# Gemini 4 Carbon vs Gemini 4 Argon: which to use for coding Authors: AutoRouter Author profile: https://autorouter.com/about Published: 2026-11-10T19:40:00.000Z Reading time: 10 min Canonical: https://autorouter.com/blog/gemini-4-carbon-vs-argon Publisher: AutoRouter, the Mac app that routes Codex, Claude Code and other AI tools to every model (https://autorouter.com) About AutoRouter: a Mac app (macOS 14 or later) that adds every leading model to Codex, Claude Code, Claude Desktop, T3 and Terminal and picks the model for each task, using your ChatGPT and Claude plans first. Install from https://autorouter.com/AutoRouter.dmg, or ask your agent to follow https://autorouter.com/setup.md. Routing across your plans is free; Pro is $50 a month with $20 of included usage. Unrelated to OpenRouter's Auto Router. > Keep everyday coding on Gemini 4 Argon, and bring in Gemini 4 Carbon, generally available since November 10, 2026, for root-cause debugging in large codebases, architecture changes, migrations and security audits. As of November 10, 2026, the two models share one price list: $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20, and Google says Carbon uses more tokens on each request. By the arithmetic of AutoRouter, the Mac app for Codex and Claude Code, one assumed agent request costs $0.0345 on either model now and $0.069 from December 1. ## Key facts - As of November 10, 2026, Gemini 4 Carbon and Gemini 4 Argon both cost $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens through November 30, 2026, and $4, $0.20 and $20 from December 1 ([Google](https://aistudio.google.com/docs/pricing#gemini-4-carbon)). - One assumed agent request (45,000 cached input tokens, 5,000 new input tokens and 2,000 output tokens) costs $0.0345 on Gemini 4 Carbon or Gemini 4 Argon through November 30, 2026 and $0.069 from December 1, at Google's list prices, by AutoRouter's arithmetic. - Google's November 10, 2026 announcement keeps most software engineering on Gemini 4 Argon and gives Gemini 4 Carbon complex, multi-step coding and debugging, where it uses more tokens than Argon ([Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)). - On Cursor's CursorBench 4.0, as of November 10, 2026, Gemini 4 Carbon at high thinking scored 60.4% at $9.22 per task, Claude Fable 5.2 at max effort 60.0% at $32.79, Claude Opus 5.5 at max effort 57.8% at $13.43 and Gemini 4 Argon at high thinking 52.9% at $4.57 ([Cursor](https://cursor.com/cursorbench)). - Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench for Gemini 4 Carbon, three security benchmarks it says were evaluated without guardrails ([Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/), October 28, 2026). As of November 10, 2026, keep everyday coding on Gemini 4 Argon, and bring in Gemini 4 Carbon for the hardest tasks: root-cause debugging across a large codebase, architecture changes, multi-package migrations and security audits. That is [Google's own advice](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/), published the day Carbon became generally available. The two models [share one price list](https://aistudio.google.com/docs/pricing#gemini-4-carbon): $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20. Google says Carbon uses more tokens than Argon on each request, so any difference in cost comes from token counts. [AutoRouter](https://autorouter.com), the Mac app for Codex and Claude Code, wrote this post, has Gemini 4 Carbon in its model picker and worked the prices through: one assumed agent request costs $0.0345 on either model now and $0.069 from December 1. Every outside figure below comes from the maker's or publisher's own page, linked where it is used and checked on November 10, 2026. ![AutoRouter's launch poster for Gemini 4 Carbon. It reads: On Tuesday, Nov 10, Google released Gemini 4 Carbon. Google's strongest model, built for its hardest problems. Now in AutoRouter. Gemini 4 Carbon. The picture is a cube of carbon atoms joined in a lattice, white on a dark ground.](https://autorouter.com/models/gemini-4-carbon.png) ## What is Gemini 4 Carbon? Gemini 4 Carbon is the model Google calls its strongest, built for cybersecurity work and long agentic tasks in software engineering, knowledge work and science. Google [announced it on October 28, 2026](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) and gave it first to cybersecurity defenders through its Fairwind Program. Gemini 4 Argon [launched