Gemini 4 Carbon

• Oct 28, 2026

  Google releases Gemini 4 Carbon, its strongest model, to cyber defenders first

Google's second Gemini 4 model, announced 2026-10-28 two weeks after Argon and described as building on it; access starts with a subset of cyber defenders and phases out to more trusted partners through the Fairwind Program, with developers, businesses and consumers promised access 'as soon as possible', and the model is in the U.S. government's voluntary pre-release testing — hence availability "restricted". Pricing is Google's introductory $2/$10 per MTok (cached input 95% off), rising to $4/$20; Google's blog gives no end date for the discount. Google's evaluation PDF (deepmind.google/models/evals-methodology/gemini-4-carbon, highest thinking setting) uses only untracked suites: Artificial Analysis Intelligence Index v5.0 59 (Argon 50, GPT-6 Astra 57, Claude Opus 5.5 58), Vals Index v2.2 70.4%, SWE-sweep 9.6%, FrontierSWE 2 61.2%, Terminal-Bench 4.0 65.2% (not the tracked 2.1), Agents' Last Exam v1.1 77.4%, OSWorld 2.1 offline 81.5%, AutomationBench 53.9%, Vending-Bench 2 $16,484, Chartography 72.5%, Blueprint-Bench 2 58.6%, GraphWalks 256K–1M 96.2%, RiemannBench 80.0%, Terminal-Bench Science 0.1 64.8%, CWE-bench 71.0%, ExploitGym 54.8% and SEC-bench Pro 88.2% (183-task May 2026 JavaScript-engine split, pass@1 over 3 attempts); the ExploitGym and SEC-bench Pro runs are Google's own, some with mitigations off. No GPQA Diamond, SWE-bench, HLE, ARC-AGI-2 or GDPval-AA v2 figure was published. LMArena Elo 1528 (±9, preliminary, 4,403 votes) is the 'Gemini 4 Carbon (High)' text-arena listing, rank 1 ahead of Argon (now 1520); it is 4th on WebDev Arena at 1768. costPerTask $7.76 and output speed 99.8 tok/s are Artificial Analysis' figures for the only variant it lists, Carbon (High), and AA prices the index run at the $4/$20 list rate, not the introductory one (so the per-task cost at launch pricing is about half). AA's FAQ gives a 103.13s time to first token on Google's API but its first-answer-token chart shows no value, so timeToFirstTokenSec is null. contextWindow, maxOutput and knowledgeCutoff are null: Google published no model card or Gemini API model page, and AA's '1M' context is not an official figure. No official model string has been published, so apiIds is empty. predecessorId is null: Google says Carbon builds on Argon but names no model it replaces.