Google made Gemini 4 Carbon available on November 10 to all paying Gemini API developers and enterprises, and began rolling it out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and the Antigravity coding platform. The price is $2 per million input tokens and $10 per million output until November 30, then $4 and $20. Until now Carbon was a model you could read about but not buy: its cyber variant went first to defenders in the Fairwind Program, and as of last week access was limited to companies Google selected. Now it is a line item, which reopens a comparison most teams thought they had finished.
Priced like Argon, which is the point
The introductory rate is identical to Gemini 4 Argon's, and so is the destination: Google's pricing page moves Argon from $2/$10 to $4/$20 on December 1, the day after Carbon's discount expires. Google's two top models therefore share one meter. What separates them is how many tokens each burns per task, and Google is candid that Carbon uses more tokens than Argon to reason and verify, still recommending Argon for most software engineering and knowledge work.
Against rivals the arithmetic is blunt. Claude Fable 5.2 lists at $8/$40 and GPT-6 Astra at $10/$50. At full price Carbon costs half of Fable per token and 40% of Astra; through November, a quarter and a fifth. When this site covered Fable 5.2 yesterday, Carbon was the cheaper model nobody could buy. Now it is simply the cheaper model. At $4/$20 it matches Claude Opus 5.5's per-token price, as The Batch pointed out last week.
A first place measured in tenths
Google's headline coding claim is first place on CursorBench 4.0, and here the referee matters: Cursor runs it, on ambiguous multi-file tasks taken from real Cursor sessions. Cursor's leaderboard has Carbon at 60.4% on high effort, 60.1% on medium and 58.6% on low. Fable 5.2 at max effort scores 60.0%, Opus 5.5 at max 57.8%, and Argon at high 52.9%. Cursor itself warns that small score differences may not be statistically meaningful, and 0.4 points is small.
The cost column says more. Carbon's high-effort run costs $9.22 per task against $32.79 for Fable 5.2 at max, which is where Google's "more than 2X lower cost per task" line comes from. Yet Opus 5.5 at high effort reaches 56.0% for $3.97. Cursor prices tokens at each model's published rates but does not say whether Carbon's figure uses the introductory or the standard rate, which is a factor of two.
The rest of the scorecard needs labels. Google says Carbon also ranks first on Perplexity's WANDR and Zapier's AutomationBench; third-party tests, but the ranks reach us through Google. On November 6, The Batch reported Carbon first on the Artificial Analysis Intelligence Index v5 at 59 and on Arena's text leaderboard at 1,528 Elo, three days before Fable 5.2 shipped.
What changes in your code
The model ID is gemini-4-carbon, listed as GA on Google Cloud's Gemini Enterprise Agent Platform from November 10 in the global, us and eu multi-regions, with a March 2026 knowledge cutoff. The context window is 1,048,576 tokens, shared between input and output; a single segment returns at most 65,536 output tokens when streaming and 32,768 when not. Images, audio, video and documents go in; only text comes out.
Three details will bite a naive swap. Thinking levels are LOW, MEDIUM (the default) and HIGH, and passing thinking_budget or thinking_level="MINIMAL" returns an HTTP 400. Sampling is fixed at temperature 1.0, topP 0.95 and topK 64, and custom values are ignored, so anyone lowering temperature for determinism is adjusting nothing. Caching is implicit only, with a 4,096-token minimum; explicit caching, tuning, the Live API and RAG Engine are unsupported. The model answers on GenerateContent, the Interactions API and Chat Completions, under Standard and Priority pay-as-you-go, batch and Provisioned Throughput.
On the developer API, access is paid tier only. None of the pages I read publish Carbon-specific rate limits or retire an older Gemini model, so plan quota from your console rather than a press release.
Six million lines of Lean, or three
The same day, Google DeepMind introduced AlphaProof Code, a framework in which humans choose what to prove and agents powered by Gemini Carbon do the proving. Its first target was Signal. In four weeks the agents proved that Signal's compiled code, not merely the protocol design, meets key security properties, such as keeping messages protected before and after a device is temporarily compromised. Agents flagged about one percent of the codebase for human review; Lean checked the rest.
The limits deserve equal billing. The post's body puts the output at more than six million lines of Lean, while the same page's description says more than three million lines of machine-checked proof, and nothing reconciles the two. The proofs are only as good as the specifications, which models drafted and human experts reviewed. The compiled code is reasoned about through what the post calls a "Lean processor model," left unexplained. No new Signal bugs turned up; as a test, the team reinserted two previously fixed bugs and got machine-checked proofs that those versions were insecure. The verification artifacts are being open-sourced; AlphaProof Code itself has no stated public availability.
The 30,000-foot read
Strip away the rankings and Carbon's launch is a pricing decision. Google has put its strongest model on the same meter as its workhorse and set that meter at half of Fable 5.2 per token and less than half of GPT-6 Astra. The benchmark lead is real but thin, a statistical tie on the one prominent coding test Google does not run. The price gap is not thin. For engineers choosing a model the procedure has not changed: run your own tasks, measure cost per completed task rather than per token, and remember that Carbon's list price doubles on December 1.
The leaderboard is outside anyone's control; it may already have moved since Fable 5.2 arrived. The harness is not. AlphaProof Code points at the more durable shift, models that write proofs as well as code, where a checker rather than a launch post decides who was right. Watch whether the Signal artifacts survive outside review, and whether Carbon's appetite for tokens eats its price advantage once December's meter starts running.
Sources: Google, Gemini 4 Carbon is now available (November 10 availability, surfaces, introductory and post-November 30 pricing, parity with Argon, token use versus Argon, Fairwind history, benchmark rankings, cost-per-task claim); Google, Gemini 4 Carbon for developers (paid API tier, AI Studio and Antigravity access); Google Cloud, Gemini 4 Carbon model page (model ID, GA date, regions, context and output limits, cutoff, thinking levels, fixed sampling, caching, APIs, consumption options); Gemini Developer API pricing (Argon's December 1 price change); Claude API docs, Fable 5.2 (Fable 5.2 list price); OpenAI API docs, GPT-6 Astra (Astra list price); Cursor, CursorBench 4.0 (scores, cost per task, task source, cost method, variance warning); The Batch, DeepLearning.AI (pre-launch access, Intelligence Index and Arena ranks, Opus 5.5 price parity); Google DeepMind, AlphaProof Code (Signal verification, duration, review share, line counts, specification review, bug reinsertion test, open-sourcing). CAVEAT: WANDR and AutomationBench ranks are as reported by Google, not checked on each leaderboard. CursorBench is run by Cursor, not Google; its Carbon cost per task does not state which price tier was applied. Third-party index ranks predate Fable 5.2's November 9 release. DeepMind's page gives two different line counts (over six million and over three million); neither has been independently reviewed. Google's pricing page, as fetched, did not yet list Carbon, so Carbon prices rest on Google's announcement. No Carbon rate limits or Gemini deprecations were found in the official pages consulted.