on October 13, 2026](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/). As of November 10, 2026, Google's launch post and [model page](https://aistudio.google.com/docs/models/gemini-4-carbon) give these details, and [Argon's page](https://aistudio.google.com/docs/models/gemini-4-argon) lists the same window, inputs and thinking levels: - **Where it runs:** the paid tier of the Gemini API, as `gemini-4-carbon`, in Google AI Studio and Android Studio, and Gemini for business. Google AI Ultra subscribers get it in Google Antigravity, the Gemini app and AI Mode in Search. - **Size:** one window of 1,048,576 tokens shared by input, thinking and output. A response stops at 65,536 tokens unless you use [long decode](https://aistudio.google.com/docs/long-decode). - **Input:** text, image, video, audio and PDF. It outputs text. - **Thinking:** `low`, `medium` or `high`, with `medium` the default. Thinking tokens bill as output tokens. ## Which coding work belongs on Carbon, and which stays on Argon? Most coding stays on Gemini 4 Argon, and Gemini 4 Carbon takes the tasks that need the deepest reasoning. Google's [November 10, 2026 post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) says Argon "remains the ideal choice for the majority of software engineering and knowledge work". Its [developer docs](https://aistudio.google.com/docs/whats-new-gemini-4-carbon#choosing-between-models) split the work this way: | Kind of work | Google points to | Why (Google, November 10, 2026) | | ------------------------------------------------- | ---------------- | ------------------------------------------------------------------------------------------------------ | | Interactive coding and day-to-day feature work | Argon | The docs say Argon answers faster and spends fewer tokens per request than Carbon. | | Fast production loops | Argon | The docs say to keep Argon where speed matters. | | Root-cause debugging in a complex codebase | Carbon | Carbon spends more thinking tokens to explore and verify its work, which raises token use and latency. | | Architectural design and multi-package migrations | Carbon | The launch post lists both as work Carbon does well, with several agents working together. | | Security auditing | Carbon | The docs list it, and Google's security benchmarks below cover this kind of work. | | You aren't sure yet | Argon | Google describes it as the choice for most software engineering. | Google's docs also say to test at all three thinking levels and pick the lowest that meets your quality bar. [Which model to use for which coding task](/blog/which-model-for-which-task) maps the rest of your work to a model tier. ## What does Gemini 4 Carbon cost next to Argon and other frontier models? Gemini 4 Carbon costs the same per token as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens through November 30, 2026, and $4 and $20 from December 1 ([Google's pricing page](https://aistudio.google.com/docs/pricing#gemini-4-carbon), as of November 10, 2026). From December 1, both match Claude Opus 5.5's input and output prices. | List price per million tokens, November 10, 2026 | Gemini 4 Carbon and Gemini 4 Argon | Claude Opus 5.5 | Claude Fable 5.2 | GPT-6 Astra | | ------------------------------------------------ | ------------------------------------------------------------------------------------------- | -------------------------------------------------------------- | ---------------------------- | ----------------------------------------------------- | | Maker | Google | Anthropic | Anthropic | OpenAI | | Input | [$2 through November 30, then $4](https://aistudio.google.com/docs/pricing#gemini-4-carbon) | [$4](https://platform.claude.com/docs/en/about-claude/pricing) | $8 | [$10](https://developers.openai.com/api/docs/pricing) | | Cached input | $0.10 through November 30, then $0.20 | $0.20 | $0.20 | $1 | | Cache write | None listed; [caching is automatic](https://aistudio.google.com/docs/caching) | $5 (5-minute), $8 (1-hour) | $10 (5-minute), $16 (1-hour) | $12.50 (30-minute), $25 (12-hour) | | Output, thinking included | $10 through November 30, then $20 | $20 | $40 | $50 | | Higher prices for long prompts | None listed | None | None | Over 272,000 input tokens: $20 input, $75 output | The arithmetic below prices the assumed agent request this blog's pricing posts use: 45,000 cached input tokens, 5,000 new input tokens and 2,000 output tokens, thinking