# Google ships Gemini 4 Carbon to cyber defenders first, at half the price of Claude Opus 5.5

Google's own benchmark table puts Carbon one point ahead of Claude Opus 5.5 on the Artificial Analysis Intelligence Index, but developers and businesses cannot use it until Google finishes a vetted-access phase that has no end date.

- Published: 2026-10-29T06:14:22.974Z
- Canonical: https://polylog.news/ai/2026-10-29/google-ships-gemini-4-carbon-to-cyber-defenders-first-at-hal
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Google DeepMind](https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/), [Polylog editors](https://polylog.news), [Unite.AI](https://www.unite.ai/staged-rollout-sends-googles-gemini-4-carbon-to-cyber-defenders-first/), [Vals AI](https://www.vals.ai/models/google_gemini-4-carbon)

Google DeepMind announced [Gemini 4 Carbon](https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/) on Oct. 28. Koray Kavukcuoglu, a senior vice president at Google DeepMind and Google's chief AI architect, called it the company's strongest model, with state-of-the-art cybersecurity and long-horizon capabilities, [according to Unite.AI](https://www.unite.ai/staged-rollout-sends-googles-gemini-4-carbon-to-cyber-defenders-first/). Carbon follows Gemini 4 Argon, which Google announced on Sept. 30 and rolled out widely about two weeks before Carbon.

The release terms are the main change. Google will give Carbon first to a subset of cyber defenders and then widen access to trusted partners through its Fairwind Program. It said it will keep working on guardrails before it offers the model to developers, businesses and consumers ([WinBuzzer](https://winbuzzer.com/2026/10/28/gemini-4-carbon-ai-leads-test-for-fixing-vulnerable-code-xcxwbn/)). Under the Fairwind terms Google set for Argon, member organizations may give access only to internal security, incident-response or penetration-testing teams, and Google alone decides who qualifies ([TechWyse](https://www.techwyse.com/news/ai-search/gemini-4-argon-restricted-launch-cyber-defenders), [Pebblous](https://blog.pebblous.ai/blog/gemini-4-argon-fairwind-vetted-access/en/)). Carbon has a 1-million-token context window. Its introductory API price is $2 per million input tokens and $10 per million output tokens, and Google plans to raise it to $4 and $20 on a date it has not announced ([eesel](https://www.eesel.ai/blog/gemini-4-carbon-pricing)).

Google frames Carbon's security value around three stages of defensive work: finding a flaw, proving it can be exploited, and fixing it. It reported 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. Google produced all three figures itself. The ExploitGym paper reported that the strongest configurations it tested, Anthropic's Claude Mythos Preview and OpenAI's GPT-5.5, produced working exploits for 157 and 120 of its 898 instances ([arXiv](https://arxiv.org/abs/2605.11086)). Google's figure is substantially higher than those results, but the two come from different test setups (harnesses), so the comparison is not direct. On patching, Google earlier reported 68% for Argon on CWE-bench v1, a benchmark that Collinear AI maintains and runs itself. If the benchmark versions match, Carbon gains about three points over Argon.

The lead in general capability is narrow. The Russian-language Telegram channel [AI ML Big Data](https://t.me/ai_machinelearning_big_data/11237) reported that Carbon ranked first on the Artificial Analysis Intelligence Index v5.0 with 59 points. Google's comparison table gives Claude Opus 5.5 58 points and OpenAI's GPT-6 Astra 57 ([Kingy AI](https://kingy.ai/blog/gemini-4-carbon/)). In independent testing, [Vals AI](https://www.vals.ai/models/google_gemini-4-carbon) ranks Carbon second of 29 models on its index at 70.36%, fourth of 57 on Terminal-Bench 4.0 at 63.13%, and eighth of 121 on Vibe Code Bench. Google's widest reported margin is in long-context retrieval: 96.2% on GraphWalks between 256,000 and 1 million tokens, against 71.8% for GPT-6 Astra and 66.8% for Opus 5.5.

Price is the clearer advantage once access opens. Anthropic charges $4 per million input tokens and $20 per million output tokens for Opus 5.5 ([LLM Stats](https://llm-stats.com/models/claude-opus-5-5)). Carbon's introductory rate is half of both, for a model that scores one point higher on the Artificial Analysis index. Open-weight competition continues: Moonshot AI released Kimi K3.1 Pro on Oct. 23 and promised to publish its weights by Nov. 4 ([Wavect](https://wavect.io/blog/open-weight-llm-comparison-2026/)).

## What this means

Google is using a cyber-capability claim to justify vetted access. That gives Google control over who gets its strongest model first, and the vetted groups are large security teams that are likely to become early enterprise customers. The one-point lead on the index is within the range that reruns and harness changes can move, so the framing of a new "number one model" overstates what changed. Two things are new. Releasing the general frontier model to cyber defenders first is now Google's default sequence, after Argon. And the $2/$10 introductory price is aimed directly at Anthropic, which gets 80% of its revenue from metered usage, according to its stock-listing filing. The closest precedent is OpenAI's staged release of GPT-2 in 2019: OpenAI withheld the full model, then published it within about nine months with no major documented misuse. Gating did not hold back capability for long, but it did set a norm. The likelier outcome is that Google opens Carbon to the general API within weeks to a few months at the introductory price, as Argon moved from restricted to wide release in about two weeks. That depends on no public misuse incident coming out of the Fairwind group. The read is wrong if Google keeps Carbon restricted into next year or sells general users a separate version with weaker cyber capability. If Carbon opens at $2/$10, expect Anthropic or OpenAI to answer on price within the 30-day horizon of the frontier price-war thread.

## What to watch

- When Carbon reaches the general Gemini API and whether it keeps the $2/$10 introductory rate, because that decides whether the price pressure on Opus 5.5 is real or only announced.
- Whether Collinear AI or another outside group publishes its own Carbon runs on CWE-bench or ExploitGym, which would show whether Google's cyber scores hold up outside Google's own harness.
- Whether Moonshot AI publishes Kimi K3.1 Pro weights by its Nov. 4 promise, since a strong open-weight release in the same weeks would set a lower price floor than either US lab.