Google DeepMind announces Gemini 4 Carbon, releasing it first to a limited set of cyber defenders at Argon's introductory API price
Google DeepMind announced Gemini 4 Carbon, two weeks after Gemini 4 Argon. This week it becomes available only to a subset of cyber defenders. Access will widen gradually to trusted partners through the Fairwind Program. General availability to developers, businesses and consumers is "as soon as possible" with no date. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period it rises to $4 and $20. Google self-reports these results: 59 on Artificial Analysis Intelligence Index v5.0 (#1), 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench, #4 on WebDev Arena (1768, +89 elo), 0.0% successful cheating on Terminal-Bench 4.0, and 0.0% attack success on Gray Swan IPI. Google also says it is engaged in the U.S. government's voluntary pre-release testing process.
Entities: Google DeepMind, Google, Gemini 4 Carbon, Gemini 4 Argon, Koray Kavukcuoglu, Fairwind Program
1 primary
What happened
Google DeepMind announced Gemini 4 Carbon, two weeks after Gemini 4 Argon. This week it is available only to a subset of cyber defenders, with access widening gradually to trusted partners through the Fairwind Program. General availability has no date ("as soon as possible"). Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later. All benchmark figures are Google's own, including 59 on the Artificial Analysis Intelligence Index v5.0, 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench.
Why it matters
For now the model is usable by very few people, so most developers and businesses cannot act on it. Security teams outside the gated group are the most affected, because a model said to be strong at both finding and exploiting vulnerabilities is being released to defenders first. That makes access a real competitive and security question. The planned price doubling after the introductory period matters for anyone budgeting around it. Its role in US government voluntary pre-release testing suggests regulators will be watching, but Google's description does not show how much that testing influenced the release.
What is noise
"Strongest model", "state-of-the-art" and "improved alignment and safety" are Google's own claims, and no independent party has checked the benchmark scores. The 0.0% cheating and 0.0% attack success figures are strikingly clean and depend heavily on how the tests were defined. The comparison to an unnamed lab's model, and the framing about the "pace of the frontier", are marketing. Missing context includes who the defenders are, how many there are, and whether the #1 ranking on one index holds up elsewhere (it is only #4 on WebDev Arena).
Watch next
- 01Independent evaluations of Carbon on cyber benchmarks and the Artificial Analysis index, to see whether the self-reported 88.2% on SEC-bench Pro and 59 index score hold up
- 02A general availability date, and whether the price actually moves from $2/$10 to $4/$20 per million tokens when the introductory period ends
- 03Reports from Fairwind Program partners or government testers, and any competitor response such as a similarly gated cyber model, showing whether staged access becomes a standard pattern
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680
- OpenAI launches GPT-6 Sol and Luna at half the token price of GPT-5.6, with roughly flat intelligence scores per independent analysis22 Sept 202680