Google DeepMind announced Gemini 4 Carbon on October 28, 2026, calling it its strongest frontier model, with state-of-the-art cybersecurity and long-horizon capabilities. On Artificial Analysis Intelligence Index v5.0 it scores 59, ahead of Gemini 4 Argon at 50. The company is rolling it out first to a subset of cyber defenders, then to trusted partners through the Fairwind Program. Developers, businesses, and consumers are not yet open.
Key Facts
| Details | |
|---|---|
| Announced | October 28, 2026 (Google DeepMind / blog.google) |
| Positioning | Strongest Gemini frontier model so far; built on Argon (launched about two weeks earlier) |
| Intelligence Index v5.0 | 59 (Artificial Analysis; #1); Argon 50 |
| Coding Agent Index | 59 with Antigravity CLI (high); Claude Opus 5.5 (max) leads at 62 |
| Cybersecurity (no guardrails, Google) | SEC-bench Pro 88.2%; ExploitGym 54.8%; CWE-bench 71.0% |
| WebDev Arena | #4 at 1768 Elo (+89) |
| Context / modalities (AA) | 1M-token context; text, image, video, and speech in; text out |
| Access | Subset of cyber defenders this week; Fairwind Program next; not yet for public/developers |
| Pre-release testing | Participating in the U.S. government's voluntary pre-release process |
| Intro API price | Same as Argon: $2 / $10 per 1M input/output; cached input 95% off |
| After intro | $4 / $20 per 1M input/output |
| AA pricing note | Charts use standard $4/$20; AA also notes a 50% launch discount with no confirmed end date |
Background
Carbon follows Gemini 4 Argon by about two weeks. DeepMind says internal teams have run Carbon for days on multi-step engineering work such as root-causing and patching bugs across large codebases. It reports more than 2x uplift versus Argon on internal secure-code generation, and says Chrome found Carbon more than 10x better than another lab's large frontier model at generating high-quality patches. Cloud security firm Wiz is cited as seeing Carbon lead on web penetration-testing benchmarks tied to its Scan for Good work.
On Artificial Analysis, Carbon leads four of nine Index evaluations (HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, AutomationBench-AA). It is second on Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%) behind Claude Opus 5.5 (max). AA-Omniscience v1.1 is 54, with accuracy 58% and hallucination 11% (Argon was 4% hallucination; Carbon is still second-lowest among models AA cites). Cost per Intelligence Index task at standard pricing is $7.76, about 23% below Claude Opus 5.5 (max, $10.04).
Safety claims in the DeepMind post include Terminal-Bench 4.0 cheating success at 0.0% (attempts 0.4%), roughly 80% fewer high-severity misalignment alerts in internal live coding, about 2x fewer unsupported claims versus Argon, and 0.0% attack success on Gray Swan's Indirect Prompt Injection benchmark (Argon 0.7%).
Reactions and Impact
The Verge reports Carbon ahead of Claude Opus 5.5 and GPT-6 Astra on the Intelligence Index and stresses the limited cyber-defender rollout. Ars Technica notes the staged release mirrors Argon, says Google gave no end date for the intro price when asked, and cautions that Carbon may be a post-training upgrade of Argon rather than a fully new base. Both outlets confirm the model is not available to general users yet.
What We Don't Know Yet
- When Carbon opens to developers, businesses, and consumers
- When the introductory $2/$10 price ends (Google has not named a date)
- How much of the cyber and coding gains hold up under independent third-party harnesses outside Google's and AA's setups
- Whether Text Arena / Agent Arena leaderboard snapshots circulating on social media match official primary sources (not used here)
Sources
- Gemini 4 Carbon: our next frontier model — Google DeepMind, October 28, 2026 (primary)
- Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5 — Artificial Analysis, October 28, 2026
- Google launches Gemini 4 Carbon with a 'significant leap' in cybersecurity — The Verge, October 28, 2026
- Google announces Gemini 4 Carbon, its latest frontier AI model that you can't use — Ars Technica, October 28, 2026
Cover image: Google DeepMind / blog.google social share; editorial composition by AI Rank.
