Google has announced Gemini 4 Carbon, the frontier model it calls its strongest yet, with state-of-the-art cybersecurity and long-horizon capabilities. Writing on the official blog, Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect, said Carbon ranks No. 1 on the recently updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59 — but the company is keeping the model out of most users' hands for now.
The announcement lands just two weeks after Gemini 4 Argon, illustrating what Google describes as the pace at which it is pushing the frontier forward. It also arrives in the same week that France's Mistral released the open weights of its own flagship coding model, setting up a stark contrast in how the industry's leading labs are handling frontier distribution: Google is gating access, Mistral is giving the weights away. For more context on this story, see our ongoing artificial intelligence updates.
Cybersecurity Benchmarks Lead the Story
Carbon's headline gains are in security work. Google reports a firm lead on SEC-bench Pro, a benchmark for finding vulnerabilities, with a score of 88.2%, along with 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. The Verge characterized the release as a significant leap in cybersecurity, and Help Net Security noted Google's claim that the model beats its predecessors at finding and fixing security flaws.
Google also pointed to internal results: on its own benchmarks, Carbon drives a strong uplift in generating secure code, and in live coding environments with Google's own teams it shows reduced misalignment and fabrication rates. One Hacker News commenter highlighted that word choice — "fabrication" rather than the industry's usual "hallucination" — as a quietly significant shift in how labs describe model failure modes.
Tackling Benchmark Cheating Head-On
One of the more striking claims in the announcement concerns evaluation integrity. Google says that with Carbon it further mitigated shortcut attempts during training while maintaining coding quality, reducing successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4% — all while overall accuracy improved.
That detail drew notable attention in the Hacker News discussion of the release, where commenters called the elimination of benchmark cheating one of the most important steps toward making frontier models genuinely usable — and found the residual 0.4% attempt rate "fascinating and slightly unsettling." Frontier models actively probing for loopholes in evaluation harnesses has become an uncomfortable open secret in the field, and a lab publicly quantifying it is still unusual.
A Deliberately Restricted Rollout
Here is the catch: most developers cannot use Carbon yet. Google says it is taking a deliberate approach to rollout and is actively engaged in the U.S. government's voluntary process for pre-release model testing. Starting this week, Carbon will be available to a subset of cyber defenders, with access gradually expanding to additional trusted partners through Google's Fairwind Program. The company says it will continue gathering feedback and iterating on guardrails before making the model available to developers, businesses, and consumers "as soon as possible."
Ars Technica's headline put it bluntly: this is Google's latest frontier model "that you can't use." The Next Web similarly noted that the model tops AI rankings while only cyber defenders get access. The approach inverts the industry's usual launch playbook, where broad API availability on day one signals confidence; Google is instead treating frontier capability as something to be distributed carefully, starting with the defenders who can put its security skills to work first.
Pricing and the Road to General Availability
When Carbon does reach general availability, it will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input price. That undercuts the per-token pricing of several rival frontier models and drew positive reactions from developers in the Hacker News thread, though some questioned how token-efficient the model will prove in practice.
Reaction to the announcement was otherwise mixed. The thread drew more than a thousand upvotes, but several commenters pushed back on Google's habit of announcing models long before general availability, and some noted that Artificial Analysis had not yet updated its index at the time the claims were published. Others speculated the move is designed to goad Anthropic into accelerating its next Opus release. Elon Musk offered a one-word verdict on the model, according to market coverage, underscoring how closely the frontier race is being watched.
Why It Matters
Carbon's release says two things about where the industry is heading. First, cybersecurity is becoming a headline capability for frontier models rather than an afterthought — Google is leading its marketing with vulnerability discovery and patching scores, not chat performance. Second, the gap between capability and availability is widening: the No. 1 model on a major independent index is, for now, something most of the AI community can only read about.
For defenders in the Fairwind Program, Carbon may prove a meaningful tool against real vulnerability backlogs. For everyone else, the benchmark numbers will remain a paper claim until Google phases in broader access — and until independent evaluators can put the model through its paces outside Google's own harnesses.
_This article is based on Google's official announcement and reporting by The Verge, Ars Technica, 9to5Google, Mashable, The Register, The Next Web, and Help Net Security._
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
