from Flash's speed translate directly to the ability to serve the service at scale without infrastructure bottlenecks.


6. Where 3.5 Flash Leads the Field

Gemini 3.5 Flash does not just beat its predecessors. At launch it beat every model then available, including models that cost 3-10x more, on specific benchmark categories. Understanding where it leads (and why) is critical for choosing the right model for your use case.

Agentic Task Completion

On GDPval-AA (a benchmark that evaluates real-world agentic performance with Elo ratings), Gemini 3.5 Flash scores 1,656 Elo, compared to 1,314 for Gemini 3.1 Pro. That is a 342 Elo gap, which in competitive rating systems represents a massive capability difference. The only model that scores higher is GPT-5.5 at approximately 1,769 Elo, but GPT-5.5 costs 3.3x more on input and 3.3x more on output.

Tool Use at Scale

On MCP Atlas (which evaluates how well models use tools at scale), Gemini 3.5 Flash scores 83.6%, beating Gemini 3.1 Pro (78.2%) and leading the entire field at launch. This benchmark is particularly relevant because AI agents interact with the real world through tool use: calling APIs, querying databases, executing code, browsing the web. A model that is excellent at tool use is a model that makes excellent agents. The benchmark name (MCP Atlas) reflects the growing importance of the Model Context Protocol standard that we covered in our guide to the 2026 MCP specification.

Financial Agent Tasks

On Finance Agent v2, Gemini 3.5 Flash scores 57.9% compared to Gemini 3.1 Pro's 43.0%. This is a 14.9 percentage point improvement and represents the largest single-benchmark gap between Flash and Pro. Financial agent tasks require numerical reasoning, data extraction from financial documents, and multi-step analysis, exactly the kind of structured work that agents are deployed for in enterprise settings.

Multimodal Understanding

On MMMU-Pro (multimodal understanding), Gemini 3.5 Flash scores 84.0%, which Google reports as the highest recorded score. On CharXiv Reasoning (reasoning about scientific charts and figures), it scores 84.2%, also leading the field at launch. These multimodal capabilities matter for agents that need to process images, screenshots, charts, and other visual content as part of their workflow.

For businesses deploying AI agents, these benchmarks map directly to real capabilities. An agent that scores well on MCP Atlas will use your APIs correctly. An agent that scores well on Finance Agent v2 will process your financial data accurately. An agent that scores well on CharXiv will correctly interpret the charts in your reports. Our analysis of the cost of agentic AI explored the relationship between model capability and agent deployment economics.

Why Agentic Benchmarks Are the New Standard

There is a structural reason why Google chose to evaluate Flash primarily on agentic benchmarks rather than the traditional academic benchmarks that dominated previous model releases. Traditional benchmarks (MMLU, MATH, HumanEval) measure isolated capabilities: knowledge recall, mathematical reasoning, code generation. They test the model in a vacuum. Agentic benchmarks (MCP Atlas, GDPval-AA, Finance Agent) measure the model's ability to accomplish real tasks end-to-end, including planning, tool selection, error recovery, and multi-step execution.

The difference matters because a model that scores 95% on MMLU but cannot reliably use a tool (because tool use requires different capabilities: following complex instructions, managing state across calls, interpreting tool outputs, recovering from errors) is useless for agent deployments. Conversely, a model that scores 85% on MMLU but leads on MCP Atlas is a better agent model, even though it "scores lower" on the traditional benchmark. Google's choice to prioritize agentic benchmarks reflects the reality that models are increasingly evaluated by what they can accomplish, not what they know.

This shift in evaluation methodology has implications beyond Google. Expect Anthropic, OpenAI, and other providers to increasingly report agentic benchmark scores rather than (or in addition to) traditional academic benchmarks. The benchmark landscape is evolving to match how models are actually used in production. For businesses evaluating models, this means asking "how well does it perform on tasks similar to mine?" rather than "what is its MMLU score?"


7. Where 3.5 Flash Falls Short

Honesty about limitations is more useful than cheerleading. Gemini 3.5 Flash has clear weaknesses relative to both its predecessor (3.1 Pro) and the expensive flagships.

Pure Reasoning

On ARC-AGI-2 (abstract reasoning), Gemini 3.5 Flash scores 72.1% compared to Gemini 3.1 Pro's 77.1% and GPT-5.5's 85.0%. This is a 5-point gap against its own predecessor and a 13-point gap against the best model. For tasks that require deep abstract reasoning (novel problem-solving, mathematical proof construction, complex logical deduction), 3.5 Flash is not the right choice.

Frontier Knowledge Tests

On Humanity's Last Exam (HLE), designed to test the frontier of AI knowledge, Gemini 3.5 Flash scores 40.2% compared to 3.1 Pro's 44.4%. A 4.2-point deficit on a test designed to push the limits of AI capability. For research applications that need the absolute best reasoning, 3.1 Pro or the competitor flagships remain superior.

Long-Context Retrieval

On MRCR v2 (128k) (long-context multi-needle retrieval), Gemini 3.5 Flash scores 77.3% compared to 3.1 Pro's 84.9%. A 7.6-point gap. For applications that need to find specific information buried deep in long documents (legal discovery, compliance review, academic research across large corpuses), 3.1 Pro's better retrieval accuracy across the same 1M window makes it the better choice.

SWE-Bench Software Engineering

On SWE-Bench Verified (the industry-standard benchmark for real-world software engineering), Claude Opus 4.7 scores 87.6% and GPT-5.5 scores 88.7%. Gemini 3.5 Flash's SWE-Bench Pro score of 55.1% uses a different variant, but the general pattern is clear: for complex software engineering tasks (multi-file refactors, debugging production codebases), the expensive flagships remain substantially better. Our guide to Claude Opus 4.7 covers Anthropic's flagship capabilities in detail.


