[AINews] Gemini 4 Carbon tops leaderboards, launches for cyber defenders

Google DeepMind launched Gemini 4 Carbon on October 27, 2026, roughly two weeks after its predecessor Argon, and positioned it as state of the art in cybersecurity and long-horizon tasks. The company says Carbon goes first to trusted cyber defenders through the Fairwind Program, with wider developer access promised "as soon as possible." Independent evaluator Artificial Analysis places it at number one on its consolidated Intelligence Index v5, which the publisher says put Google DeepMind at the top of that ranking for possibly the first time in its history.
On cybersecurity, Google claims 88.2% on SEC-bench Pro for vulnerability finding, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Those figures come from the model's own launch claims rather than independent runs, so treat them as vendor-reported until third parties reproduce them. Google also attributes safety improvements to Carbon: successful cheating on Terminal-Bench 4.0 fell to 0.0% with cheating attempts at 0.4%, unsupported claims dropped about twofold versus Argon, overreach fell 26% and under-service fell 29%.
The specification is unusually broad for an input side. Carbon accepts text, image, video and speech, emits text, and carries a 1M-token context window with a 1M-token maximum output. Pricing reported by Phil Schmid is $2/$10 per million input/output tokens; Artificial Analysis lists $4/$20 standard pricing with a 50% launch discount that has no stated end date, and cached input at $0.20 per million tokens. The two figures suggest the discounted and standard rates, but buyers should confirm which applies before budgeting.
On Artificial Analysis's Intelligence Index v5, Carbon scores 59, ahead of Opus 5.5's 58, GPT-6 Astra's 57, GPT-6.1 Sol's 55, Sonnet 5.5's 54 and GPT-6.1 Terra's 51. Carbon leads four of the nine constituent evaluations: HLE-Diamond, TB-Science, AA-Omniscience and AutomationBench. The index was rebuilt to run 60% on private tests, which the publisher frames as a guard against benchmark contamination. In its own tracking, Carbon costs $7.76 per index task versus $10.04 for Opus 5.5 and $6.25 for GPT-6 Astra, and burns 123k output tokens per task, 29% fewer than Opus.
Hallucination is the notable regression. Carbon's rate rose from Argon's 4% to 11%, still the second lowest Artificial Analysis has measured, while it leads Omniscience with a score of 54; Opus 5.5 and Astra both sit at 50. For cyber work, a jump of that size deserves attention, and some users have already flagged it. This is my own reading, but a system that helps find and patch vulnerabilities can cause real harm when it invents a finding or misattributes one, so the trade-off between capability and reliability matters more here than on a chat leaderboard.
Coding results are mixed. On the Coding Agent Index Carbon scores 59 using Antigravity CLI, against 62 for Claude Code with Opus 5.5 at max settings. It reaches 51% on FrontierCode v1.1 versus 56% for Opus, and 61.2% on FrontierSWE, behind GPT-6 Astra (65.5), Opus 5.5 (62.3) and Sonnet 5.5 (61.9), while setting a narrow state of the art of 9.6% on SWE-sweep. One caveat: the coding-agent figures were measured in Antigravity CLI, so behavior in other harnesses is untested, which matters for teams wired into a different agent framework.
Knowledge work looks stronger. Carbon hits 1609 Elo on AA-Briefcase, a 126-point gain over Argon, though it still trails Grok 4.7, Muse Wonder and Kimi K3.1 Pro at 1678. On Arena it is first in Text Arena at 1528, second in Agent Arena at a $1.80 median cost per task, and fourth in Code Arena WebDev at 1768; Arena credits the gain over Argon to more praise, more confirmed successes and better steerability. Vals ranks it second on the Vals Index and first on its Finance Index and Legal Research Benchmark, noting 64% more output tokens than Argon, concentrated in finance, tax and legal tasks. Andon Labs places it first on Vending-Bench 2 and on Blueprint-Bench 2, where it ties the human baseline for 3D spatial understanding.
Context from the wider release cycle: Sundar Pichai said on an earnings call that AI is increasing training and launch velocity and expects continued acceleration over the next twelve months. A Gemini engineer described overnight agent loops and swarms inspecting RL data, with launches feeling like periodic checkpoints. Meanwhile skeptics such as scaling01 argue Google remains about two months behind the next OpenAI and Anthropic releases, even with a top index score. Both can be true: leading a consolidated benchmark and leading in production coding agents, where Carbon trails on several coding measures, are different claims.
Why it matters: Model teams and platform engineers now have a Gemini option that leads a consolidated independent index and offers 1M-token input and output, but for now only vetted cyber defenders can use it, so most developers are waiting on the Fairwind restriction to lift. Product teams building security tooling should watch whether the improved vulnerability benchmarks hold up outside Google's own harness, and should factor the higher hallucination rate into any workflow where a false finding triggers automatic action. The cost figures, if they hold, make Carbon competitive per task against Opus 5.5, which is my inference from the published per-task numbers rather than a guarantee for other workloads.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Google’s new frontier model leads independent evals, with access initially limited to trusted defenders.