Gemini 4 Carbon: Benchmarks, Pricing, Access, Release

Google has announced Gemini 4 Carbon, a new frontier model that the company says now ranks first on the most widely cited independent leaderboard for large language models. Announced on October 28, 2026, Carbon scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, one point ahead of Anthropic’s Claude Opus 5.5 and two points ahead of OpenAI’s GPT-6 Astra, according to Google and Artificial Analysis.

Most people cannot use it yet. Following the template set by its predecessor, Carbon is rolling out first to vetted cyber defenders, with broader access for developers, businesses and consumers promised but not dated.

The benchmark lead

According to Google’s announcement blog and Artificial Analysis data reported by VentureBeat, Neowin and others:

• Gemini 4 Carbon: 59
• Claude Opus 5.5: 58
• GPT-6 Astra: 57
• Gemini 4 Argon: 50 (high reasoning)

The index was recently updated to version 5.0 with harder evaluations, and scores do not carry over between versions. Gemini 4 Argon, released at the start of October, scored 53 on the earlier v4.3.2 index but 50 on v5.0. That makes the jump from Argon to Carbon roughly nine points on a like-for-like basis, a large gain for a model released about four weeks later.

Artificial Analysis also reported that Carbon posts the top score on its AA-Omniscience knowledge benchmark at 54. The same testing found Carbon’s hallucination rate rose to 11%, up from 4% for Argon, according to coverage of the results.

Several outlets noted that the margin over rival models is small and that rankings on aggregate indexes frequently shift as labs ship updates.

Coding and agentic results

Google described “significant gains across coding and agentic tasks,” according to 9to5Google. The detailed numbers are mixed against competitors:

• FrontierSWE 2: Carbon scored 61.2%, behind GPT-6 Astra (65.5%) and Claude Opus 5.5 (62.3%).
• Terminal-Bench Science: Carbon scored 64.8%, ahead of Astra (63.3%) and Opus 5.5 (59.0%).
• WebDev Arena: Carbon reached 1769 Elo, ranking fourth.
• Terminal-Bench 4.0: Google said Carbon reduced successful cheating on the benchmark to 0.0%, addressing a known problem in which agentic models exploit evaluation harnesses rather than solving tasks.

Wccftech reported that Carbon led every other model on all but two of Google’s published benchmarks.

Cybersecurity is the headline capability

Google is positioning Carbon primarily as a security model. The company reported:

• SEC-bench Pro (vulnerability discovery): 88.2%
• CWE-bench: 71.0%
• ExploitGym: 54.8%

Google said cyber benchmarks were run without guardrails, reflecting the configuration offered to vetted defenders rather than the public. Wccftech, citing Google’s figures, reported that Carbon performed more than ten times better than any publicly available model at generating high-quality security patches.

Those capabilities explain the staged launch. Models that can find and exploit vulnerabilities at this level are dual-use, and all three leading labs have moved to restrict their most capable cyber features. OpenAI took a similar tiered approach, as covered in OpenAI’s Astra restricted cyber release.

Access: defenders first

According to Google, the rollout proceeds in stages:

1. A subset of cyber defenders receives access first, beginning the week of the announcement.
2. Access expands to more trusted partners through Google’s Fairwind Program.
3. Developers, businesses and consumers follow “as soon as possible,” with no date given.

Artificial Analysis’s chart marks Carbon as lacking a publicly available API, according to R&D World. That mirrors Argon, which spent about two weeks in the Fairwind preview before opening to paid Gemini subscribers and API customers on October 13, as detailed in our Gemini 4 Argon rollout coverage.

Pricing

Reports on pricing are not fully consistent. Multiple outlets report an introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%, rising to $4 and $20 after an introductory period that Google has not dated. Artificial Analysis lists $4/$20 as standard pricing with a 50% launch discount, which works out to the same effective rate.

If those figures hold, Carbon would launch at the same introductory price as Argon, which Neowin framed as undercutting rival flagships “at a fraction of the price.” Google has not published separate pricing for the unrestricted defender variant, according to MindStudio’s analysis.

Speed and cost to run

Artificial Analysis put the cost of running its full index on Carbon at $7.76 per task. It also measured a time to first token of 97.5 seconds on Google’s API, compared with a median of 3.6 seconds for comparable models, reflecting the extended reasoning the model performs before answering. For long-horizon agent work that latency matters less; for interactive chat it would be a significant constraint at launch settings.

Where it fits

Carbon is the second model under the Gemini 4 name and arrives as competition at the frontier has compressed to single-digit index points. Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra had traded the top position through the autumn. The announcement also came the same week Alphabet reported third-quarter results in which Google Cloud revenue grew 103%, while Search growth slowed as AI Mode monetized at lower rates than traditional search.

What it means

Google has regained the benchmark lead, narrowly. A one-point margin on an aggregate index is within the range that a single rival update can erase. The more meaningful signal is the nine-point jump from Argon to Carbon in about a month, which suggests Google’s release cadence has accelerated to match or exceed its rivals.

Security-first launches are becoming the norm for frontier models. Google, OpenAI and Anthropic have all converged on staged releases in which the most capable cyber features go to vetted defenders first. That gives defenders a head start on patching, but it also means the public benchmark leader is, for now, a model the public cannot use.

The tradeoffs are real. A higher hallucination rate than Argon and a time to first token measured in tens of seconds are notable costs. Enterprises evaluating Carbon for production will weigh accuracy on knowledge tasks and responsiveness alongside raw index scores.

What to watch next. The date Carbon reaches paid API users and Gemini subscribers; whether the introductory pricing has a fixed end date like Argon’s November 30 deadline; independent replication of Google’s cyber benchmark figures; and responses from Anthropic and OpenAI, both of which have shipped point releases within weeks of competitor launches this year. The winners in the near term are the security teams inside the Fairwind Program; developers will have to wait.