Google’s Gemini 4 Carbon takes benchmark lead with access limited to cyber defenders

Google's Gemini 4 Carbon scored 59 on version 5.0 of the Artificial Analysis intelligence index at its high reasoning setting, topping GPT-6 Astra at 57 and GPT-6.1 Sol at 55, but access is limited to cyber defenders. Google reports Carbon scores 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym and 71.0% on CWE-bench for patching, with those external benchmarks run with guardrails disabled. Artificial Analysis recorded 200 million output tokens across Carbon's index evaluation — about three times the 64 million median — at a weighted average cost of $7.76 per task, with a median 97.5 seconds to first output token on Google's API versus a 3.6-second median for comparable models.

Opus 5.5 has a score of 58 https://artificialanalysis.ai/ intelligence . The gap Carbon follows opened in May. At its I/O conference, Google promised Gemini 3.5 Pro https://blog.google/innovation-and-ai/sundar-pichai-io-2026/ in June. That model slipped repeatedly, and Bloomberg reported its coding performance fell short of internal goals https://www.bloomberg.com/news/articles/2026-07-16/google-gemini-launch-delayed-as-tech-falls-short-of-internal-goals . When Google announced Argon on September 30, a spokesperson told Reuters https://www.reuters.com/legal/litigation/google-announces-gemini-4-flagship-ai-model-after-months-delays-2026-09-30/ that the company no longer plans to release 3.5 Pro and that Argon is larger than Google’s earlier Pro models. Cheaper Flash models filled in, from 3.5 Flash at I/O to 3.8 Flash on September 2, and Gemini 3.1 Pro Preview stayed Google’s newest Pro model until Argon arrived on September 30 https://www.rdworldonline.com/googles-overdue-gemini-4-argon-reaches-the-frontier-with-mixed-results-for-science/ . Carbon scores 59 at its high reasoning setting on version 5.0 of the independent evaluator’s index https://artificialanalysis.ai/models/gemini-4-carbon . In the same Artificial Analysis ranking https://artificialanalysis.ai/leaderboards/models , GPT-6 Astra scores 57 and GPT-6.1 Sol 55 in their listed maximum-reasoning configurations. The index combines nine evaluations spanning coding, scientific reasoning, knowledge and other tasks, among them Terminal-Bench 4.0, FrontierCode 2.0, HLE-Diamond and Terminal-Bench-Science 0.1. Argon scores 50 on the same version, nine points below Carbon. Artificial Analysis had scored Argon at 53 on the earlier v4.3.2 index, so scores carry over only within one index version. Google reports https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ Carbon scores 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench a defensive cybersecurity and AI evaluation benchmark from Collinear AI for patching. Google evaluated these external benchmarks with guardrails disabled, a condition that limits direct extrapolation to the safeguarded product. In its own testing, Google says Carbon more than doubles Argon’s success rate at generating secure code, and its Chrome team found Carbon “more than 10X better” at high-quality patches than another lab’s large frontier model, which Google leaves unnamed. The company also reports fourth place on WebDev Arena at 1,768 Elo, an 89-point improvement. Google built Carbon for long-horizon engineering work and says its internal teams have seen the model run for days, persisting through dead ends on multi-step tasks. On Gray Swan’s indirect prompt-injection benchmark, which Argon previously led at a 0.7% attack success rate, Google reports 0.0% for Carbon. The performance lead comes with substantial token consumption, a trait Carbon shares with Argon, which generated 110 million output tokens https://artificialanalysis.ai/models/gemini-4-argon on the same index, about 1.7 times the median. For Carbon, Artificial Analysis recorded https://artificialanalysis.ai/models/gemini-4-carbon 200 million output tokens across its index evaluation, about three times the 64 million median for comparable models, and calculated a weighted average cost of $7.76 per task using standard pricing. That puts Carbon below Opus 5.5 at maximum reasoning $10.04 and above GPT-6 Astra $6.24 and GPT-6.1 Sol $1.27 . Artificial Analysis measured a median 97.5 seconds before Carbon’s first output token on Google’s API, against a 3.6-second median for comparable models. Google plans introductory rates https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ footnote-1 of $2 per million input tokens and $10 per million output tokens, the same as Argon’s, rising to $4 and $20 after the promotion. Cached inputs receive a 95% discount. The release extends the restricted cyber-defender rollout that accompanied Argon, which Google opened first to its Fairwind Program, launched September 2 for government authorities, critical infrastructure operators and software maintainers. Google plans to give selected defenders access to Carbon this week https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ and expand through Fairwind to more trusted partners. Broader developer, business and consumer access remains a future phase, which Google says will follow feedback and further work on guardrails. Google also says it is taking part in the U.S. government’s voluntary process for pre-release model testing.