8. Gemini 3.5 Flash vs. Previous Gemini Models

Understanding where 3.5 Flash fits in the Gemini lineage helps you decide whether to upgrade from older models.

MetricGemini 3.5 FlashGemini 3.1 ProGemini 3 FlashGemini 2.5 Pro
Input $/1M$1.50$2.00$0.50Legacy
Output $/1M$9.00$12.00$3.00Legacy
Context window1M1M1M1M (2M planned)
Output speed~289 tok/s~122 tok/sSlowerSlower
Terminal-Bench 2.176.2%70.3%LowerLower
GDPval-AA (Elo)1,6561,314LowerLower
MCP Atlas83.6%78.2%LowerLower
ARC-AGI-272.1%77.1%--
HLE40.2%44.4%--

The progression tells a clear story. Gemini 3 Flash (the previous speed tier at $0.50/$3.00) already beat Gemini 2.5 Pro on 18 of 20 benchmarks. Gemini 3.5 Flash now beats Gemini 3.1 Pro on agentic benchmarks. Each generation's Flash model is catching up to, and now surpassing, the previous generation's Pro model. That trend continued in an unexpected direction: rather than a 3.5 Pro establishing a new ceiling, Google shipped three further Flash models and no 3.5 Pro at all, and the next model above Gemini 3.1 Pro was not a Pro but Gemini 4 Argon, generally available from 13 October 2026.

The chart above tells the story quantitatively. Gemini 2.5 Flash achieved approximately 72% of Gemini 2.5 Pro's coding performance. By generation 3.0, Flash reached 82%. Generation 3.1 narrowed to 89%. And now generation 3.5 Flash has crossed 100%, surpassing its own predecessor Pro (108% of 3.1 Pro on Terminal-Bench). This trajectory, the Flash line overtaking the Pro line, is the single most important trend in Google's model development strategy. It means the speed-optimized tier is no longer a compromise. It is the primary model for the majority of production use cases.

For developers currently using Gemini 3.1 Pro, the migration decision depends on your use case. If you are building agents, the Flash line (3.5 Flash at launch, 3.6 or 3.8 Flash today) is faster, cheaper, and better on every metric that matters. If you are building reasoning-heavy applications (research tools, complex analysis), 3.1 Pro's advantages on ARC-AGI-2, HLE, and long-context retrieval may be worth the higher cost and slower speed.

The Context Window Trade-off

One practical consideration for migration: context is not actually a trade-off. Google lists Gemini 3.1 Pro at an input limit of 1,048,576 tokens and a 65,536-token output limit, the same 1M tokens in and the same output cap as 3.5 Flash. For most agent use cases, 1M tokens (roughly 750,000 words or 15,000 pages) is vastly more than needed. But for specific use cases that require processing extremely large documents (entire codebases, full legal depositions, complete financial filings across multiple years), neither model takes more than 1M tokens in one call, so the job needs architectural changes (chunking, retrieval-augmented generation, or multi-pass processing) whichever you pick. The Gemini that changed the shape of the window is Gemini 4 Argon: its 1,048,576 tokens are a unified input-and-output budget, and with long decode it can produce up to 1,048,576 output tokens in a single response, 16x the 65,536-token cap that 3.5 Flash and 3.1 Pro share.

For context, most production AI applications use less than 50,000 tokens of context per call. The 1M window in Flash provides 20x headroom above typical usage. The scenarios where more than 1M genuinely matters are rare: processing a 1,000-page PDF in a single call, loading an entire medium-sized codebase into context, or maintaining a conversation history spanning thousands of exchanges. If you are not doing these things, the context window is irrelevant to your decision.


9. Gemini 3.5 Flash vs. Claude Opus 4.6 and 4.7

The comparison with Anthropic's flagships is the most important competitive matchup because Anthropic holds six of the top ten places on the LMArena overall leaderboard, where Claude Opus 5.5 on high effort sits third at 1,509 Elo and Claude Opus 4.6 on high effort at 1,503, and Claude Opus 4.7 scores 87.6% on SWE-Bench Verified. These are the models most businesses consider as "the best available" - Anthropic.

Where Gemini 3.5 Flash Wins

  • Speed: 289 tok/s vs. ~25 tok/s (11.6x faster)
  • Price: $1.50/$9.00 vs. $5.00/$25.00 (3.3x cheaper input, 2.8x cheaper output)
  • Agentic benchmarks: Leads on MCP Atlas, Finance Agent v2, MMMU-Pro, CharXiv
  • Tool use: Higher scores on structured tool interaction benchmarks

Where Claude Opus Wins

  • SWE-Bench Verified: Opus 4.6 at 80.8%, Opus 4.7 at 87.6% (vs. 55.1% SWE-Pro for Flash)
  • GPQA Diamond: Opus 4.6 at 91.3% (vs. ~90.4% for Flash, close but Opus leads)
  • LMArena Elo: Opus 4.6 at 1,496 standard and 1,503 on high effort, against 1,478 for Gemini 3.5 Flash on high effort
  • Writing quality: Anthropic's models are generally preferred for long-form writing in human evaluations

The Practical Decision

If you need an AI for writing, analysis, and complex software engineering, Claude Opus remains the better choice despite costing 3x more. The SWE-Bench gap (87.6% vs 55.1% on different variants) is significant for production codebases. If you need an AI for agents, tool use, and high-volume automated tasks, Gemini 3.5 Flash delivers comparable or better results at a fraction of the cost and an order of magnitude faster.

