Gemini 4 Carbon Scores 59 on the Intelligence Index, but Only Cyber Defenders Get It First

News Summary

Google announced Gemini 4 Carbon on October 28, 2026 (about 10:10 AM Pacific Time), roughly two weeks after the broader release of Gemini 4 Argon. Google DeepMind describes Carbon as its strongest model so far, built for long-horizon engineering work that needs an AI system to keep going for hours or days. It is not publicly available yet. Selected cybersecurity defenders get access first, and developers, businesses and consumers will follow later, with no date given.

What Google Announced

Carbon is a new model in the Gemini 4 family, announced on October 28, 2026, two weeks after Gemini 4 Argon became globally available on October 13, 2026 (Pacific Time announcements, as reported by industry outlets). Google says Carbon is aimed at much harder, long-horizon engineering problems that need persistent, agentic workflows, meaning the model plans, tries, checks its own work and continues across many steps.

Google says internal teams ran Carbon for multiple days on complex tasks, and that the model recovered from dead ends on its own instead of stopping. The company also reports gains in scientific reasoning, multimodal understanding and full-stack web development. It describes Carbon as able to trace subtle bugs and patch them across large codebases and low-level systems.

Staged Access Through the Fairwind Program

Carbon is not a general launch. Selected cyber defenders receive access during the week of the announcement. More trusted partners are then being added through Google DeepMind's Fairwind Program, which reportedly includes more than 650 partners worldwide. Priority goes to governments and national cybersecurity authorities, critical infrastructure operators and core technology platforms, and academic labs focused on defensive benchmarking can apply. Applicants go through security and ethics vetting.

Google says it will gather feedback and refine safeguards before widening access to developers, businesses and consumers. It is also taking part in the U.S. government's voluntary pre-release testing process. A third-party check of a paid Gemini API key on the evening of October 28, 2026 (India Standard Time) listed Argon but not Carbon, which fits the limited rollout.

Benchmark Results

On the Artificial Analysis Intelligence Index v5, Carbon scores 59, up from 50 for Argon, which puts it at the top of the index according to The Decoder. On the Artificial Analysis Coding Agent Index it also scores 59 versus 50 for Argon, though Claude Opus 5.5 still leads that particular index at 62.

Other reported numbers:

• WebDev Arena: up 89 Elo to roughly 1,769, fourth place.
• Text Arena: first place at 1,528 points versus 1,520 for Argon.
• Agent Arena: second place across more than 21,900 sessions, with a median task cost of $1.80.
• AA-Briefcase (agentic knowledge work): 1,609 Elo, up 126 from Argon.
• Vals AI: 70.36% accuracy, second overall, and first on its Finance Index and Legal Research benchmark.

Artificial Analysis also noted trade-offs. Answer accuracy rose from 48% to 58%, but the hallucination rate also rose from 4% to 11%. Carbon generates 2.3 times Argon's reasoning tokens, and some agentic tasks such as finance and tax work take more than twice as long. Its cost per task is $7.76 versus $7.68 for Argon.

Cybersecurity Focus

Google's own table shows Carbon first in 15 of 16 cybersecurity tests. The headline results are 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym at ten hours of runtime (validating exploitation) and 71.0% on CWE-Bench (patching the root cause of flaws). Google states these evaluations were run without guardrails, so they show raw capability and not behavior under full safety protections.

Google also makes internal claims that Carbon more than doubles Argon's rate of producing secure code, and that its Chrome team saw more than a tenfold improvement in high-quality patches compared with another lab's frontier model, which was not named. These internal claims have not been independently verified.

Safeguards and Safety Measures

Carbon keeps the layered protections introduced with Argon: real-time misuse defenses, monitoring of the model's reasoning, and resistance to prompt injection. Google reports that on Gray Swan's indirect prompt injection benchmark no evaluated attack succeeded, down from a 0.7% success rate for Argon, though this only covers the attacks tested. On Terminal-Bench 4.0 there were no successful cheating cases, with cheating attempts at 0.4%.

In live coding with internal teams, high-severity misalignment alerts fell by almost 80%, and Google says Carbon makes roughly half as many unsupported claims as Argon. The intended behavior is to disclose missing evidence or blocked actions instead of inventing results.

Pricing

Introductory API pricing is $2 per million input tokens and $10 per million output tokens, matching Argon, with cached input at a 95% discount. Later list pricing is $4 and $20 per million tokens. Artificial Analysis notes the introductory 50% discount runs through the end of 2026.

What to Watch

Nearly all performance figures come from Google or from benchmark operators and were not independently reproduced for this report. The key open questions are when Carbon reaches the general developer API, how its guardrails change results once the cybersecurity tests are repeated with protections on, and whether the higher accuracy justifies the higher hallucination rate and slower runtimes on long agentic tasks.