← All Comparisons
Claude Fable 5.1 vs GPT-6 Astra
| GPT-6 Astra | Claude Fable 5.1 |
|---|
| Pricing | ||
|---|---|---|
| Input / output per 1M | $10 / $50 | $10 / $50 |
| Cached input per 1M | about $1 | $0.25 |
| Cache writes per 1M | about $12.50 | $12.50 / $20 |
| vs predecessor | 2.5x over GPT-5.6 Sol | Same as Fable 5 |
| Coding | ||
| DeepSWE v1.1 | 74.1% | 67.4% |
| Terminal-Bench 4.0 | 57.9% | 55.8% |
| CursorBench 3.2.0 | Not reported | 73.4% |
| AA Coding Agent Index | 67 | 70 |
| Math and reasoning | ||
| FrontierMath Tier 4 | 97.6% | 87.8% |
| ARC-AGI-3 | 99.9% | Not reported |
| Knowledge | ||
| GPQA Diamond | 96.0% | 93.7% |
| HLE, no tools | Not reported | 60.9% |
| HLE, with tools | Not reported | 65.0% |
| Agentic work | ||
| OSWorld | 72.6%* | 77.9% / 41.7%* |
| AutomationBench | Not reported | 31.4% |
| Context and controls | ||
| Context window | about 1.05M tokens | 1M tokens |
| Max output | 128K tokens | 128K tokens |
| Knowledge cutoff | April 2026 | June 2026 |
| Effort dial | Low - Max | Low - High |
| Safety and access | ||
| Restricted program | Daybreak | Trusted access (Mythos 5.1) |
| Public-model limit | No exploit PoCs on standard API | Cyber tasks rerouted from Fable 5.1 |
Scores are vendor-reported launch-table figures from September 2026 unless marked independent (AA). OSWorld rows compare different task setups. Some OpenAI-chart Claude rows use reduced-safeguard Mythos rather than shipped Fable 5.1. Prices are list rates - task cost depends on effort level, caching, and harness.