Gemini 4 Carbon is the new #1 model — gated to cyber defenders only
Beats Opus 5.5 and GPT-6 Astra on a harder index, yet has no public API and an 11% hallucination rate — as DeepSeek opens a security-tuned Flash the same week.
Copy markdown
Carbon takes #1 — 59 on a tougher index
Google says Gemini 4 Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, edging Claude Opus 5.5 (58) and GPT-6 Astra (57), and winning 15 of 17 rows in its own benchmark comparison. If you benchmark-shop your model picks, the top of the leaderboard just moved — on paper.
No public API, no GA date — Argon is still the one you call
Access is gated to cybersecurity defenders in Google's "Fairwind" program plus Antigravity internally; there's no public endpoint and no general-availability date. Intro pricing is $2/$10 per million tokens (doubling to $4/$20 later), but for now the Gemini you can actually ship on is Argon.
Smartest on the test, but it hallucinates more
Carbon posts the top AA-Omniscience score (54) yet its hallucination rate climbs from Argon's 4% to 11%, and its headline cyber results were measured "without guardrails." Before wiring a frontier model into agents or retrieval, that's a tradeoff worth testing yourself.
DeepSeek ships an 'abliterated' Flash for security work
DeepSeek V4.1 Flash Cyber (Oct 29) is a refusal-stripped build of V4.1 Flash 1017, aimed at security tooling and agents — vision, tool calls, JSON, ~1M-token context, from $0.30/$1.00 per million via gateways. If red-team or security automation keeps hitting refusals, it's a cheap, long-context option built to comply.