included. Google lists no cache-write price for either model, so new input bills at the input price. Caching is implicit, and the example assumes all 45,000 repeated tokens hit the cache, which Google's [caching guide](https://aistudio.google.com/docs/caching) does not promise. ```text Assumed request: 45,000 cached input, 5,000 new input, 2,000 output tokens Google's list prices per million tokens as of November 10, 2026 Gemini 4 Carbon and Gemini 4 Argon share these prices Through November 30, 2026 45,000 cached input x $0.10 = $0.0045 5,000 new input x $2.00 = $0.0100 (no cache-write price listed) 2,000 output x $10.00 = $0.0200 Total $0.0345 From December 1, 2026 45,000 cached input x $0.20 = $0.0090 5,000 new input x $4.00 = $0.0200 2,000 output x $20.00 = $0.0400 Total $0.0690 Each extra 1,000 thinking or output tokens Through November 30: + $0.01 From December 1: + $0.02 ``` At the launch price, $20 of API usage pays for 579 of these requests on Carbon or Argon, and for 289 from December 1. The same request costs $0.074 on Claude Opus 5.5 and $0.139 on Claude Fable 5.2, as worked in [Claude Fable 5.2 vs Opus 5.5: which to use for coding](/blog/claude-fable-5-2-vs-opus-5-5). On GPT-6 Astra it costs $0.2075: $0.045 for cached input, $0.0625 for new input at the 30-minute cache-write price, and $0.10 for output. The arithmetic gives every model the same token counts, and real counts differ: Google says Carbon spends more thinking tokens on each request than Argon. ## What does Google claim about Gemini 4 Carbon, and who measured it? Google claims Gemini 4 Carbon leads three security benchmarks and Cursor's coding benchmark. As of November 10, 2026, Cursor's own page shows that coding lead at 0.4 points. **Security.** In its [October 28, 2026 post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/), Google reports 88.2% on SEC-bench Pro, which tests finding vulnerabilities in real software, 54.8% on ExploitGym, which tests exploiting them, and 71.0% on CWE-bench, which tests patching them. Google says these were evaluated without guardrails, and its November 10 post says Carbon leads them when accessed through the Fairwind Program. These scores describe security work. None of the three measures feature work or refactoring. **Coding.** [Google says](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) Carbon ranks first on CursorBench 4.0, a set of long coding problems that Cursor runs, at less than half the cost per task of the next model from another maker. Cursor's page and Artificial Analysis's model pages list these results as of November 10, 2026, each model at its highest setting: | Result, as of November 10, 2026 | Gemini 4 Carbon (`high`) | Gemini 4 Argon (`high`) | Claude Opus 5.5 (`max`) | Claude Fable 5.2 (`max`) | GPT-6 Astra (`max`) | | ---------------------------------------------------------------------- | ---------------------------------------------------------- | --------------------------------------------------------- | ---------------------------------------------------------- | ----------------------------------------------------------- | ------------------------------------------------------ | | [CursorBench 4.0](https://cursor.com/cursorbench) score, run by Cursor | 60.4% | 52.9% | 57.8% | 60.0% | Not listed | | CursorBench 4.0 cost per task | $9.22 | $4.57 | $13.43 | $32.79 | Not listed | | Artificial Analysis Intelligence Index v5.0.1 | [60](https://artificialanalysis.ai/models/gemini-4-carbon) | [50](https://artificialanalysis.ai/models/gemini-4-argon) | [59](https://artificialanalysis.ai/models/claude-opus-5-5) | [65](https://artificialanalysis.ai/models/claude-fable-5-2) | [57](https://artificialanalysis.ai/models/gpt-6-astra) | - **Cost per task depends on the workload.** Carbon cost twice as much per task as Argon on CursorBench 4.0, $9.22 against $4.57. - **These are top settings.** At `medium`, Carbon's default, Cursor lists 60.1% at $8.49 per task. Claude Opus 5.5 at `high` scored 56.0% at $3.97 per task on the same board. - **Google's ranking claim has aged.** On October 28, Google said Carbon ranked first on the Artificial Analysis Intelligence Index at 59. Artificial Analysis now lists it fourth of 74 entries at 60, and lists Claude Fable 5.2, released November 9, first at 65. Cursor notes that small differences in its scores may not be statistically meaningful. To judge your own work, run one hard task from your backlog on Carbon and on the model you use now. ## How do you try Gemini 4 Carbon in Codex or Claude Code? Use a gateway that lists the model, such as OpenRouter, or pick Gemini 4 Carbon in the model picker if you use AutoRouter. As of November 10, 2026, Google's docs have no guide for running Carbon inside either app. - **Codex.