Google Announces Gemini 4 Carbon, but General API Access Is Not Live Yet

Google announced Carbon on October 28

Google announced Gemini 4 Carbon on October 28, 2026 as a new frontier model for sustained coding, agentic work and cybersecurity defense. The company says internal teams have used it on engineering tasks that run for days, including finding and patching subtle bugs across large codebases and low-level systems. Those are launch claims, not evidence that Carbon is ready for a reader’s repository or operating environment.

The immediate product fact is narrower: this week, Google plans to give a subset of cyber defenders access and then phase in more trusted Fairwind partners. Its announcement says developer, business and consumer availability will follow “as soon as possible.” No public API model identifier, console route, region list, quota, context limit, maximum output, tool matrix or general-availability date appears in the reviewed announcement.

The stated price is for a future launch contract

Google says Carbon will launch at the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Cached input is intended to cost 95% less than ordinary input, which implies $0.10 per million cached-input tokens if the published relationship is applied exactly.

That is useful planning data, but not yet a complete billable API contract. The announcement does not specify cache eligibility, retention, context tiers, media pricing, batch discounts, grounding or tool fees, regional processing, service levels or whether long-running agent sessions carry additional charges. Teams should preserve the quoted date and verify the live price attached to the eventual callable ID before comparing cost per task.

Google reports gains in coding and cyber evaluations

Google reports that Carbon scores 59 on Artificial Analysis Intelligence Index v5.0 and reaches fourth place at 1768 on WebDev Arena after an 89-point gain. For cyber work, it reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. The announcement explicitly says the three external cyber evaluations were conducted without guardrails.

The company also reports internal results: more than twice Argon’s secure-code success rate, more than ten times another unnamed frontier model’s patch quality in a Chrome-team comparison, and lower fabrication and misalignment rates in internal coding environments. AccessAllGPT did not reproduce any result or inspect prompts, sample selection, budgets, retries, scoring, model versions or failure cases. Treat every figure as a vendor-reported evaluation result, not a production reliability guarantee.

Unguarded benchmark capability is not deployed capability

A score obtained without guardrails does not establish what approved customers can do through the managed product. It may describe underlying model capability while the deployed system blocks actions, changes tools, adds monitoring or requires additional authorization. That distinction matters especially for vulnerability validation and exploitation tasks, where the same capability can support defense or misuse.

Google says Carbon retains a layered safety stack spanning model training, misuse defenses, prompt-injection resistance and chain-of-thought monitoring. It also reports 0.0% successful cheating on Terminal-Bench 4.0, 0.4% cheating attempts, 0.0% attack success on the cited Gray Swan indirect-prompt-injection benchmark and internal reductions in severe misalignment alerts. These are Google’s own claims. Finite benchmark results cannot prove absence of novel prompt injection, policy evasion or unsafe autonomous behavior.

Fairwind is a controlled preview, not public API availability

Google DeepMind describes Fairwind as an early-access program for governments, critical infrastructure operators, core technology platforms and defensive academic labs. It says the program has more than 650 partners, vets applicants, restricts use to defensive or academic purposes and requires controls such as user-level authentication, phishing-resistant multifactor authentication and access tracking.

There is also a documentation mismatch worth preserving: the Fairwind page still described exclusive access to Gemini 4 Argon when AccessAllGPT reviewed it on October 29, while the Carbon announcement said selected defenders would start receiving Carbon this week. That does not negate the announcement, but it means the public program page had not yet become account-level evidence of Carbon access. Prospective users should require approval and a visible Carbon entitlement rather than infer access from an existing Fairwind relationship.

What builders should do next

Track Carbon now if long-running code repair, vulnerability discovery or defensive patching is important to your team. Freeze a small evaluation set before access arrives: representative repositories, known vulnerabilities, clean controls, ambiguous findings, patch tests, prompt-injection cases and tasks where the right answer is to stop. Score accepted findings and patches, false positives, regressions, unsafe actions, intervention, elapsed time, tokens and total task cost.

Wait for the account-visible model ID, context and output limits, supported tools, guardrail behavior, data terms, regions, quotas, dated pricing and a stable snapshot policy before planning production. If early access is offered, constrain it to authorized, isolated targets with least privilege, complete traces, human review and deterministic gates before any patch or exploit action. A launch announcement justifies a bounded evaluation plan; it does not justify an automatic migration.

Disclosures

AccessAllGPT received no model or Fairwind access, credentials, credits, briefing, evaluation logs, data, review, payment or compensation for this article. Google and Google DeepMind did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank Gemini 4 Carbon. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google or Google DeepMind. Publication-wide relationships are listed on the disclosures page.