Gemini 4 Carbon is live: benchmarks, pricing and how (not) to access it

Google unveiled Gemini 4 Carbon on 28 October 2026 — its latest frontier AI model, positioned around cybersecurity and autonomous software engineering, and following Gemini 4 Argon by just two weeks [CNA/Reuters, The Information]. The announcement cycle called it the company's strongest model yet. The catch most headlines skipped: you cannot simply start using it. Access is initially restricted to a subset of cybersecurity defenders through Google's Fairwind Program while safeguards are tested [TechDogs]. This piece separates what was actually announced from what is a company claim, and gives you a practical plan that doesn't involve waiting for a model you can't yet buy.

The five-minute picture: what launched, exactly

• Gemini 4 Carbon, announced 28 Oct 2026 — a frontier model focused on cybersecurity tasks (vulnerability discovery, exploitation testing, software patching) and long-running engineering work [CNA/Reuters].
• Second major Gemini model in a month: Gemini 4 Argon launched roughly two weeks earlier, with Carbon positioned above it [The Information, CNA/Reuters].
• Gated rollout: available first to selected cybersecurity defenders via Google's Fairwind Program; Google says it is updating safeguards before extending access to developers, businesses and consumers. No public availability date announced [TechDogs].
• Introductory API pricing stated at $2 per million input tokens and $10 per million output tokens — matching Argon's introductory rate; cached input quoted at a 95% discount, with post-intro rates reported rising to $4/$20 [TechDogs, CNA/Reuters]. Pricing is announced but not yet broadly purchasable, because access itself is restricted.
• US pre-release testing: Google confirmed it is participating in the US government's voluntary pre-release model testing process [TechDogs].

What Google claimed on benchmarks — and the caveats attached

Reported via TechDogs from Google's announcement: 88.2% on SEC-bench Pro (identifying real software vulnerabilities), 54.8% on ExploitGym (evaluating whether a vulnerability can be exploited) and 71.0% on CWE-bench (generating patches) — and a first-place score of 59 on version 5.0 of the Artificial Analysis Intelligence Index. Google also said Carbon generates secure code at more than twice Argon's success rate on internal benchmarks, and that Google's Chrome team observed it performing more than ten times better than another lab's frontier model at generating high-quality security patches.

Three caveats Google itself acknowledged, per the same coverage, matter before you repeat these numbers as fact: (1) the cybersecurity evaluations ran without guardrails, so they are capability measurements, not expected production performance; (2) the '2x Argon' and '10x other model' figures are internal or team-specific observations, not independent third-party benchmarks; (3) on reliability, Google reported almost 80% fewer high-severity misalignment alerts and roughly half the unsupported claims versus Argon in its own live-coding testing, plus a 0% attack success rate on Gray Swan's indirect prompt-injection benchmark under tested conditions — which is encouraging but not immunity. Treat every figure here as 'Google-reported' until independent evals land.

Why the access is gated (and why that's the real story)

A model that can find a vulnerability, validate whether it is exploitable, and write the patch has obvious defensive value — and an obvious dual-use problem: the same capability map helps an attacker. That is why Google is starting with trusted cybersecurity defenders through Fairwind rather than flipping a public switch, and why it is routing the model through the US government's voluntary pre-release testing. Koray Kavukcuoglu (SVP, Google DeepMind; Chief AI Architect) described it as a deliberate rollout, with broader developer/business/consumer access coming only after safeguards are updated [TechDogs]. For smaller organisations the promised upside — security assistance without a big in-house team — is real but contingent on that broader release actually arriving [Small Business Trends via TechDogs].

How it positions against the other frontier labs

Coverage framed the launch as stepping up competition with OpenAI and Anthropic specifically in coding and cybersecurity capability races [CNA/Reuters, The Information]. No credible head-to-head exists yet, because Carbon is gated — so the honest comparison right now is about accessibility, not capability:

• Gemini 4 Carbon (Google) — announced cybersecurity/coding flagship; restricted access via Fairwind; announced intro API pricing $2/$10 per million tokens. Not generally available. We could not verify its general availability, exact context window or safety-case details on a public product page as of 2026-10-30 — that gap itself is the signal: it's not for you yet.
• Gemini 4 Argon (Google) — the immediately prior model, the baseline every Carbon claim is measured against; also the currently reachable end of Google's frontier tier via the Gemini API (verify current model names, limits and prices on Google's own pricing page — they move).
• OpenAI and Anthropic — Reuters framing places both in the same coding/cybersecurity race; neither is displaced by this launch because Carbon has no open API surface to compare against yet. Pick by what you can actually use today for your workload.

What to do today: a no-panic plan

1. Do not hold a project for Carbon. There is no announced public availability date; anything on your roadmap should run on tools you can buy now.
2. If you are on Google's stack, build against the current reachable frontier tier (Argon-era models via the Gemini API) so a later Carbon switch is a model-name change, not a rewrite.
3. If cybersecurity assistance is the draw, treat 'AI patch review' as the realistic near-term use with today's tools — as an assistant to a human reviewer, never as the reviewer.
4. If you are a security team specifically, Fairwind-style gated programs are the actual path — watch Google's official announcement pages rather than aggregator coverage for eligibility openings.
5. Budget note when gated models do open up: announced intro pricing of $2/$10 per million tokens applies to API usage, not chat apps, and post-intro rates of $4/$20 were reported — model your token volume before assuming 'cheap'.

The honest verdict

Gemini 4 Carbon is a significant announced capability with a trivial amount of practical availability — the two facts most coverage kept in separate paragraphs. Google-reported benchmarks are strong, the dual-use caution behind the gated rollout is legitimate, and the pricing is set but not yet actionable for anyone outside the initial defender cohort. Nothing you should change this week: keep building on what you can access, keep every '10x' figure labelled as a vendor claim until independent evaluations exist, and re-check this space when Google opens Fairwind-style access further.

If you're choosing among AI assistants you can actually use today, our framework for picking an AI writing assistant for a solo blog covers the general-assistant decision that is still fully open.