For many businesses, the answer is both: Claude Opus for the hard problems and human-facing interactions, Gemini 3.5 Flash for the automated agent infrastructure. Our guides to Claude Code pricing and the Anthropic ecosystem provide the full picture on Anthropic's pricing and products.

The R&D World Analysis

As noted by R&D World, Gemini 3.5 Flash "scores within two points of Anthropic's flagship at a third of the price." This framing captures the disruption perfectly. In previous generations, speed-tier models scored 10-20 points below flagships. A 2-point gap means the quality difference is negligible for most tasks, while the 3x cost reduction and 10x speed improvement are enormous. The implication is that the market for expensive flagship models is shrinking to only those tasks where the last 2% of capability genuinely matters. For everything else, Flash-class models are now sufficient - R&D World.

Claude Sonnet 4.6 and Claude Haiku 4.5 Comparison

The comparison with Anthropic's mid-tier and speed-tier models is also important. Claude Sonnet 4.6 at $3.00/$15.00 sits between Flash ($1.50/$9.00) and Opus ($5.00/$25.00) on pricing, and its successor Claude Sonnet 5 now undercuts Flash on input at $2.00/$10.00, a rate Anthropic made permanent instead of letting it step up to $3.00/$15.00 as originally scheduled, and the current Claude Sonnet 5.5 keeps the same $2/$10 while halving cached input to $0.10. Sonnet 4.6 scores 79.6% on SWE-Bench Verified and 74.1% on GPQA Diamond, both of which are below Flash's capabilities on agentic benchmarks. Sonnet's advantage is in general-purpose writing quality and coding tasks where human preference (rather than benchmark scores) matters. For developers choosing between Flash and Sonnet, the question is whether you value agentic capability (Flash wins) or general-purpose quality (Sonnet may win for some tasks).

Claude Haiku 4.5 at $1.00/$5.00 is cheaper than Flash on both input and output. At ~93 tok/s, it is also reasonably fast (though 3x slower than Flash). Its successor Claude Haiku 5.5 cuts that to $0.10/$0.50 for prompts up to 100,000 tokens ( $0.50/$2.50 above that), which makes Haiku 5.5, not 4.5, the right choice if you need the absolute cheapest Anthropic option. For simple classification, extraction, and routing tasks, Haiku's lower cost may outweigh Flash's capability advantages.


10. Gemini 3.5 Flash vs. GPT-5.5 and GPT-4.1

OpenAI's model lineup creates a different competitive dynamic because it spans from GPT-4.1 Nano ($0.10/$0.40) to GPT-5.5 Pro ($30/$180), covering a wider price range than any other provider - OpenAI. The range has since filled out rather than shifted: GPT-6 Astra sits above GPT-5.5 at $10.00/$50.00, the GPT-6.1 models now cover the middle (Sol at $2/$10, Terra at $1/$5) with the GPT-5.6 family still listed beside them (Sol at $4.00/$20.00, Terra at $2.00/$12.00, Luna at $0.20/$1.20), GPT-6 Luna sits at the low end at $0.10/$0.50, and both GPT-5.5 and GPT-4.1 still hold their launch prices.

GPT-5.5 Comparison

GPT-5.5 leads Gemini 3.5 Flash on SWE-Bench Verified (88.7%), ARC-AGI-2 (85.0%), and GDPval-AA (~1,769 Elo). But it costs $5.00/$30.00 vs. $1.50/$9.00, making it 3.3x more expensive on both input and output. It also runs at ~63 tok/s, which is 4.6x slower than Flash.

Gemini 3.5 Flash leads GPT-5.5 on MCP Atlas, Finance Agent v2, MMMU-Pro, and CharXiv Reasoning. The pattern is clear: GPT-5.5 wins on traditional software engineering and abstract reasoning. Gemini 3.5 Flash wins on tool use, financial analysis, and multimodal understanding. For our complete analysis of GPT-5.5, see the GPT-5.5 benchmarks guide and complete overview.

GPT-4.1 Comparison

GPT-4.1: The Real Price Competitor

GPT-4.1 at $2.00/$8.00 is the closest price competitor to Gemini 3.5 Flash. Its 1M context window matches Flash's. At $2.00 input (vs $1.50 for Flash) and $8.00 output (vs $9.00 for Flash), the total cost for a typical agent interaction is roughly comparable. GPT-4.1 launched in April 2025, more than a year before Flash, and its benchmark data shows lower performance on the agentic benchmarks where Flash excels. The GPT-4.1 Nano variant at $0.10/$0.40 is dramatically cheaper but targets simple classification and extraction rather than agentic work.

For developers currently on GPT-4.1, Flash offers measurably better agentic performance (MCP Atlas, GDPval-AA) at a similar price point. The migration cost is primarily engineering effort to switch API providers, which for well-abstracted codebases is minimal. For developers locked into the OpenAI ecosystem through Assistants API, Function Calling patterns, or enterprise agreements, the switching cost may be higher.

The Pricing Trend Over Time

A broader perspective helps contextualize Flash's pricing. Two years ago, GPT-4 Turbo (the then-current frontier model) cost $10.00/$30.00 per million tokens. Today, Gemini 3.5 Flash delivers significantly better performance on every benchmark at $1.50/$9.00. That is a 6.7x cost reduction on input and 3.3x on output for a model that is measurably better. The trajectory is clear: frontier-class AI capability is getting cheaper by roughly 3-5x per year. Businesses planning multi-year AI strategies should budget for today's costs but expect significant future reductions.