** A custom provider has to speak OpenAI's Responses API, the only `wire_api` value Codex supports ([config reference](https://learn.chatgpt.com/docs/config-file/config-reference)). Google's [OpenAI compatibility page](https://aistudio.google.com/docs/openai) documents a Chat Completions endpoint and no Responses endpoint. [OpenRouter lists the model](https://openrouter.ai/google/gemini-4-carbon) as `google/gemini-4-carbon` at Google's input and output prices, and publishes a [Codex guide](https://openrouter.ai/docs/cookbook/coding-agents/codex-cli). - **Claude Code.** Claude Code connects to a gateway through one of the [API formats its docs list](https://code.claude.com/docs/en/llm-gateway-protocol#api-formats), such as Anthropic Messages. Google's Gemini API docs describe none of them, so requests pass through a gateway that translates. OpenRouter publishes a [Claude Code guide](https://openrouter.ai/docs/cookbook/coding-agents/claude-code-integration). AutoRouter has not run Carbon through OpenRouter for this post, so follow OpenRouter's guides for the exact settings. [Try a new model in Codex or Claude Code without another API key](/blog/try-new-models-in-codex-and-claude-code) covers both routes. ## How does AutoRouter handle Gemini 4 Carbon? Gemini 4 Carbon is in AutoRouter's model picker, where you can choose it yourself or name it in a plain-English rule. AutoRouter, the Mac app for Codex and Claude Code, [puts new models in the picker you already use](https://autorouter.com/#product) right away, with no keys or endpoints to set up. - **Choose it.** Pick Gemini 4 Carbon from the model menu in Codex, or with `/model` in Claude Code, and your prompts go straight to it. - **Name it in a rule.** Rules are plain English, for example "Use Gemini 4 Carbon for security audits." Your rules come before AutoRouter's own. - **Or select AutoRouter.** It works out how hard each task is and only considers models that clear the quality bar for it. Hard work goes to the most capable model your plans cover, and to credits only when none is strong enough. - **See what happened.** Every request shows its model, what it cost and why. Models outside your ChatGPT and Claude plans come with AutoRouter Pro, at $50 a month with $20 of included usage each month. Paid models bill at the provider's price plus 10%, which AutoRouter earns. The AutoRouter team tests every new model as it becomes available. AutoRouter runs on macOS 14 or later and is in beta, and requests pass through its servers on their way to each provider ([AutoRouter privacy](https://autorouter.com/privacy)). ## Questions ### Is Gemini 4 Carbon better than Claude Fable 5.2 at coding? It depends on the task. As of November 10, 2026, Carbon at `high` and Fable 5.2 at `max` are 0.4 points apart on Cursor's [CursorBench 4.0](https://cursor.com/cursorbench), 60.4% to 60.0%, and Carbon's cost per task is $9.22 against $32.79. On the Artificial Analysis Intelligence Index, which also covers knowledge work and science, [Fable 5.2](https://artificialanalysis.ai/models/claude-fable-5-2) is listed at 65 and Carbon at 60. In Claude Code, Anthropic's [models overview](https://platform.claude.com/docs/en/models/overview) says to start most work on Claude Opus 5.5. ### What does Gemini 4 Carbon cost after November 30, 2026? From December 1, 2026, Gemini 4 Carbon costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, twice the launch price ([Google's pricing page](https://aistudio.google.com/docs/pricing#gemini-4-carbon), as of November 10, 2026). Gemini 4 Argon moves to the same prices on the same day. The assumed request above goes from $0.0345 to $0.069. ### Should you raise Argon's thinking level or switch to Carbon? Google's docs give no fixed order, so test both. They say to run your evaluations at `low`, `medium` and `high` and pick the lowest level that meets your quality bar ([What's new in Gemini 4 Carbon](https://aistudio.google.com/docs/whats-new-gemini-4-carbon)). Moving from Argon to Carbon needs only the model ID changed to `gemini-4-carbon`. ## Sources - [Gemini 4 Carbon is now