Gemini 4 Carbon: Price, Benchmarks, and What It Costs After November 30

Google made Gemini 4 Carbon generally available on November 10, 2026. The API model ID is gemini-4-carbon. Until November 30 it costs $2 per million input tokens and $10 per million output. On December 1 that doubles to $4 and $20.

On the two main independent indexes, Carbon comes close to Claude Fable 5.2 without passing it: 60 against 62 on Artificial Analysis when both run at high, which is Carbon’s top setting but not Fable’s (Fable 5.2 at max scores 65), and 70.45% against 71.31% on Vals. It costs less per task at high, and it is slower.

We also ran Carbon ourselves on one coding task. It passed all 59 hidden tests at both settings, as Fable 5.2 and Opus 5.5 did, and at medium it took about three times as long as either of them. The other numbers come from Google or outside evaluators, and we say which is which.

What changed on November 10

Google announced Carbon on October 28 and gave access only to cybersecurity defenders through its Fairwind Program. The November 10 release opens it to paying API customers and to Google AI Ultra subscribers. Google also set the date the introductory price ends, published a model card and safety report, and started a preview of /swarm in Google Antigravity, which runs a team of agents from one prompt.

Google says verified defenders in its Fairwind Program get specialized cybersecurity capabilities through Carbon for cyber. It has not said whether the rest of the public model differs from what defenders received in October. That matters for reading the outside scores below, because Vals first tested Carbon on October 28 and does not say which version its later figures use.

Gemini 4 Carbon vs Gemini 4 Argon

The two models share a token price, a context window, and tools, so moving between them is a change of one string. Google’s developer guide sends “the majority of software engineering and knowledge work” to Argon and keeps Carbon for the hardest multi-step tasks. It also suggests pairing them, with Carbon on planning and final checks and Argon or Flash on the parallel implementation work.

The Vals gap is one point, and each score has a margin close to one point. The Artificial Analysis gap is ten points. The two evaluators weight different work, so test on yours before you pick by score alone.

What a task really costs

Artificial Analysis runs nine evaluations for its Intelligence Index and reports what each model spends. It prices Carbon at the December rate of $4 and $20. These are the figures listed on November 10:

Carbon at high ranks fourth of 74 entries, and the three above it are all Fable 5.2 settings.

Google says every Carbon token price is half until November 30, including cache reads. If the token mix stays the same, that makes a task about $3.88 at high, $3.30 at medium, and $3.60 at low. That is our arithmetic, and Artificial Analysis does not publish it.

Three comparisons change with the date:

• Carbon at high against Fable 5.2 at high. Two points lower. It costs 15% less from December 1 and about 57% less until November 30, and it takes 46% longer.
• Carbon at medium against Fable 5.2 at medium. Two points higher. Until November 30 it costs about 13% less. From December 1 it costs 73% more. In both periods it takes almost three times as long.
• Carbon against Opus 5.5. Carbon at high is one point above Opus 5.5 at max and 23% cheaper at the December price. Opus 5.5 at xhigh scores 57 for $6.05 in about 16 minutes, close to Carbon at medium on score and faster.

One result looks odd, but it appears on both indexes. Argon at high costs more per task than Carbon at high on Artificial Analysis, $9.95 against $7.76, and more per test on Vals, $13.37 against $10.51, at the same token price. Neither evaluator explains it. Do not assume Argon will be the cheaper model on your bill until you have checked it on your work.

Speed is Carbon’s weak side. Artificial Analysis ranks the high setting 61st of 72 on performance, with 66.85 seconds to the first answer token and 80 output tokens per second. It also ranks it 70th of 73 on verbosity.

Which thinking level to use

Carbon has three levels: low, medium, and high. The API default is medium. Google’s own benchmark tables use the highest setting, so the default you get on the API is not the setting behind the headline numbers.

Start with medium. It scores two points less than high on Artificial Analysis, for 15% less money and 8% less time. Google recommends it for interactive coding and agentic tool use. Medium also had a lower hallucination rate than high on AA-Omniscience, 4% against 11%.

Use high for the work Carbon exists for. Google lists multi-file root-cause debugging, security patching, formal math, and architecture design. It costs about 18% more per task than medium and takes about 8% longer, and the first answer token arrives about three times later, 66.85 seconds against 23.19.

Skip low if your goal is saving money. It scored four points less than medium and cost 9% more on this index. The published data does not show why. Google positions low for fast classification, routing, and extraction, which is short work and not what this index measures.

Compare that with Fable 5.2, where medium costs 58% less than high. On Carbon, the levels move the bill by much less. If cost is your main concern, the bigger decision is which model to use.