11. Gemini 3.5 Flash vs. Open-Weight Models

The open-weight landscape provides alternatives that run on your own infrastructure, eliminating API costs entirely (but introducing compute costs).

Llama 4 Maverick (Meta's flagship open-weight model) scores 80.5% on MMLU-Pro and 69.8% on GPQA Diamond. It achieves 1,417 Elo on LMArena, placing it in the top tier of open-weight models but below all current closed-source flagships - Meta AI. These scores are competitive with closed models from 6-12 months ago but lag behind the current frontier. The advantage of Maverick is zero API cost: if you have the GPU infrastructure, you pay only for compute. The disadvantage is that you need significant GPU resources (multiple A100s or H100s) to run it at reasonable speed.

Llama 4 Scout is the smaller, faster variant scoring 74.3% on MMLU-Pro and 57.2% on GPQA Diamond. Scout is designed for deployment on more modest hardware (single GPU or even CPU inference with quantization). For businesses that need on-premise AI with no cloud dependency, Scout provides a self-hosted alternative to Flash, though with meaningfully lower capability on every benchmark.

DeepSeek V4-Pro offers a compelling value proposition at $1.32/$3.96 at peak hours and $0.66/$1.98 off-peak with a 1M context window, 80.6% on SWE-Bench Verified, 90.1% on GPQA Diamond, and 87.5% on MMLU-Pro - DataCamp. These are higher than Gemini 3.5 Flash on these specific benchmarks. DeepSeek's cached input pricing ($0.044/1M tokens at peak, $0.022 off-peak) is also dramatically lower than Flash's ($0.15). For pure reasoning tasks at minimum cost, DeepSeek is a strong alternative. However, agentic benchmark data for DeepSeek V4-Pro is limited, and the company's Chinese origin may create compliance concerns for some enterprises. We profiled DeepSeek and other cost-efficient models in our agent swarm cost guide.

Mistral Large 3 at $0.50/$1.50 is the budget option with a 262K context window and 73.1% on MMLU-Pro. It scores lower than Flash on all available benchmarks but costs 3x less on input and 6x less on output. For European businesses with data sovereignty requirements, Mistral's EU-based infrastructure is a compliance advantage that no US-based model can match.

The Build vs. Buy Decision for Open-Weight Models

The choice between Flash (API) and an open-weight model (self-hosted) comes down to scale and control. At low to medium volume (under 10 million tokens/day), Flash's API pricing is almost always cheaper than self-hosting because you avoid the fixed costs of GPU infrastructure, cooling, maintenance, and engineering time. At high volume (over 100 million tokens/day), self-hosting Llama 4 Maverick on H100 GPUs can achieve lower per-token costs, but requires significant upfront investment and operational expertise.

The control argument cuts both ways. Self-hosting gives you data sovereignty (tokens never leave your infrastructure), latency control (no network round-trips to API endpoints), and independence from provider pricing changes. But it also gives you operational burden (GPU procurement, driver updates, model serving infrastructure, scaling, monitoring) that API-based models eliminate. For most businesses, the operational simplicity of API-based models like Flash is the deciding factor. For businesses with strict data residency requirements, high volume, and existing GPU infrastructure, open-weight models remain a viable alternative.

The chart above illustrates the cost-efficiency frontier. Gemini 3.5 Flash delivers 6.1 intelligence points per dollar of output cost, nearly 3x more efficient than Claude Opus 4.6 (2.2) and 3.6x more efficient than GPT-5.5 (1.7). Only DeepSeek V4 Flash (10.7) is more cost-efficient, but at a lower absolute intelligence level. Flash sits at the optimal point on the efficiency frontier: high enough intelligence for production agentic work, low enough cost for enterprise-scale deployment.


12. Cost Analysis: What Agents Actually Cost to Run

Raw API pricing tells you the cost per token. What matters in practice is the cost per task for the agent workflows you actually run. Let me model three common agent scenarios.

Scenario 1: Email Processing Agent

An agent that reads incoming emails, classifies them, drafts responses, and files them. Assume 50 emails/day, average 500 tokens input and 200 tokens output per email.

ModelDaily Input CostDaily Output CostDaily TotalMonthly Cost
Gemini 3.5 Flash$0.04$0.09$0.13$3.90
Claude Sonnet 4.6$0.08$0.15$0.23$6.90
Claude Opus 4.6$0.13$0.25$0.38$11.40
GPT-5.5$0.13$0.30$0.43$12.90

Scenario 2: Research Agent (Heavy Use)

An agent that searches the web, reads documents, synthesizes information, and generates reports. Assume 20 tasks/day, average 10,000 tokens input and 2,000 tokens output per task.

ModelDaily Input CostDaily Output CostDaily TotalMonthly Cost
Gemini 3.5 Flash$0.30$0.36$0.66$19.80
Claude Sonnet 4.6$0.60$0.60$1.20$36.00
Claude Opus 4.6$1.00$1.00$2.00$60.00
GPT-5.5$1.00$1.20$2.20$66.00

Scenario 3: Coding Agent (Continuous)

An agent that reads codebases, generates code, runs tests, and iterates. Assume 100 interactions/day, average 5,000 tokens input and 1,000 tokens output per interaction.

