Google ships Gemini 4 Carbon: #1 on Artificial Analysis at 59, built for security work and multi-day tasks

Google launched Gemini 4 Carbon on October 28, positioning it as its strongest model for cybersecurity and long-horizon work. Two weeks earlier, Google had shipped Gemini 4 Argon — Carbon compresses the Gemini 4 series cadence to one frontier model per fortnight. Third-party leaderboards backed the claim on day one: on Artificial Analysis’ Intelligence Index v5.0, Carbon took the top spot at 59 (the site’s own banner reads “Gemini 4 Carbon is the new #1”), ahead of the previous leaders Claude Opus 5.5 at 58 and GPT-6 Astra at 57. The security scores are the headline: SEC-bench Pro (finding vulnerabilities) 88.2%, ExploitGym (exploitation) 54.8%, CWE-bench (patching) 71.0%, with Google adding that Carbon’s misalignment and fabrication rates on secure-code generation also dropped on internal benchmarks.

Key points

• One generation per two weeks: Carbon arrives fourteen days after Argon; Google frames it as built on Argon, aimed at much harder long-horizon engineering problems.
• Runs for days: Google’s internal teams have watched Carbon persist through dead ends for days on multi-step tasks like root-causing and patching subtle bugs across large codebases and low-level systems — the concrete claim behind “long-horizon”.
• Priced at Argon’s level: $2 per million input tokens, $10 per million output tokens, cached input at 95% off the input price. For throughput-bound security-scanning workloads, the cached tier is what makes repeated scans affordable.
• Staged availability, not a full launch: access opens this week to a subset of cyber defenders under the US government’s voluntary pre-release testing process, then widens through the Fairwind Program to trusted partners; no date for general availability.

Context

Carbon continues Google’s element-naming line, following Argon. The security triad is the claimed differentiator: the model must find vulnerabilities (SEC-bench Pro), exploit them (ExploitGym) and write the patch (CWE-bench). Google says safeguards are strengthening in tandem and notes cooperation with the government pre-release process — frontier capability plus offensive/defensive security is becoming the standard release narrative.

Analysis

• The #1 spot has an asterisk: 59 vs 58 vs 57 is a one-point spread on a composite index — top billing is partly a marketing position. The security-benchmark leads are visibly larger and more directional, which is where a buyer should look.
• “Runs for days” awaits independent replication: long-horizon capability is currently Google’s own account; wait for independent long-horizon evaluations before treating it as settled.
• Staged rollout is capability rationing: ordinary users won’t touch Carbon soon; developers should watch the Vertex/AI Studio opening cadence. For security teams, “cyber defenders first” means the offensive/defensive capability gap opens inside a few institutions before it reaches the market.

Summary

Gemini 4 Carbon is Google’s second frontier model in two weeks: security benchmarks and long-horizon work are the pitch, the composite-index lead is thin, pricing is unchanged, and access is staged. Real-world verdicts wait on independent evaluation — the immediate signal is that Google has made two-week model generations its stated rhythm.

Read against this year’s board, the intent is clear: Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Astra have traded the top spot on the composite index, and Google is answering with release density, aiming for a permanent presence on the leaderboard over any single long-lived #1. Each Carbon-class release resets the news cycle and forces competitors to respond on Google’s clock.

Security benchmarks are the fulcrum of that strategy: enterprise buyers pay most readily for the explicit ability to find and patch vulnerabilities, and that is where the three-way composite gap is widest. It is also the segment where staged access reads as prudence rather than limitation — a cyber-defense model released without government pre-release testing would draw scrutiny that a coding model would not. Our earlier coverage of the Oracle–Google TPU partnership recorded the same company positioning on compute distribution in parallel — model cadence on one line, compute supply on the other.