Google DeepMind announced Gemini 4 Carbon on October 28, 2026. Google calls it its strongest model, built for cyber defense and long-running agent jobs, and it now tops the Artificial Analysis Intelligence Index v5.0 with 59 points. Almost nobody can use it. Artificial Analysis lists Gemini 4 Carbon with “No publicly available API”, and from this week only a small group of cyber defenders gets in.
Carbon lands two weeks after Gemini 4 Argon, the model it builds on, and it takes part in the US government’s voluntary pre-release testing.
Where it leads and where it does not
The Decoder reports Google’s own slide puts Carbon first on 15 of 16 tests. The independent lab Artificial Analysis scores Carbon (high) on 59 against 50 for Argon, with wins in four of nine sub-tests. Claude Opus 5.5 still beats it on agentic coding in Antigravity, by three points.
Google’s cyber scores were measured without the model’s safety guardrails:
- SEC-bench Pro: 88.2 percent, for finding vulnerabilities.
- ExploitGym: 54.8 percent, for exploiting them.
- CWE-bench: 71.0 percent, for fixing the root cause.
Carbon leads the Arena.ai Text Arena with 1,528 and reaches number four in WebDev Arena with 1,768. The Vals AI Index v2.2 puts Carbon second on 70.36 percent, behind Claude Sonnet 5.5 on 70.85.
The bill for a long-running agent
Google lists an introductory price identical to Argon: $2 per million input tokens and $10 per million output, with cached input at a 95 percent discount on the input price. Artificial Analysis works from its regular list rate: $4 input, $20 output, ignoring the 50 percent intro discount that runs to the end of 2026.
Per task Artificial Analysis lands at $7.76 for Carbon, against $7.68 for Argon and $10.04 for Claude Opus 5.5 max, 23 percent cheaper than Opus. Carbon writes 123,000 output tokens per task, 29 percent fewer than Opus. Vals measured $10.51 per test and about 48 minutes of runtime. Arena.ai, over 21,900 real sessions, reports a median task cost of $1.80, 72 percent under Opus 5.5 max.
The speed matters just as much. Vals AI found Carbon is slower than Argon, not only better. Its finance and tax tasks take more than twice as long, and a legal research task runs over 50 minutes. Carbon also thinks out loud more, with 2.3 times as many reasoning tokens and 64 percent more output tokens than Argon. That is still less than half the output of Claude Sonnet 5.5 or Opus 5.5.
Agents that run for days spend more tokens and more hardware per job, and it shows in parts: memory prices are up 500 percent in a year because AI data centers have booked 2027 output ahead of everyone else.
The safety numbers are Google’s own
The cyber scores come from external benchmarks run without the safety filters, and Google’s own internal results cannot be checked from outside. Google says Carbon is more than twice as good as Argon at generating secure code, that its Chrome team reports patches more than 10 times better than another lab model, and that serious misalignment alerts fell close to 80 percent. Google also puts cheating on Terminal-Bench 4.0 at 0.0 percent success.
Developers, businesses and consumers are promised Carbon “as soon as possible”. Until then the Fairwind Program decides who is next.
Sources: Google Blog, The Decoder, Artificial Analysis
Leave a comment