ModelDaily Input CostDaily Output CostDaily TotalMonthly Cost
Gemini 3.5 Flash$0.75$0.90$1.65$49.50
Claude Sonnet 4.6$1.50$1.50$3.00$90.00
Claude Opus 4.6$2.50$2.50$5.00$150.00
GPT-5.5$2.50$3.00$5.50$165.00

With caching enabled (reusing system prompts and tool definitions), Gemini 3.5 Flash's costs drop further. If 80% of input tokens are cached, the input cost drops by ~85%, making the coding agent scenario roughly $18/month instead of $49.50. Caching is the hidden advantage that makes Flash's economics unbeatable for agent workloads. For more on coding agent economics and framework comparisons, see our top 50 AI coding agent frameworks.

Enterprise Scale Economics

The per-task costs seem small, but they compound rapidly at enterprise scale. A company with 1,000 employees, each running 50 agent interactions per day, generates 50,000 daily interactions. At the coding agent cost profile ($1.65/day per user on Flash vs $5.00/day on Claude Opus):

  • Gemini 3.5 Flash: 50,000 interactions x $0.033 = $1,650/day or $49,500/month
  • Claude Opus 4.6: 50,000 interactions x $0.100 = $5,000/day or $150,000/month
  • GPT-5.5: 50,000 interactions x $0.110 = $5,500/day or $165,000/month

The monthly difference between Flash and the flagships is $100,000-$115,000. Over a year, that is $1.2-1.4 million in savings. This is why model pricing is not an academic exercise. For enterprise AI deployments, the choice of model directly impacts the P&L. And at these scales, the capability difference between Flash and the flagships needs to be dramatic to justify the cost premium. For most agent workloads (email processing, document summarization, data extraction, scheduling, monitoring), Flash's performance is indistinguishable from the flagships on the tasks that actually occur.

The Caching Strategy

The 90% caching discount ($0.15/1M vs $1.50/1M for input tokens) creates a specific optimization strategy. Agent systems should be designed to maximize cache reuse. This means structuring your prompts with static elements (system prompt, tool definitions, reference documents, few-shot examples) at the beginning, where they can be cached, and dynamic elements (user input, conversation history, current task context) at the end. The static portion gets cached after the first call, and subsequent calls pay only $0.15/1M for that content instead of $1.50/1M.

For a typical agent with a 5,000-token system prompt and 2,000 tokens of tool definitions, the cached input saves approximately $0.009 per call. Over a million calls (which a busy enterprise agent system handles in weeks, not months), that is $9,450 in savings from caching alone. Google designed this pricing structure specifically to encourage the kind of repetitive, high-volume usage that agent systems generate.


13. The Thinking Mode: Dynamic Reasoning Control

Gemini 3.5 Flash ships with dynamic thinking mode on by default. This means the model shows its reasoning process, similar to chain-of-thought prompting, but with configurable thinking levels that let you control the trade-off between quality, cost, and latency.

The thinking levels are "standard" and "extended." Standard thinking adds minimal latency and cost while improving response quality on complex tasks. Extended thinking dedicates more computation to harder problems, producing better results at higher cost and latency. The key innovation is that this is dynamic: the model automatically adjusts thinking depth based on task complexity rather than requiring you to choose a thinking mode upfront.

For agent systems, this is practical rather than academic. Simple tasks (classifying an email, extracting a name from text) run at maximum speed with minimal thinking. Complex tasks (analyzing a financial report, debugging code) automatically get more reasoning time. The system scales its compute investment to match the difficulty of each task, which is exactly the behavior you want from an agent that handles a heterogeneous mix of easy and hard work throughout the day.

Cost Implications of Thinking Mode

The thinking tokens count toward your output token usage, which means extended thinking on complex tasks costs more than standard thinking on simple tasks. This creates a natural economic incentive: you pay more for harder reasoning. In practice, this means the $9.00/1M output pricing is a blended average. Simple tasks that require minimal thinking cost less per task than the headline rate suggests. Complex tasks that trigger extended thinking cost more.

For developers who want precise cost control, the thinking levels can be configured explicitly rather than left to the model's dynamic adjustment. Setting "standard" thinking for all calls gives predictable costs but potentially worse results on hard tasks. Setting "extended" for all calls gives better results but higher costs and latency. The dynamic default is the sweet spot for most agent systems where task difficulty varies throughout the day.

Comparison with Competitor Thinking Models

The thinking mode approach differs from competitor implementations. OpenAI's o-series (o1, o3) uses a separate model family for chain-of-thought reasoning, which means you need to choose between a thinking model (slow, expensive, better reasoning) and a non-thinking model (fast, cheap, less reasoning). Anthropic's Claude includes an "extended thinking" option that can be toggled per request. Google's approach with 3.5 Flash integrates thinking natively with dynamic adjustment, which eliminates the need for developers to implement their own task-routing logic to select thinking vs non-thinking models.

For agent systems, Google's approach is the most developer-friendly because the model itself decides when to think harder. This removes an entire category of engineering complexity (task difficulty classification and model routing) from the agent developer's responsibility.


14. When to Use Gemini 3.5 Flash (and When Not To)

Based on the benchmark data, pricing, and speed analysis, here are concrete recommendations. Since 8 October 2026 a call to gemini-3.5-flash is served by Gemini 3.6 Flash, so read them as recommendations for the Flash tier, and point new code at 3.6 Flash or 3.8 Flash.

Use Gemini 3.5 Flash when:

You are building AI agents that need to call tools, search the web, process documents, and take actions. The combination of leading agentic benchmarks, fastest speed, and competitive pricing makes the Flash tier the default choice for agent infrastructure. Platforms like O-mega, which build autonomous AI workforces that handle end-to-end business operations, evaluate models like Flash on exactly these agent-relevant criteria.