What we measured ourselves

We ran one coding task through Atomic Agent, our open-source agent, on OpenRouter with the provider pinned to Google AI Studio. The task is a stock library of about 1,400 lines with 13 planted bugs and one missing feature, scored by 59 hidden tests. Each Carbon setting ran three times on November 10, at the introductory price. The Claude rows are from our Fable 5.2 vs Opus 5.5 test on the same task and the same build of the agent.

The Carbon high cost is the average of two runs. The longest run recorded only 23 of its 69 steps, so we left its cost out, and the true average is probably higher.

If the token mix stays the same, the December price makes these about $1.90 at medium and about $2.60 at high.

The date changes the comparison at both settings, and at medium it changes which Claude model Carbon matches on cost. This month Carbon costs about the same as Opus 5.5 at medium and takes three times as long. From December it costs about the same as Fable 5.2 at medium and also takes three times as long.

At high, Carbon was far cheaper than Fable 5.2 at either price, and it finished sooner on average, 20.4 minutes against 26.6. That is the opposite of the speed gap on Artificial Analysis, so do not read much into one task. Opus 5.5 at high took about half the time of Carbon at high.

Carbon worked in many small steps, 56 on average at medium against 6 for Fable 5.2 and 10 for Opus 5.5. Most of its input came from the cache, which is what kept the cost down.

It is one task, and every setting scored 59 out of 59, so it says nothing about the ceiling of any model. Runs also varied: Carbon at high took between 14.6 and 29.6 minutes. The six Carbon runs ran in parallel on one machine with one API key on November 10. The Claude runs were on November 9, the Opus 5.5 medium runs went one after another, and the other nine Claude runs were partly or fully in parallel, so treat the time comparisons as rough.

Pricing

Prices from Google’s pricing page. OpenRouter lists google/gemini-4-carbon at $2 and $10 with cached input at $0.10.

Four details affect the bill:

• There is no free tier. Carbon is paid only on the API.
• Thinking bills as output. A task that thinks a lot pays $20 per million for those tokens from December 1.
• Caching is automatic. Explicit caching is not supported. Implicit caching gives 95% off cached input.
• Search costs extra. Grounding with Google Search includes 5,000 free requests a month across Gemini 3 and newer models, then costs $14 per 1,000.

If you are building a budget for a product, use the December column. A cost estimate made this month at the listed price will be half of what you pay next month.

Independent results from Vals

Vals AI first published Carbon at high effort on October 28, at the $4 and $20 list price. The scores below are from its leaderboard as of November 9, and some have changed since October 28, for example Tax Agent Bench from 48.86% to 51.78%.

The top of the Vals Index is close to a tie. Fable 5.2 has 71.31%, Sonnet 5.5 has 70.85%, Carbon has 70.45% with a margin of 0.97 points, and Opus 5.5 has 70.14%.

Cost differs more than score. Vals lists Carbon at $10.51 per test, against $17.94 for Fable 5.2, $16.18 for Sonnet 5.5, and $24.77 for Opus 5.5. These models do not all run at the same effort setting on Vals, so read those as what each entry cost in that run.

The pattern by field is the useful part. Carbon is first on the Vals Finance Index, Legal Research Bench, and Tax Agent Bench, where Fable 5.2 ranks 27th. Carbon is seventh on Finance Agent v2.1, so the finance lead has limits. On coding it is ninth on Vibe Code Bench, where Fable 5.2 is first with 93.77%.

Benchmarks from Google

Google’s model evaluation, October 2026. Gemini scores use the highest thinking setting; other models use their maximum setting where a result exists and the best available result otherwise. Google compiled this table from its own runs and outside sources: Terminal-Bench comes from Artificial Analysis, FrontierSWE 2 from Proximal, AutomationBench from Zapier, and Vending-Bench 2 from Andon Labs. Fable 5.2 came out afterwards and is not in the table.

Carbon does not win everything in Google’s own table. Astra leads FrontierSWE 2 and Opus 5.5 leads OSWorld-2.1.

Fable 5.2 can be placed next to two of these rows. On FrontierSWE, Anthropic reports 69.9% for Fable 5.2 against 61.2% for Carbon. On AutomationBench, Carbon is ahead, 53.9% to 51.4%. Google takes both rows from public leaderboards (Proximal and Zapier), and Anthropic’s Opus 5.5 and Fable 5.1 scores match Google’s on both. The Fable 5.2 scores are from Anthropic’s announcement, so treat the comparison as a rough one.

