Gemini 4 Carbon Takes #1 on AA Intelligence Index v5.0, Leads 11 of 15 Benchmarks
Google DeepMind announced Gemini 4 Carbon on October 28, launching it at the top of the Artificial Analysis Intelligence Index v5.0 with a score of 59 — one point ahead of Claude Opus 5.5 (58), two ahead of GPT-6 Astra (57), and nine ahead of its own predecessor Gemini 4 Argon (50). The model is not yet publicly available; Google is rolling it out initially to a subset of cyber defenders through the Fairwind Program while completing pre-release testing with the U.S. government’s voluntary safety process.
Benchmark Summary
| Benchmark | Carbon | GPT-6 Astra | Claude Opus 5.5 | Gemini 4 Argon |
|---|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 57 | 58 | 50 |
| Vals Index v2.2 | 70.4% | 65.0% | 70.2% | 69.2% |
| GraphWalks (256K–1M BFS F1) | 96.2% | 71.8% | 66.8% | 84.2% |
| AutomationBench | 53.9% | 41.4% | 42.5% | 51.3% |
| Agents’ Last Exam v1.1 | 77.4% | 73.1% | 71.6% | 69.0% |
| Terminal-Bench 4.0 (AA) | 65.2% | 59.1% | 59.6% | 57.1% |
| FrontierSWE 2 | 61.2% | 65.5% | 62.3% | 55.0% |
| OSWorld-2.1 offline | 81.5% | 80.9% | 87.2% | 74.7% |
| WebDev Arena ELO | 1768 | — | — | — |
Carbon leads 11 of the 15 benchmarks Google disclosed. Its two notable losses: FrontierSWE 2 (where GPT-6 Astra posts 65.5% against Carbon’s 61.2%) and OSWorld-2.1 offline (where Claude Opus 5.5 scores 87.2% against Carbon’s 81.5%).
Long Context is the Widest Gap
The most decisive margin is GraphWalks, a long-context retrieval task that runs inputs from 256K to 1M tokens. Carbon scores 96.2% — a 12-point gain over its own predecessor Argon and a 24-point gap over GPT-6 Astra. Google framed this as a “significant improvement in long-horizon tasks,” and the numbers bear that out. At 1M tokens, Carbon is operating in a tier that neither Astra nor Opus 5.5 currently matches.
Cybersecurity Is the Launch Rationale
Google described Carbon as built for “cybersecurity and long horizon capabilities” and the benchmark profile makes that claim concrete. On three external cybersecurity evaluations run without guardrails:
- SEC-bench Pro (finding real vulnerabilities): 88.2%, new state-of-the-art
- ExploitGym (validating exploitability): 54.8%, new state-of-the-art
- CWE-bench (patching root causes): 71.0%, above Opus 5.5’s 67.0%
Neither Opus 5.5 nor Fable 5.1 has a published SEC-bench Pro score. Google also cited internal results: more than 2x uplift in generating secure code over Argon, and Google’s Chrome team reporting Carbon performs more than 10x better than another lab’s frontier model on patch quality.
The restricted launch follows from this framing. Access starts with cyber defenders this week, expands through Fairwind Program partners, then eventually reaches developers and consumers.
Cost Position
Carbon launches at $2 per million input tokens and $10 per million output tokens — matching Argon’s introductory price. That intro period runs through November 30, 2026, after which pricing moves to $4/$10 standard rates. Cached input tokens are priced at 95% off.
On a cost-per-task basis, Artificial Analysis reports $7.76 for Carbon versus $10.04 for Claude Opus 5.5 — a 23% reduction for comparable or slightly higher overall intelligence. That efficiency argument carries more weight than the raw benchmark margin: Carbon’s AA Index lead over Opus 5.5 is one point; the cost gap is nearly $2.30 per task.
Where Carbon Trails
FrontierSWE 2 is the benchmark Carbon loses. GPT-6 Astra scores 65.5%, Carbon 61.2%, Opus 5.5 62.3%. For teams building strictly on agentic coding scaffolds, Astra retains an edge on this harness.
OSWorld-2.1 offline is the other gap: Opus 5.5 at 87.2% sits 5.7 points clear of Carbon. Anthropic’s model remains the strongest for desktop GUI automation tasks.
Where It Ranks on WebDev Arena
Carbon landed at 1768 ELO on the WebDev Arena leaderboard — #4 overall, a jump of 89 points from Argon. Google credited improvements in output quality and artifact generation for the gain. The number situates Carbon in the top tier without displacing whatever currently holds the top three spots on that board.
What’s Actually Available
As of October 28, Google has not published a public API model identifier for Carbon. Developers can check the Gemini API release notes for updates; Argon (gemini-4-argon) remains the available Gemini 4 entry point. Carbon’s general availability timeline is described as “as soon as possible.”