Google’s Gemini 4 Carbon takes No. 1, but access is limited

Google has announced Gemini 4 Carbon, a new frontier model it calls its strongest yet, with state-of-the-art results in cybersecurity and long-horizon work.

The catch is who can use it. Carbon is not in the Gemini app or the public API. Google says it is making Carbon available this week to a subset of cyber defenders, then gradually phasing in more trusted partners through its Fairwind Program, with no date given for developers or consumers.

That puts Carbon on the same cautious path as Gemini 4 Argon, which Google also introduced to vetted defenders before widening access.

The benchmark picture

Carbon’s headline result is independent. Artificial Analysis, which tracks models across a mix of tests, ranks it first on version 5.0 of its Intelligence Index with a score of 59, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57, and Gemini 4 Argon at 50.

The rest is mostly Google’s own reporting, with comparison figures as relayed by launch coverage:

Read the table for what it says about direction, not just rank:

• Where Carbon leads. Security and very long inputs. The jump from 84.2% to 96.2% on GraphWalks suggests it holds up much better when a task spans hundreds of thousands of tokens.
• Where it trails. Software-engineering tasks on FrontierSWE 2 and desktop computer use on OSWorld-2.1. Coverage notes these are the only two benchmarks in Google’s set where another model clearly wins.
• Caveats on the security numbers. Google ran the SEC-bench Pro and ExploitGym scores itself using its Antigravity harness, and ExploitGym was run with mitigations off. Vendor harnesses and settings can move agentic results by many points, so treat these as Google’s claims until outside labs reproduce them. Google also publishes its evaluation methodology on its DeepMind evals page.

Why a security-first rollout

Carbon is being offered first to defenders because the same abilities that find and patch vulnerabilities can be used to exploit them. Google describes the Fairwind Program as a vetted channel for governments, national cyber agencies, critical-infrastructure operators and major technology platforms, with background checks for applicants. Google also says Carbon is in the US government’s voluntary pre-release testing.

For most professionals, that means you cannot use Carbon today and should not expect to soon. What you can use is what it signals: Google believes the model is strong enough in security work to justify gated access, and it is tuning guardrails before wider release.

What is still unclear

• Public availability. There is no date for the Gemini app, AI Studio or the API.
• Pricing. One guide lists $2 per million input tokens and $10 per million output tokens, matching Argon’s introductory rates, but with no general access, that figure is unconfirmed.
• Independent replication. Most security and coding numbers come from Google.
• Official video. We did not find an official Google or DeepMind launch video for Carbon to embed, so the announcement pages above are the primary sources.

What this means for working professionals

If you work in security, IT, legal or compliance, Carbon is the clearest sign yet that frontier labs will ship their strongest models to defenders first and the public later. For everyone else, the practical takeaways are smaller but real:

• Long documents. The GraphWalks result points to better handling of very large contexts, which matters for contracts, case files and research corpora, once access arrives.
• Model choice is task-specific. A model that tops an overall index can still lose on computer use or software engineering. Test on your own work before switching tools.
• Treat the first numbers as provisional. Expect revisions once independent labs test Carbon on their own harnesses.