Google announces Gemini 4 Carbon, tops Artificial Analysis index but limits access

Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, with what it describes as state-of-the-art cybersecurity and long-horizon capabilities. Google says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The model arrives roughly two weeks after Gemini 4 Argon.

Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. It says Carbon more than doubles Argon's success rate in generating secure code on internal benchmarks, and that the Chrome team found it more than 10 times better than another lab's large frontier model at producing high-quality patches. Carbon also rose 89 Elo points to fourth place (1768) on WebDev Arena.

Availability is narrow. Starting this week, Carbon goes to a subset of cyber defenders, with access phased in to more trusted partners through the Fairwind Program. Google says it is taking part in the US government's voluntary pre-release testing process and is iterating on guardrails before offering the model to developers, businesses and consumers. It gave no date for that wider release.

Pricing matches Argon's introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper. Artificial Analysis, as summarized by Latent Space, lists standard pricing at $4 and $20 with a 50% launch discount that has no end date.

Carbon is the first model from Google DeepMind to take the top spot on the Artificial Analysis index, according to Latent Space, which puts it ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). The margins are small, though. Ars Technica notes that Carbon beats Argon by slim margins on every test Google supplied, that frontier models score in the same neighborhood, and that Carbon could even be a post-training upgrade of Argon. That is the publication's speculation, not something Google has said.

Independent results are mixed. Latent Space's roundup shows Carbon behind Claude Code with Opus 5.5 on the Coding Agent Index (59 versus 62) and behind GPT-6 Astra, Opus 5.5 and Sonnet 5.5 on FrontierSWE. It leads on several knowledge and reasoning evals, and uses fewer output tokens per task than Opus. Its hallucination rate on Artificial Analysis's Omniscience test rose from 4% to 11% versus Argon, though that remains the second lowest measured. The coding-agent results were measured in Google's Antigravity CLI, so other harnesses are untested.

The staged rollout reflects the dual-use nature of the cyber results. Google frames the gains as defensive, but the same benchmarks measure exploitation, which is why it is routing the model through vetted defenders and government testing before general release.

The main open question is when Carbon reaches developers and consumers. Google has made no promises on timing; Ars Technica suggests that, if it follows Argon, the wait could be a couple of weeks, and that AI Ultra subscribers and Gemini API developers may get it first.

Pricing is also unsettled. Ars Technica reports the discounted rate will double when the introductory period ends, and Google declined to give an end date, suggesting none has been decided. Argon's discounted rates run through the end of November.