available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/), Google, November 10, 2026 - [Gemini 4 Carbon: our next frontier model](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/), Google, October 28, 2026 - [Gemini 4 Argon: Built to solve your hardest problems](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/), Google, October 13, 2026 - [Gemini Developer API pricing](https://aistudio.google.com/docs/pricing), Google, checked November 10, 2026 - [Gemini 4 Carbon](https://aistudio.google.com/docs/models/gemini-4-carbon) and [Gemini 4 Argon](https://aistudio.google.com/docs/models/gemini-4-argon), Gemini API docs, Google - [What's new in Gemini 4 Carbon](https://aistudio.google.com/docs/whats-new-gemini-4-carbon), Gemini API docs, Google - [Context caching](https://aistudio.google.com/docs/caching), [Long decode](https://aistudio.google.com/docs/long-decode) and [OpenAI compatibility](https://aistudio.google.com/docs/openai), Gemini API docs, Google - [CursorBench](https://cursor.com/cursorbench), Cursor, checked November 10, 2026 - [Gemini 4 Carbon](https://artificialanalysis.ai/models/gemini-4-carbon), [Gemini 4 Argon](https://artificialanalysis.ai/models/gemini-4-argon), [Claude Opus 5.5](https://artificialanalysis.ai/models/claude-opus-5-5), [Claude Fable 5.2](https://artificialanalysis.ai/models/claude-fable-5-2) and [GPT-6 Astra](https://artificialanalysis.ai/models/gpt-6-astra), Artificial Analysis, checked November 10, 2026 - [Pricing](https://platform.claude.com/docs/en/about-claude/pricing) and [Models overview](https://platform.claude.com/docs/en/models/overview), Claude Platform docs, Anthropic - [API pricing](https://developers.openai.com/api/docs/pricing), OpenAI - [Configuration reference](https://learn.chatgpt.com/docs/config-file/config-reference), Codex docs, OpenAI - [LLM gateway protocol](https://code.claude.com/docs/en/llm-gateway-protocol), Claude Code docs, Anthropic - [Google: Gemini 4 Carbon](https://openrouter.ai/google/gemini-4-carbon), OpenRouter's listing, and its [Codex](https://openrouter.ai/docs/cookbook/coding-agents/codex-cli) and [Claude Code](https://openrouter.ai/docs/cookbook/coding-agents/claude-code-integration) guides - [AutoRouter](https://autorouter.com) and [AutoRouter privacy](https://autorouter.com/privacy) ## Read next - [Muse Ember vs Claude Haiku 5.5: Meta's low-cost model for coding](https://autorouter.com/blog/muse-ember-vs-claude-haiku-5-5.md): Muse Ember is Meta's smallest Muse model, released on November 9, 2026 at $0.10 per million input tokens and $0.50 per million output tokens, the same list prices as Claude Haiku 5.5 and GPT-6 Luna on prompts up to 100,000 tokens. Meta says it matches Haiku 5.5's best score on the Artificial Analysis Intelligence Index at about a third of the cost per task, and Meta's own docs keep multi-file changes on a larger model. AutoRouter, the Mac app for Codex and Claude Code, has Muse Ember in its model picker. - [Claude Fable 5.2 vs Opus 5.5: which to use for coding](https://autorouter.com/blog/claude-fable-5-2-vs-opus-5-5.md): Keep everyday coding on Claude Opus 5.5 and move to Claude Fable 5.2, released November 9, 2026, for major refactors, migrations, long unattended runs and tasks where Opus 5.5 at high effort still falls short. As of November 10, 2026, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, twice Opus 5.5's price, with the same $0.20 cache-read price. By the arithmetic of AutoRouter, the Mac app for Codex and Claude Code, one assumed agent request costs $0.139 on Fable 5.2 and $0.074 on Opus 5.5. - [Claude Opus 5.5 vs Sonnet 5.5: which to use for coding](https://autorouter.com/blog/claude-opus-5-5-vs-sonnet-5-5.md): Start coding work on Claude Opus 5.5, released September 22, 2026 and Claude Code's default model, interface design and animation included, and move a task down to Claude Sonnet 5.5 when it is well scoped, such as a bug fix or a feature built against a spec. As of November 10, 2026, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, twice Sonnet 5.5's $2 and $10. By the arithmetic of AutoRouter, the Mac app for Codex and Claude Code, one assumed agent request costs $0.074 on Opus 5.5 and $0.037 on Sonnet 5.5. Tags: [Models](https://autorouter.com/blog/tags/models.md), [Comparisons](https://autorouter.com/blog/tags/comparisons.md), [Costs](https://autorouter.com/blog/tags/costs.md) Drafted with AI from the sources linked above, then checked and edited by the AutoRouter team.