Google announced Gemini 4 Carbon on October 28, 2026, and the new model took first place on a freshly revised benchmark ranking a day later. Artificial Analysis places Gemini 4 Carbon first on Intelligence Index v5.0 with 59 points, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Argon scores 50 on the same index.
The lead is narrow, and the index itself is new. Artificial Analysis launched Intelligence Index v5 on October 27, 2026, with Claude Opus 5.5 in first place at 58 points. Gemini 4 Carbon arrived a day later and moved one point ahead. Because the test suite was rebuilt, v5 scores cannot be compared with older index scores. The ranking combines nine evaluations, covering knowledge work, coding, science and general capability.
The main change in v5 is how much of the test is hidden. Artificial Analysis said 60 percent of the revised index uses private tests or withheld answers, a change intended to reduce benchmark contamination. Contamination means a model has effectively seen test questions during training. The revision adds FrontierCode 2.0 and Terminal-Bench-Science and replaces HLE with HLE-Diamond.
Gemini 4 Carbon task results
Carbon leads four of the nine evaluations, including the scientific workflow test, according to reporting on the ranking. Claude Opus 5.5 still leads repository coding, and GPT-6 Astra leads research physics. No single model leads all nine, so a buyer who cares about one kind of work should look at the evaluation that matches it.
On the separate Coding Agent Index, which tests models inside coding tools, Carbon in Google’s Antigravity CLI scores 59, behind Claude Code with Opus 5.5 at 62. That index has three parts: FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA. Artificial Analysis places Carbon second on the first two, with 51 percent and 64 percent, and its SWE-Atlas-QnA-AA result is 64 percent, behind Muse Code with Muse Wonder Full at 69 percent.
Terminal-Bench-Science rises from 52.9 percent for Argon to 64.8 percent for Carbon, and Terminal-Bench 4.0 rises from 57.1 percent to 65.2 percent. On AA-Briefcase, a knowledge work evaluation, Gemini 4 Carbon reaches 1,609 Elo, a gain of 126 points over Argon, but it trails Kimi K3.1 Pro at 1,678 Elo and Claude Sonnet 5.5 at 1,812 Elo.
Gemini 4 Carbon cybersecurity focus
Cybersecurity is the clearest focus of the announcement. Google splits security into finding a vulnerability, checking exploitability, and producing a patch. Google reported 88.2 percent on SEC-bench Pro for vulnerability discovery, up from 83.4 percent for Argon. It reported 54.8 percent on ExploitGym, up from 42.0 percent, and 71.0 percent on CWE-bench for patching, up from 68.0 percent.
Those three steps are separate skills. Finding suspicious code does not prove a defect can be exploited, and demonstrating an exploit does not prove a proposed repair works. Google said the external evaluations were conducted without guardrails, so customers may not obtain the same scores under deployment restrictions.
Google also reported internal reliability gains. The company said Carbon roughly halves unsupported assertions in internal testing, compared with Argon, and severe alignment alerts fell almost 80 percent in its internal coding environments. At the same time, an external hallucination measure on the AA-Omniscience test rose from 4 percent for Argon to 11 percent for Gemini 4 Carbon, counting wrong answers among responses that were not fully correct. Google said its teams have seen Carbon run for days on long engineering tasks.
Gemini 4 Carbon access and pricing
Access to Gemini 4 Carbon is limited to selected cybersecurity defenders in Google’s Fairwind Program, leaving most developers and businesses waiting for a broader release. Google said the broader release will come as soon as possible after gathering feedback and iterating on guardrails, and after voluntary United States government testing.
Google said Carbon will launch at $2 per million input tokens and $10 per million output tokens, with a 95 percent discount for cached inputs. Those rates match Argon’s introductory rates and are temporary. Later prices are $4 per million input tokens and $20 per million output tokens, with no end date specified for the introductory period.
Artificial Analysis marks Carbon as lacking a publicly available application programming interface, which matches the limited rollout. Argon, opened more broadly on October 13, remains the model most customers can evaluate. The practical question for most teams is when Fairwind access widens and whether the security scores hold once guardrails are in place.