You need high-volume, cost-sensitive API usage. At $1.50/$9.00 with 90% caching discounts, the economics of Flash support millions of daily interactions without budget concerns.

You need fast response times. At 289 tok/s, user-facing applications feel responsive even with complex multi-turn conversations.

You are building multimodal agents that process images, audio, and video alongside text. The MMMU-Pro and CharXiv scores demonstrate frontier-level multimodal understanding.

Do not use Gemini 3.5 Flash when:

You need the absolute best performance on hard reasoning tasks (ARC-AGI-2, Humanity's Last Exam). Use Gemini 4 Argon, Claude Opus 5.5, or GPT-6 Astra instead.

You need maximum software engineering capability. SWE-Bench scores show a significant gap between Flash and the flagships. Use Claude Opus 5.5 or GPT-6 Astra for complex codebase modifications.

You need perfect long-context retrieval. The 7.6-point gap on MRCR v2 against 3.1 Pro matters for legal discovery, compliance review, and similar needle-in-a-haystack tasks.

You need the best creative writing quality. LMArena human preference rankings consistently favor the flagships (Gemini 4 Argon, Claude Opus) for long-form, nuanced writing. Flash is built for efficiency, not prose artistry.

Industry-Specific Recommendations

Different industries have different model requirements. Here is how 3.5 Flash fits specific verticals.

Financial Services: Flash's leading score on Finance Agent v2 (57.9%, 14.9 points above 3.1 Pro) made it the strongest choice at launch for financial AI agents: portfolio monitoring, transaction classification, risk alerting, and compliance screening. The speed advantage matters in finance because market-relevant information is time-sensitive. An agent that processes market events in 1 second captures opportunities that a 10-second agent misses.

E-commerce: Universal Cart at I/O 2026 runs on Flash, which tells you Google considers it sufficient for shopping agent tasks: product comparison, compatibility checking, price monitoring, and checkout assistance. For e-commerce businesses building their own shopping agents, Flash's tool-use capabilities (MCP Atlas: 83.6%) and cost structure make it the natural choice.

Healthcare and Life Sciences: The MMMU-Pro score (84.0%) demonstrates strong multimodal understanding, which is relevant for processing medical images, lab reports, and clinical documentation. However, the reasoning gap on HLE (40.2% vs flagships) suggests that for complex diagnostic reasoning, a flagship model may produce more reliable results. The Google Co-Scientist program specifically targets scientific research with Gemini models.

Software Development: Flash's Terminal-Bench score (76.2%) is strong for routine coding tasks: code generation, test writing, documentation, and code review. For complex refactoring, debugging production issues, or architecturally significant changes, the SWE-Bench gap against Claude Opus 4.7 (87.6%) and GPT-5.5 (88.7%) is material. For a detailed comparison of AI coding tools, see the Claude Agent SDK guide.

Customer Service: High-volume customer interactions benefit most from Flash's speed and cost. At 289 tok/s, response latency is imperceptible. At $1.50/$9.00, the cost per customer interaction is pennies. The tool-use capability enables agents that can look up order status, process returns, schedule appointments, and escalate complex issues, all within the customer conversation.


15. Gemini 3.5 Pro: What We Know

Gemini 3.5 Pro never shipped. Google pointed at a June 2026 release after I/O, moved the date, and then stopped talking about it. Nothing named 3.5 Pro appears in the Gemini API model list or on the API pricing page, and there is no model card and no published benchmark score for it. Google skipped the name instead: its flagship is now Gemini 4 Argon, generally available since 13 October 2026 at an introductory $2/$10 that the pricing page doubles from 1 December 2026, while Gemini 3.1 Pro remains in preview at $2.00/$12.00 for prompts up to 200k tokens and $4.00/$18.00 above that. Fifteen days after Argon, Google announced Gemini 4 Carbon, which it says ranks first on version 5.0 of the Artificial Analysis Intelligence Index at 59, though on 28 October 2026 it went only to a subset of cyber defenders, ahead of a wider release.

What Google shipped instead was more Flash. Three of them in six weeks: Gemini 3.6 Flash, Gemini 3.7 Flash, and on 2 September 2026 Gemini 3.8 Flash alongside a restricted cyber-defense variant. All three launched at the same $0.75/$3.75 introductory price. On 8 October 2026 Google then deprecated both 3.5 Flash and 3.7 Flash, routing their API requests to 3.6 Flash and 3.8 Flash respectively. A business choosing a Gemini workhorse today is therefore not waiting on a Pro release. It is choosing between 3.6 Flash and 3.8 Flash, both at half or less of what 3.5 Flash cost, with Gemini 4 Argon above them for the hardest work.

What the Missing Pro Means for the Market

The published numbers say the workhorse tier absorbed the flagship work. Google's Gemini 3.8 Flash model card puts 3.8 Flash at 89.4% on Terminal-bench 2.1 against 76.2% for 3.5 Flash, 61.4% on Vals Finance Agent v2 against 57.9%, 86.2% on CharXiv against 84.2%, and 54.9% on HLE-Verified against 40.2%. That last figure is the one to sit with: the reasoning deficit this guide documents for 3.5 Flash has been closed inside the Flash tier, not by a Pro model.

The same card runs 3.8 Flash against the flagships of its launch week. It edges Claude Opus 5 on Terminal-bench 2.1 (89.4% vs 89.1%), Vals Finance Agent v2 (61.4% vs 58.6%) and Harvey's Legal Agent benchmark (10.0% vs 6.7%), while Opus 5 keeps a wide lead on computer use (OSWorld-2.0, 75.4% vs 59.0%) and on the harder Terminal-bench 4.0 (51.8% vs 19.1%). A Flash model trading blows with a $5/$25 flagship at $0.75/$3.75 is the clearest evidence yet that the interesting competition now happens below the flagship tier rather than at it.

