Reviews AI Models
Gemini 4 Carbon Review: DeepMind Frontier Model Built for Cybersecurity and Agentic Tasks
At a glance
Gemini 4 Carbon is Google DeepMind's newest frontier model focused on vulnerability detection, patching and long-horizon agentic work with staged access starting late October 2026.
Gemini 4 Carbon is Google DeepMind's latest frontier model, announced on October 28, 2026. It succeeds Gemini 4 Argon and targets defensive cybersecurity workflows plus multi-day autonomous engineering tasks.
The model leads the Artificial Analysis Intelligence Index v5.0 with a score of 59. It records 88.2 percent on SEC-bench Pro for vulnerability finding and 71.0 percent on CWE-bench for patching. External tests without guardrails place it ahead of Claude Opus 5.5 and GPT-6 Astra on several metrics.
Capabilities, Benchmarks and API Pricing
Gemini 4 Carbon shows more than double the secure code generation performance of Argon and produces over ten times as many high-quality patches according to Chrome team internal results. It gained 89 Elo points on WebDev Arena, reaching fourth place. Context handling reaches 1 million tokens in reported evaluations, supporting improved results on GraphWalks and scientific tasks.
API pricing matches the introductory rates set for Argon: two dollars per million input tokens and ten dollars per million output tokens, with cached input available at a 95 percent discount. Rates are scheduled to double after the introductory period, though no fixed end date has been published for Carbon. Access begins with the Fairwind Program for vetted partners in government and critical infrastructure, followed by phased expansion.
Independent evaluations note second place on the Coding Agent Index behind Claude Opus 5.5. One metric recorded a slight rise in hallucination rate compared with Argon.
Security Measures, Access and Real-World Feedback
Safety claims include a zero percent attack success rate on the Gray Swan indirect prompt injection benchmark and zero successful cheating attempts on Terminal-Bench 4.0. Internal monitoring shows an 80 percent drop in high-severity misalignment alerts and roughly half the unsupported claims compared with the prior model.
Registration for the Fairwind Program requires vetting, and broader availability for developers and consumers remains unspecified beyond the statement that it will arrive as soon as possible after guardrail feedback. The model is not yet present in the public Gemini API or consumer products.
User reports from early partners highlight strong results on vulnerability lifecycle tasks, while some testers note occasional increases in unsupported statements. Complaints center on the narrow initial access window and lack of a clear timeline for wider release.
Competitor Comparison, Limitations and Current Standing
Direct comparisons show Gemini 4 Carbon ahead of GPT-6 Astra on the overall Intelligence Index but behind Claude Opus 5.5 on agentic coding benchmarks. Pricing sits between the two competitors' published API tiers at introduction.
| Model | Intelligence Index | SEC-bench Pro | Intro API Price (in/out) |
|---|---|---|---|
| Gemini 4 Carbon | 59 | 88.2% | $2 / $10 |
| Claude Opus 5.5 | 58 | — | $3 / $15 |
| GPT-6 Astra | 57 | — | $2.5 / $12 |
Limitations include the staged rollout that restricts testing to a small set of partners and the absence of public API endpoints. Some external benchmarks remain incomplete because full access has not been granted. Independent evaluations are ongoing and may alter current rankings once broader data appears.
The model demonstrates measurable gains in defensive security metrics and agentic endurance while carrying the constraints of limited availability and pending wider verification.
Frequently asked questions
When will Gemini 4 Carbon become available to the public?
Access starts with vetted partners through the Fairwind Program and expands in phases. No exact date has been set for general developer or consumer availability.
How does Gemini 4 Carbon pricing compare with earlier Gemini models?
Introductory rates match Argon at two dollars per million input tokens and ten dollars per million output tokens before scheduled increases.
What benchmarks show the largest gains for Gemini 4 Carbon?
SEC-bench Pro vulnerability detection reached 88.2 percent and CWE-bench patching reached 71.0 percent, with internal Chrome team results showing more than ten times the high-quality patches of Argon.
Is Gemini 4 Carbon currently in the public Gemini API?
No. The model remains outside the public API and consumer products during the initial guarded rollout phase.
Filed under
Comments 2