Terminal-Bench shows why you should read the source. In Google’s table, which uses Artificial Analysis’s run, Carbon beats Opus 5.5 by 5.6 points, 65.2% to 59.6%. In the Vals run of the same benchmark Opus 5.5 scores 65.15% and Carbon 63.13%, which reverses the order. Vals also marks 11.11% of the Opus runs on this test as fallbacks. Neither evaluator explains the gap, and the winner can change between setups.

Google also says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than 2 times lower cost per task than the next best external model. It did not give the CursorBench or WANDR scores in the announcement.

On Arena, Carbon at high was first on November 9 with a preliminary 1526 ±10, against 1522 ±5 for Argon at high. Those two are statistically indistinguishable.

Google describes Carbon as built for high factual accuracy. On AA-Omniscience, Artificial Analysis lists a hallucination rate of 11% for Carbon at high, against 4% for Carbon at medium and 4% for Argon at high.

Limits to plan for in agents

• Output limit. Input, thinking, and output share a 1M token window, but a request stops at 65,536 output tokens unless you use a background request with a higher max_output_tokens on the Interactions API, or client-side continuation on the GenerateContent API. Thinking counts against that number.
• Timeouts. Google says hard prompts can take several minutes and recommends a client timeout of at least 15 to 20 minutes, or background mode.
• Removed parameters. temperature, top_p, and top_k are ignored. thinking_budget is deprecated in favor of thinking_level.
• Cache behavior in loops. Implicit caching needs at least 4,096 tokens. Google’s docs say that in agent loops a request gets the full cache benefit only when it adds about 2,050 new tokens or more, so many tiny steps can cost more than you expect.

The knowledge cutoff is March 2026, according to the model card. Input can be text, images, video, audio, or PDF.

Where you can use it

On the API, Carbon is on the paid tier of the Gemini API, in Google AI Studio, and in Gemini for business. OpenRouter lists it as google/gemini-4-carbon.

In consumer products, the Gemini app, AI Mode in Search, and Antigravity offer Carbon to Google AI Ultra subscribers. Google AI Pro subscribers get Argon. If you do not want a subscription, pay-as-you-go through the API or OpenRouter is the way in.

Safety

Google reports no meaningful new capabilities or material performance gains over Argon in its frontier safety assessment, but it still cannot rule out that the model without safeguards reaches its critical cyber capability level. In simulated cyberattack tests, its safeguards blocked 99% of attack attempts from a red-team agent. Carbon scored a 0% attack success rate on Gray Swan’s prompt injection benchmark, against 0.7% for Argon. The cybersecurity scores Google highlights were measured without safeguards, so they do not describe the public product. If security work is your use case, test what the public model will do.

What early users report

On release day it is too early for useful opinions on quality. In an early snapshot of the Hacker News launch thread, about half of the first comments were about access and plans, and only one commenter reported trying the model, on a single paper draft.

The customer quotes in Google’s post are favorable, as you would expect. Kit Langton of OpenCode says Carbon rebuilt a desktop app’s UI overnight and is now a preferred planning model. Google selected these quotes, so give more weight to what you see in the next two weeks.

How to try Gemini 4 Carbon

Download Atomic Agent for Mac, Windows or Linux, or install the terminal version. It is open source and works with your own API keys.

curl -fsSL https://atomicagent.io/install | sh
atomic-agent

On Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell. The source is on GitHub.

Connect OpenRouter with your own key and choose google/gemini-4-carbon from the live model list; see our OpenRouter guide. To run at high, add one line to your openrouter entry under llm.providers in ~/.atomic-agent/config.json:

"extraBody": { "reasoning": { "effort": "high" } }

The setting applies to every model on that provider entry, so change it when you switch models. More in our models guide. We pinned the provider to Google AI Studio, and our test ran on a build included in the 0.7.0 release. You can also add a Gemini API key directly, and it needs to be on the paid tier. We tested only the OpenRouter route.

If you are comparing Carbon with Anthropic’s models, see our Claude Fable 5.2 review and our hands-on Fable 5.2 vs Opus 5.5 test. If you are comparing agents rather than models, see our list of open-source Claude Code alternatives.

The takeaway

Carbon is close to Fable 5.2 on independent indexes and is first on the Vals finance index, legal research, and tax. At high it costs less per task and takes longer. Which model is the better deal depends on the date: until November 30 Carbon at medium undercuts Fable 5.2 at medium with a higher index score, and from December 1 it costs 73% more on Artificial Analysis.

In our one-task test it passed everything at both settings and took about three times as long as Claude at medium.

Test it in November, when mistakes cost half as much, and set your budget with the December price. Start at medium, use high for hard problems, and do not expect low to save money on long tasks. For everyday work, try Argon first, as Google recommends.