Google’s frontier model expands beyond the Fairwind preview. Its first public evaluations show different strengths across coding, search and business workflows, while DeepMind’s AlphaProof Code project shows how Carbon was used to verify parts of Signal’s cryptography.

Google made Gemini 4 Carbon broadly available on November 10, 2026, after an October 28 release to a select group of cybersecurity defenders through its Fairwind Program. The model launched on the paid Gemini API at $2 per million input tokens and $10 per million output tokens through November 30. On the same day, Google DeepMind published research showing how Carbon-powered agents helped verify security properties of a compiled build of Signal’s cryptographic library.

What is Gemini 4 Carbon designed to do?

Carbon is based on Gemini 4 Argon, which Google made widely available on October 13. Google still recommends Argon for most software engineering and knowledge work. It positions Carbon for tasks such as debugging across a large codebase, investigating a scientific question, or revisiting an answer through several rounds of testing. Google says Carbon can spend more tokens than Argon exploring a problem and checking its work, an important consideration when estimating the cost of a long-running job.

What did AlphaProof Code verify in Signal?

The DeepMind team used AlphaProof Code, a multi-agent formal verification framework, to prove specified security properties of Signal’s PQXDH handshake and Triple Ratchet in an AArch64 build of libsignal, the protocol’s reference implementation. The researchers used the stock Rust compiler and ran the library through a command-line test harness. The proofs covered properties including forward secrecy, authentication and post-compromise security under the paper’s stated assumptions. Their scope was the research build and its modeled environment.Human experts reviewed every formal specification, which defines the property the implementation has to satisfy.

Carbon-powered agents then generated about 6.5 million lines of Lean proof code. Lean checked those proofs against the specifications. DeepMind reports that five researchers carried out the proof work over four weeks. The team also reported that agents flagged roughly 1% of the codebase for additional human inspection

Earlier work had established security properties of the protocol’s design. The AlphaProof Code study shows how researchers connected specified guarantees to the compiled machine code of the existing library. The command-line harness and its assumptions define where those guarantees apply. For software teams evaluating AI-assisted verification, the division of work matters: people decide what the proof must establish, agents construct it, and Lean checks it

How much does the Gemini 4 Carbon API cost?

For the public Carbon model, Google’s introductory rate matches Argon’s. The scheduled December 1 rate is $4 per million input tokens and $20 per million output tokens. Google Cloud’s global pricing table lists cache reads at $0.10 per million input tokens during the promotion and $0.20 afterward.

Model or pricing periodInput per 1 million tokensOutput per 1 million tokens
Gemini 4 Carbon through November 30, 2026$2$10
Gemini 4 Carbon from December 1, 2026$4$20
Claude Fable 5.2, standard API$8$40
GPT-6 Astra, standard API$10$50

Anthropic released Claude Fable 5.2 on November 9, a day before Carbon’s wider rollout. At its scheduled December rates, Carbon costs half as much per input and output token as Fable 5.2 at Anthropic’s standard rate. The economics of a completed job also depend on how many tokens each model uses, how often it calls tools, and whether a failed run needs to be retried.

Gemini 4 Carbon vs. Claude Fable 5.2 on benchmarks

The closest contest in this group was on CursorBench 4.0, which tests agents on ambiguous, multi-file coding tasks. Carbon High scored 60.4% and Fable 5.2 Max scored 60.0%, with a substantial gap in the listed costs. On FrontierSWE v2, Anthropic reported a wider score difference in Fable’s favour.

Benchmark and sourceGemini 4 CarbonClaude Fable 5.2
CursorBench 4.0, Cursor60.4%, $9.22 listed cost (High)60.0%, $32.79 listed cost (Max)
AutomationBench, Zapier53.88%, $2.12 per task (High)51.45%, $4.64 per task (Max with default fallbacks)
WANDR, Perplexity0.857, $5.86 per task0.851, $14.51 per task
FrontierSWE v2, Anthropic61.2%69.9%

The WANDR results come from Perplexity’s own published comparison. On that wide-search benchmark, Carbon achieved a similar score to Fable at roughly 40% of the reported cost per task. Zapier’s leaderboard also shows a trade-off within Google’s lineup. Argon High scored 51.29% at $1.70 per task at standard list pricing, compared with 53.88% at $2.12 for Carbon High. Whether the improvement justifies the extra cost will depend on what happens when a workflow fails.

Google’s own long-context evaluation reported 96.2% F1 for Carbon on the GraphWalks subset covering context lengths from 256,000 to one million tokens. The result measures graph traversal on that subset. Google also reports cybersecurity evaluations in its model card, including assessments relevant to vulnerability discovery and patching.  Carbon model card

Where can developers and enterprises access Gemini 4 Carbon?

Google opened access through the paid Gemini API in Google AI Studio and through its enterprise offerings. It also began rolling out Carbon to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity. GitHub began a gradual rollout to Copilot Pro+, Max, Business and Enterprise customers. Cursor lists Carbon among its supported models. Google Cloud’s specialized Gemini 4 Carbon for cyber variant remains restricted to organisations approved under the Fairwind Program.

For an enterprise team comparing Carbon with Argon or Fable, a useful trial would start with work that already takes several rounds of review, such as a code migration or a workflow that reconciles records across multiple systems. Compare how often each model completes the job correctly, how long reviewers spend checking the output, and what the run costs at December rates. Those measurements can show where a more capable model earns its place in a production workflow.


Discover more from Poniak Times

Subscribe to get the latest posts sent to your email.