For the competitive landscape, the cadence is the story. Anthropic has moved through Claude Opus 4.7, Opus 4.8 and Opus 5, all at $5/$25, then cut the price with Opus 5.5 at $4/$20, and added the Fable and Mythos lines at $10/$50. OpenAI has added the GPT-5.6 family, GPT-6 Astra above GPT-5.5, and the GPT-6.1 models below it. Google went from 3.8 Flash to Gemini 4 Argon in six weeks. Release cycles that used to run six months now run weeks, and the practical consequence for anyone deploying is that a model choice has a shelf life measured in months. Build the abstraction layer that lets you swap models, because you will.


16. First-Principles Analysis: Why Speed Beats Capability

The conventional wisdom in AI is that the most capable model is the best model. This has been true for chatbot-era products where a single model call produces a single response. But the agent era inverts this logic, and understanding why explains Google's strategic choice to release Flash before Pro.

The Agent Economics Equation

An agent completing a task makes N model calls, where N ranges from 3-5 (simple tasks) to 50+ (complex multi-step tasks). The total cost is N times the per-call cost. The total latency is N times the per-call latency (simplified, assuming sequential calls). For an agent making 20 calls per task:

  • Gemini 3.5 Flash: 20 calls x ~1 sec = 20 seconds, 20 calls x ~$0.01 = $0.20
  • Claude Opus 4.6: 20 calls x ~4 sec = 80 seconds, 20 calls x ~$0.03 = $0.60
  • GPT-5.5: 20 calls x ~1.5 sec = 30 seconds, 20 calls x ~$0.035 = $0.70

Flash completes the same task 4x faster and 3x cheaper than Opus, even though Opus is more capable on individual reasoning steps. The question is whether the capability difference matters for the specific task. If the task is "read this email and draft a response," a 90th-percentile model and a 99th-percentile model produce indistinguishable results. The extra capability is wasted. But the 4x speed difference and 3x cost difference compound across millions of tasks per day.

When Capability Matters More Than Speed

There are tasks where the capability gap between Flash and the flagships produces meaningfully different outcomes. Debugging a complex race condition in a distributed system. Synthesizing contradictory evidence from multiple legal documents. Writing a technical paper that requires deep domain expertise. For these tasks, the flagships' higher scores on SWE-Bench, ARC-AGI-2, and HLE translate to tangibly better results.

The optimal architecture for most businesses is a routing system that sends easy tasks to Flash (fast, cheap, good enough) and hard tasks to a flagship (slower, expensive, excellent). This is exactly how Google's own products work: Gemini Spark runs on Flash for routine monitoring and simple actions, but can escalate to more capable models for complex reasoning when needed. This mirrors the approach of autonomous AI platforms like O-mega, where task complexity determines which model handles each step.

The Broader Industry Implication

Gemini 3.5 Flash's release accelerates a trend that has been building for two years: the commoditization of intelligence. When a speed-tier model at $1.50/$9.00 matches or beats last-generation flagships on real-world tasks, the "intelligence premium" that flagship models can charge narrows. This creates pressure on Anthropic, OpenAI, and other providers to either reduce flagship pricing, differentiate on non-benchmark dimensions (developer experience, safety, reliability, ecosystem), or accept that flagships serve an increasingly niche market.

For businesses, this commoditization is pure upside. The capability you need for most production AI tasks is now available at commodity prices. The expensive flagships become reserved for genuinely hard problems, not the default for everything. This is the same dynamic that played out with cloud computing (commodity compute replaced premium servers for most workloads) and will likely play out the same way: rapid adoption of cost-effective options for the majority of use cases, with premium options surviving for specialized needs.

The LMArena Question

One benchmark is notably absent from Flash's dominance: LMArena (formerly LMSYS Chatbot Arena), the crowdsourced human preference ranking. The current text leaderboard, dated 27 October 2026 and covering 419 models, is now Google's at the very top: Gemini 4 Argon on high effort sits second at 1,520 Elo behind a preliminary 1,528 for the not yet widely released Gemini 4 Carbon, with Claude Opus 5.5 on high effort third at 1,509 and Claude Fable 5.1 Max at 1,504. Gemini 3.5 Flash has accumulated enough votes for a stable ranking and sits 30th at 1,478 Elo on high reasoning effort, ahead of the older Gemini 3 Flash at 1,473. The highest-placed Flash model is Gemini 3.8 Flash on high effort, 10th at a preliminary 1,496. Speed-tier models historically score lower on LMArena because the evaluation favors verbose, polished responses over concise, efficient ones, which fits 3.5 Flash sitting 42 points behind Google's own flagship.

This matters because LMArena captures something benchmarks do not: whether humans actually prefer talking to this model. For consumer-facing chatbot products, LMArena scores are arguably more important than any benchmark. For agent infrastructure (where the model talks to tools, not humans), LMArena is less relevant. The distinction reinforces the core thesis: Flash is built for agents, not for chat.

Yuma Heymans (@yumahey), who founded O-mega to build autonomous AI workforces, has observed that the most effective agent systems are not those running the single most powerful model but those that dynamically route between models based on task complexity, exactly the architecture that Gemini 3.5 Flash's speed and cost advantages make possible.


