一句话总结
This dataset compiles comparable figures for Gemini 4 Carbon (released October 28, 2026) and rivals — Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and GPT-6.1 Sol — on the Intelligence Index, terminal tasks, long context, cybersecurity and pricing; all values come from official disclosures or third-party evaluations.
Data Notes
- Index version: Artificial Analysis Intelligence Index v5.0, four domains (knowledge work, coding, science, general reasoning) at 25% each, with about 60% private data.
- Cybersecurity benchmark scores (SEC-bench Pro, ExploitGym, CWE-bench) are official values measured in an unrestricted test setting.
- Prices are API rates per million tokens; Carbon's $2/$10 launch price is a limited-time rate, with the regular price at $4/$20.
Core Data
| model | Index v5.0 | Terminal-Bench 4.0 | GraphWalks (256k-1M) | SEC-bench Pro | ExploitGym | CWE-bench | Input ($/1M) | Output ($/1M) |
|---|---|---|---|---|---|---|---|---|
| Gemini 4 Carbon | 59 | 65.2% | 96.2% | 88.2% | 54.8% | 71.0% | 2 (intro)/4 | 10 (intro)/20 |
| Claude Opus 5.5 | 58 | 59.6% | 66.8% | Not disclosed | 34.5% | 67.0% | 4 | 20 |
| GPT-6 Astra | 57 | 59.1% | 71.8% | 85.4% | 42.4% | 68.0% | 10 | 50 |
| Gemini 4 Argon | 50 | Not disclosed | Not disclosed | Not disclosed | Not disclosed | Not disclosed | 2 | 12 |
| GPT-6.1 Sol | 55 | Not disclosed | Not disclosed | Not disclosed | Not disclosed | Not disclosed | 1.5 | 8 |
Field Notes
- Intelligence Index v5.0: overall capability score out of 100; 59 is the global high as of October 28, 2026.
- Terminal-Bench 4.0: success rate on terminal tasks such as operating a computer.
- GraphWalks: information-tracking accuracy under long context (256k to 1M tokens).
- SEC-bench Pro / ExploitGym / CWE-bench: vulnerability discovery / exploitation / patching benchmarks.
- Input/Output price: API rate per million tokens; Carbon's $2/$10 intro rate applies at launch, then reverts to $4/$20.