Summary: Google announced Gemini 4 Carbon on October 28, 2026, positioning it as its strongest frontier model with state-of-the-art cybersecurity and long-horizon agentic capabilities. It tops Artificial Analysis’ Intelligence Index v5.0 at 59 and posts leading scores on vulnerability-finding, exploitation, and patching benchmarks. Access is staged — cyber defenders first, then developers — at the same intro price as its predecessor, Argon.
Why cybersecurity is the headline
Carbon’s clearest wins are across the cyberdefense lifecycle. On external benchmarks evaluated without guardrails, it scores 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for validating which flaws are exploitable, and 71.0% on Collinear’s CWE-bench for patching at the root cause. Each of those is a stage defenders actually work through: discover, prioritize, fix.
Internal results are also notable. Google reports over 2x uplift versus Argon in generating secure code, and the Chrome team says Carbon produced high-quality patches more than 10x better than another lab’s large frontier model. Wiz, Google’s cloud security platform, observes Carbon leading on web penetration testing benchmarks tied to its Scan for Good initiative.
For product builders, this signals where frontier differentiation is heading: not just general coding ability, but provable strength in a specific, high-stakes workflow. If you build security tooling, Carbon’s benchmark coverage maps almost one-to-one onto a product surface.
Built for long-horizon agentic work
Google describes Carbon as designed for sustained agentic workflows on hard engineering problems. Internal teams have reportedly seen it run for days, persisting through dead ends on tasks like root-causing subtle bugs across large codebases and low-level systems.
The external coding signal is a +89 Elo jump to 1769, landing at #4 on WebDev Arena, with Google citing improved output quality in full-stack web development and multimodal understanding. That trajectory is fast — Carbon arrives just two weeks after Gemini 4 Argon’s rollout, which itself brought chat, tasks, and an agentic IDE into one client. The cadence matters as much as any single score: plan your evaluation and migration loops accordingly.
Alignment numbers worth reading closely
The safety claims are unusually specific, which makes them useful for evaluation planning:
- Successful cheating on Terminal-Bench 4.0 dropped to 0.0%, with cheat attempts at 0.4%, while overall accuracy improved.
- Internal testing showed roughly 2x fewer unsupported claims versus Argon, with the model adding caveats into deliverables rather than hedging.
- Overreach fell 26% and under-service fell 29%, which Google ties to more reliable multi-step execution.
- On Gray Swan’s Indirect Prompt Injection benchmark, Carbon posted a 0.0% attack success rate, down from Argon’s 0.7%.
Take these as vendor-reported internal results rather than independent validation. But they’re testable claims — if your agents depend on scope discipline or injection resistance, they belong in your eval suite.
Pricing and staged rollout
Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. Google is also participating in the U.S. government’s voluntary pre-release testing process.
Availability is the catch. Starting this week, Carbon goes to a subset of cyber defenders, with broader access phased through the Fairwind Program. Developers, businesses, and consumers get it on product surfaces “as soon as possible” — no date given. If your roadmap assumes day-one API access, build in a fallback. Pricing parity with Argon also makes comparison testing cheap once access opens: same budget, two checkpoints, one clear benchmark grid.
The grounded takeaway: Carbon’s benchmark lead is real but vendor-reported, and the staged rollout means most builders should spend the wait assembling an eval harness around the security and long-horizon claims rather than rewriting plans around them.
Sources
View the original post on X