The second model of Google's Gemini 4 generation, announced October 28, 2026 by Koray Kavukcuoglu, four weeks after Gemini 4 Argon was announced and two weeks after Argon reached general availability. Google calls it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities", built on Argon and shipped on the same safety and security stack. The rollout repeats Argon's: a subset of cyber defenders get it first, then more trusted partners through the Fairwind Program, and developers, businesses and consumers only "as soon as possible", after further guardrail work. Google is again taking part in the US government's voluntary pre-release testing process. Pricing matches Argon's: an introductory $2 / $10 per million input/output tokens, with cached input 95% off, rising to $4 / $20 after the introductory period. Artificial Analysis lists the $4 / $20 rate, a 1M-token context window, and text, image, speech and video input. Architecture, parameter count and training scale are undisclosed, and Google published no model card or evaluation sheet for Carbon at launch.

Artificial Analysis ranks it first. It scores 60 on Intelligence Index v5.0.1 (high reasoning; 59 on v5.0), one point above Claude Opus 5.5 (59) and ahead of GPT-6 Astra (57), GPT-6.1 Sol (55) and Argon (50). AA scored it on v5.0 only, so it has no v4.3.2 reading. The gain over Argon is concentrated in coding and science. On AA's runs it posts the top Terminal-Bench-Science 0.1 score (64.8%, against Astra's 63.3%) and the top HLE-Diamond score (69.1%, against Opus 5.5's 67.9%). On Terminal-Bench 4.0 it reads 65.2%, up from Argon's 57.1% and second only to Kimi K3.1 Pro (67.2%). FrontierCode 2.0 rises to 40.6% from Argon's 25.7%, still behind Opus 5.5 (46.4%) and Astra (42.4%). It trails Astra on the CritPt v2 hard subset (63% to 70.5%) and Opus 5.5 on AA-Briefcase (Elo 1609 to 1800). Its AA-Omniscience accuracy is 58.3% with an 88.6% non-hallucination rate, the calibration pattern Argon set, far above Opus 5.5's 49.7%. It is verbose: 200M output tokens for the index, against a median of 64M, at $7.76 per index task (Opus 5.5: $10.04, Astra: $6.24), with a median first-token time of 97 seconds on Google's API.

Google's own figures centre on cybersecurity, all measured without guardrails: 88.2% on SEC-bench Pro (vulnerability finding), 54.8% on ExploitGym (exploitation) and 71.0% on CWE-bench (patching; Argon's sheet gave 68.0% on CWE-bench v1). It more than doubles Argon's success rate at generating secure code on internal tests, and Chrome's security team found it more than 10× better than "another lab's large frontier model" at producing high-quality patches. Wiz reports it leads on web penetration-testing benchmarks. On WebDev Arena it reads 1768 Elo (fourth, up 89). Google's alignment claims are all relative to Argon: attack success on Gray Swan's indirect prompt injection benchmark falls to 0.0% from 0.7%; successful cheating on Terminal-Bench 4.0 falls to 0.0%, with cheating attempts at 0.4%; high-severity misalignment alerts in live internal coding environments drop by almost 80%; unsupported claims fall by about half; and overreach and under-service fall 26% and 29%. Google says its teams have run it for days at a time on long multi-step engineering work, root-causing and patching subtle bugs across large codebases and low-level systems. All figures in this paragraph are self-reported.

Model Details

Context window 1,000,000
AA Intelligence 60 was 59 on v5.0
License Proprietary (API)

Benchmark Scores

Benchmark Score Mode
SEC-bench Pro 88.2% —
ExploitGym 54.8% —
CWE-bench 71.0% —
WebDev Arena (Elo) 1768 —
frontierreasoningcodingagenticcybersecuritylong-contextmultimodalproprietary

Related