Gemini 4 Carbon: Our next frontier model
AI-assisted digest, checked against the original source before publishing. How we work
DeepMind unveils Gemini 4 Carbon, a cybersecurity- and long-horizon-focused frontier model rolling out first to trusted cyber defenders.
Carbon ranks first on the recently updated v5.0 of the Artificial Analysis Intelligence Index at 59, with notable gains in coding and agentic tasks. In live coding environments used by DeepMind's own teams, the model showed reduced misalignment and fabrication rates, and internal benchmarks showed a strong uplift in secure code generation. The company says it targeted improvements across coding, agentic and scientific work, plus full-stack web development and multimodal understanding, with better output quality in web work lifting the model on WebDev Arena.
On the security side, DeepMind claims state-of-the-art performance across the defense lifecycle of vulnerability detection, validation and remediation. Carbon posts 88.2% on SEC-bench Pro for autonomously surfacing complex flaws in real-world software and 54.8% on ExploitGym for validating which vulnerabilities are actually exploitable — a step defenders use to prioritize remediation. Internal results were also strong: more than a 2x uplift in secure-code success rate versus Argon, while the Chrome team reported Carbon performing over 10X better than another lab's large frontier model when generating high-quality patches.
The long-horizon claim is central to the model's design. DeepMind says its teams have watched Carbon run for days, persisting through dead ends on complex, multi-step engineering work such as root-causing and patching subtle bugs across large codebases and low-level systems — a workload profile aimed directly at autonomous coding agents rather than single-shot completions.
Rollout is deliberately staged. DeepMind is participating in the U.S. government's voluntary pre-release testing process, and starting this week Carbon is going to a subset of cyber defenders, then phasing to additional trusted partners through the Fairwind Program, with guardrails iterated on feedback before reaching developers, businesses and consumers via product surfaces. Pricing is set on par with Argon. What to watch: whether the defender-only phase actually resolves guardrail questions on time, how quickly general availability follows, and whether the security gains hold once Carbon faces adversarial, unvetted traffic rather than curated partner environments.
Key facts
- • Pricing matches Argon: $2 per million input tokens, $10 per million output tokens.
- • Cached input tokens are priced at 95% off the standard input token price.
- • CWE-bench: 71.0% for patching vulnerabilities.
- • Wiz reports Carbon leading web penetration testing benchmarks tied to Scan for Good.
- • WebDev Arena: 1769 Elo, fourth place, a jump of 89 points.
Why it matters
Builders get a frontier model tuned for multi-day agentic engineering and security work at Argon-equivalent pricing, so long-running autonomous tasks become a plausible production pattern rather than a demo. But the defender-first Fairwind rollout means general API access waits on guardrail iteration, so teams should plan around phased availability and validate the security claims on their own wor
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime
