Google DeepMind has released Gemini 4 Carbon, its newest frontier AI model, just two weeks after shipping Gemini 4 Argon. The new model debuts at the top of the Artificial Analysis Intelligence Index v5 with 59 points — a nine-point jump over Argon's 50 — and is built for long-running, autonomous engineering work. Most users will have to wait, though: because of its advanced cybersecurity capabilities, Carbon is going through the US government's voluntary pre-testing process and is initially available only to selected security experts. For continuous coverage of the latest AI developments, follow AI Buzz Wire as the rollout expands.
A security-first rollout
Carbon's unusual release path reflects how capable the model is at both offensive and defensive cyber work. Google says its capabilities span the full cyber defense lifecycle, from detecting and validating vulnerabilities to patching them. Reuters reported the model has a distinct cybersecurity focus, and Ars Technica put it bluntly in its headline: this is a frontier model "that you can't use" — at least not yet, while it sits in the government's pre-testing process for advanced AI systems.
Google has not announced a date for general availability. Pricing, however, is already public: Carbon costs the same as Argon, keeping Google's frontier token rates unchanged through two successive model generations.
Built for multi-day autonomous work
The company says it designed Carbon for "much harder, long-horizon engineering problems" that demand persistent agentic workflows. According to Google, internal teams ran the model for multiple days straight, during which it worked through dead ends on its own and handled complex, multi-step tasks such as root-cause analysis and subtle bug fixes in large codebases and low-level software.
Google also reports improvements in scientific reasoning, multimodal understanding, and full-stack web development based on early user feedback.
Strong overall, still chasing Claude in coding
The benchmark picture is strong, but not a clean sweep. On the Artificial Analysis Intelligence Index v5, Carbon at its highest reasoning tier leads the field mainly through top scores in complex reasoning, knowledge retrieval, and scientific tasks, taking first place in four of the index's nine individual evaluations.
In agentic coding, Carbon still trails Anthropic's Claude Opus 5.5 by three points — 59 versus 62 — even after a big jump from Argon's 53. On the WebDev Arena leaderboard, Carbon gained 89 Elo points to reach 1,769, which places it fourth.
Independent evaluations tell a similar story. The evaluation platform Vals AI ranks Carbon second at 70.36 percent accuracy, behind Claude Sonnet 5.5 at 70.85 percent but ahead of Claude Opus 5.5 at 70.18 percent, with Argon fourth at 69.24 percent.
On Artificial Analysis's AA-Briefcase benchmark for agentic knowledge work, Carbon reached 1,609 Elo, a 126-point gain over Argon's 1,483. It still sits behind Kimi K3.1 Pro at 1,678, Muse Wonder at 1,659, and Grok 4.7 at its xhigh setting at 1,638.
Accuracy up, hallucinations up too
Answer accuracy on Artificial Analysis's evaluations rose from 48 percent with Argon to 58 percent with Carbon. The hallucination rate climbed from 4 to 11 percent — higher, but still the second-lowest in the ranking. For comparison, Claude Opus 5.5 and GPT-6 Astra at their maximum reasoning settings each sit at 50 percent on the same measure, according to Artificial Analysis data reported by The Decoder.
Security benchmarks: a near-perfect prompt injection score
Security is where Google claims its biggest wins. In the company's own benchmark selection, Carbon takes first place in 15 of 16 tests, pulling away from the field especially in long-context and security tasks.
On Gray Swan's indirect prompt injection benchmark, Google says Carbon pushed the attack success rate from 0.7 percent with Argon down to 0.0 percent, blocking every manipulation attempt in the test set. The Decoder cautions that this does not mean prompt injection is solved — the test appears saturated — but the result helps explain why US officials want a pre-deployment look at the model before it reaches the public.
Two frontier models in two weeks
The pace is the message. Two frontier releases inside a fortnight — Argon in mid-October, Carbon now — mark an aggressive cadence from Google DeepMind as it competes with Anthropic and OpenAI for the top of independent leaderboards and for enterprise contracts.
For now, the scoreboard reads like this: Carbon leads the overall index, Claude Opus 5.5 keeps the agentic coding crown, and the strongest cyber-capable model on the market is behind a government review door. When Carbon opens up more widely, unchanged pricing means the real contest will be decided by capability per token rather than discounts — and Google has just raised the bar on that metric.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
