Google DeepMind released its new flagship model, Gemini 4 Carbon, on 10 28, just two weeks after the previous generation, Gemini 4 Argon. The company claims significant improvements in cybersecurity (finding vulnerabilities, verifying exploitability, and patching) and long-horizon programming tasks that require running continuously for several days. Access is currently limited to a subset of cyber defenders, with plans to gradually expand to trusted partners through the Fairwind program; no timeline has been provided for when developers, enterprises, and general users will gain access.
Confidence: High (100) · Event Date 2026-10-28
Rationale
- Third-party confirmed rankings: In Artificial Analysis's Intelligence Index v5, Carbon (high reasoning tier) scored 59 points, ranking first (Argon scored 50), and took first place in four of nine benchmarks; on the Agentic Coding Index, it scored 59, still trailing Claude Code + Claude Opus 5.5's 62.
- Security evaluations (officially reported, with guardrails disabled during testing): Carbon leads in SEC-bench Pro for autonomously finding real software vulnerabilities (88.2%), ExploitGym for turning vulnerabilities into usable exploits (54.8%, and CWE-bench for patching (71.0%. In internal tests, the success rate for generating secure code was more than double that of Argon; the Chrome team stated it writes high-quality patches better than a large frontier model from another lab by more than 10 times (the lab was not named).
- On alignment, the company claims: On Terminal-Bench 4.0, the cheating success rate dropped to 0.0%, with attempted cheating at 0.4%; in internal real-world coding environments, high-severity misalignment alerts decreased by nearly 80%; unfounded assertions were about half as frequent as in Argon; and on Gray Swan indirect prompt injection tests, the attack success rate was 0.0% (Argon: 0.7%).
- Artificial Analysis's independent testing also revealed trade-offs: Knowledge question-answering accuracy rose from Argon's 48% to 58%, but the hallucination rate increased from 4% to 11% (still the second lowest on the leaderboard); at standard pricing, each exponential task costs 7.76 dollars, approximately 23% lower than Claude Opus 5.5 (max) at 10.04 dollars.
Usage
- Not yet directly available. The company states that starting this week, access will first be granted to a subset of cyber defenders, then gradually expanded to more trusted partners through the Fairwind program, with feedback collected and guardrails adjusted before rolling out to developer, enterprise, and consumer products.
- Officially announced pricing: At launch, promotional pricing identical to Argon applies, at 2 dollars per million input tokens and 10 dollars for output, with a 95% discount on cached input prices; after the promotional period, prices will be 4 dollars / 20 dollars.
Risks
- Most figures come from Google's selected benchmarks and internal tests; the Chrome team's "10 times" comparison did not specify which model was used, making verification impossible.
- Ars Technica notes that outsiders cannot determine whether Carbon is a completely new model or a post-training upgrade of Argon, as various frontier models differ by only a few points on these benchmarks.
- Reports on when the promotional pricing ends vary: Ars Technica states Google refused to provide a date, Artificial Analysis writes "no definite end date," and The Decoder reports it ends by the end of 2026.
- Google is participating in the U.S. government's voluntary pre-release testing process, and the timeline for public availability remains undetermined.
Open · Google Blog · Artificial Analysis (Independent Evaluation) · Ars Technica · The Decoder · The Verge · SiliconANGLE