17. Conclusion

Gemini 3.5 Flash was the most important model release of the first half of 2026, not because it is the most capable model in existence but because it is the model that proved the agent era was economically viable. At 289 tokens per second, $1.50/$9.00 per million tokens, and, at launch, leading scores on every agentic benchmark, it is the first speed-tier model to surpass its own company's flagship on the benchmarks that matter most for production agent systems.

The decision framework is straightforward.

If you are building agents (automated task completion, tool use, monitoring, multi-step workflows): the Gemini Flash tier is the default choice, which since 3.5 Flash's deprecation means Gemini 3.6 Flash or 3.8 Flash. Few other models match its combination of agentic capability, speed, and cost.

If you need maximum reasoning capability (research, complex analysis, hard coding problems): use Gemini 4 Argon, Claude Opus 5.5, or GPT-6 Astra. The Gemini 3.5 Pro this tier was waiting for never arrived; Gemini 4 Argon arrived in its place.

If you need the absolute cheapest option and can sacrifice some capability: DeepSeek V4.1 Flash ($0.30/$1.20 at peak, half that off-peak) or Mistral Large ($0.50/$1.50) are dramatically cheaper, though with correspondingly lower agentic scores.

If you need the best creative writing and human preference: Gemini 4 Argon now leads the generally available models on LMArena at 1,520 Elo, with Claude Opus 5.5 at 1,509, and Anthropic's models still fill six of the top ten places, so both families are the preferred choice for human-facing text quality.

The model landscape is no longer a single leaderboard. It is a matrix of trade-offs across speed, cost, capability, and specialization. Gemini 3.5 Flash occupies the most important cell in that matrix for the agent era: fast enough, smart enough, and cheap enough to make AI agents a default infrastructure choice rather than a premium luxury.

The Model Selection Decision Tree

To make the decision concrete, here is how to think about model selection for production AI in 2026.

Start by asking: Is the primary consumer of this model's output a human or another system? If a human reads the output (customer-facing chat, content generation, analysis reports), writing quality matters and LMArena human preference scores are relevant. Gemini 4 Argon (1,520 Elo on high effort) and Claude Opus 5.5 (1,509 on high effort) are the strongest choices for human-facing text quality.

If another system consumes the output (agent orchestration, tool calling, data processing, automated workflows), benchmark scores on agentic tasks and speed/cost matter more than writing quality. The Gemini Flash tier leads on this dimension, with 3.8 Flash now holding the position 3.5 Flash held at launch.

Next, ask: How many model calls does each task require? Single-call tasks (summarize this document, answer this question) tolerate slower, more expensive models because the per-task cost is low. Multi-call tasks (agent workflows with 10-50 calls per task) amplify speed and cost differences multiplicatively. Flash's advantages compound with each additional call.

Finally, ask: What is the failure cost of a wrong answer? For high-stakes decisions (medical diagnosis, legal analysis, financial trading), the capability gap between Flash and the flagships on pure reasoning benchmarks (ARC-AGI-2, HLE, GPQA) matters because errors have significant consequences. For low-stakes, high-volume tasks (email classification, appointment scheduling, data formatting), Flash's capability is more than sufficient and its speed/cost advantages dominate.

This framework, combined with the benchmark and pricing data in this guide, should provide enough information to make an informed model selection for any production use case. For ongoing updates as new models release and benchmarks evolve, we maintain a living comparison at our AI model benchmarks and pricing page.

Pricing and benchmark figures in this guide were last verified against official provider documentation on 30 October 2026. AI model capabilities, pricing, and benchmarks change rapidly. Verify current details on official provider pricing pages before making deployment decisions.

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 Opus 5.5 (Anthropic): $4 in and $20 out per million tokens, cached input $0.20, a 1M token window, listed 22 September 2026.
  • Gemini 4 Argon (Google): $2 in and $10 out per million tokens, cached input $0.10, a 1M token window, listed 13 October 2026.
  • google/gemini-3.8-flash: the feeds no longer list it; they list Gemini 4 Hydrogen in its place. The figures this article gives for it are as written.
  • 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.
  • Claude Haiku 5.5 (Anthropic): $0.10 in and $0.50 out per million tokens, cached input $0.01, a 1M token window, listed 7 October 2026.
  • GPT-6 Astra (OpenAI): $10 in and $50 out per million tokens, cached input $1, a 1.05M token window, listed 4 September 2026.
  • GPT-6.1 Sol (OpenAI): $2 in and $10 out per million tokens, cached input $0.10, a 1.05M token window, listed 29 September 2026.
  • GPT-6.1 Terra (OpenAI): $1 in and $5 out per million tokens, cached input $0.05, a 1.05M token window, listed 26 October 2026.
  • GPT-6 Luna (OpenAI): $0.10 in and $0.50 out per million tokens, cached input $0.01, a 1.05M token window, listed 22 September 2026.
  • GPT-5.5 (OpenAI): $5 in and $30 out per million tokens, cached input $0.50, a 1.05M token window, listed 24 April 2026.
  • GPT-5.5 Pro (OpenAI): $30 in and $180 out per million tokens, a 1.05M token window, listed 24 April 2026.
  • GPT-4.1 (OpenAI): $2 in and $8 out per million tokens, cached input $0.50, a 1.05M token window, listed 14 April 2025.
  • openai/gpt-4.1-nano: the feeds no longer list it; they list GPT-5.4 Nano in its place. The figures this article gives for it are as written.
  • Grok 4.7 (SpaceXAI): $2 in and $6 out per million tokens, cached input $0.50, a 500K token window, listed 21 September 2026.

In this topic: AI models

Which AI model should your agents run on? Every guide on it