Claude Fable 5.2 vs GPT-6 Astra: OpenAI Should Be Really Worried
Anthropic once again leads the chart with its latest flagship models. The company has officially announced Fable 5.2 and Mythos 5.2, a significant jump over their predecessors. OpenAI is also set to announce its latest models very soon, but here is how the latest Fable 5.2 compares with OpenAI's flagship GPT-6 Astra model.
Benchmarks
| Benchmark | Claude Fable 5.2 | GPT-6 Astra | Difference |
|---|---|---|---|
| FrontierCode 2.0 | 49.1%Xhigh effort | 42.8% | Fable +6.3 points |
| FrontierSWE v2 | 69.9% | 65.5% | Fable +4.4 points |
| AutomationBench | 51.4% | 41.4% | Fable +10.0 points |
| Terminal-Bench-Science 0.1 | 74.6% | 68.1% | Fable +6.5 points |
| OSWorld 2.1 | 89.3% | 79.5% | Fable +9.8 points |
| RoboDojo Sim | 64.9% | 29.8% | Fable +35.1 points |
| AA-Briefcase v1.1 rubric | 66.1% | 52.0% | Fable +14.1 points |
| AA-Briefcase v1.1 Elo | 1785 Elo | 1569 Elo | Fable +216 Elo |
| Remote Labor Index | 41.3%High effort | 20.8% | Fable +20.5 points |
| Humanity's Last Exam | 74.2% | 57.2% | Fable +17.0 points |
| Chartography | 73.4% | 71.0% | Fable +2.4 points |
Scores from Anthropic, at each model's highest available effort unless a cell says otherwise. Fable 5.2 leads on all 11.
Artificial Analysis benchmarks
| Benchmark | Fable 5.2, Max effort | GPT-6 Astra, Max effort | Difference |
|---|---|---|---|
| Intelligence Index v5.0.1 | 65 | 57 | Fable +8 |
| AA-Briefcase v1.1 | 1785 Elo | 1569 Elo | Fable +216 Elo |
| AutomationBench-AA | 76.4% | 68.5% | Fable +7.9 points |
| Terminal-Bench 4.0 | 59.6% | 59.1% | Fable +0.5 points |
| FrontierCode 2.0 | 48.8% | 42.4% | Fable +6.4 points |
| AA-Omniscience v1.1 | 67 | 42 | Fable +25 |
| GDP.pdf | 24.2% | 31.0% | Astra +6.8 points |
| HLE-Diamond | 80.2% | 63.8% | Fable +16.4 points |
| CritPt v2 | 76.5% | 70.5% | Fable +6.0 points |
| Terminal-Bench-Science 0.1 | 73.3% | 63.3% | Fable +10.0 points |
Fable 5.2 wins 9 of these 10. GPT-6 Astra wins GDP.pdf, and Terminal-Bench 4.0 is almost a tie.
Cost and speed
| Metric | Fable 5.2, Max effort | GPT-6 Astra, Max effort | Difference |
|---|---|---|---|
| Cost per Intelligence Index task | $26.54 | $6.24 | Astra about 76% cheaper |
| Cost to run the full Index | $22,191 | $5,661 | Astra about 74% cheaper |
| Output tokens per task | 274K | 52K | Astra about 81% fewer |
| Reasoning tokens per task | 193K | 35.8K | Astra about 81% fewer |
| Output tokens for the full Index | 275M | 54.3M | Astra about 80% fewer |
| Output speed | 103 tokens/s | 38 tokens/s | Fable about 2.7x faster |
| Time to first token | 231.55s | 232.57s | Nearly tied |
| Average time per task | 44.1 minutes | 22.6 minutes | Astra about 49% faster |
Fable 5.2 writes much faster, but at Max effort it uses about five times as many tokens per task as GPT-6 Astra, so each task takes longer and costs about four times as much.
Pricing and specs
| Spec | Claude Fable 5.2 | GPT-6 Astra | Difference |
|---|---|---|---|
| Input | $8 | $10 | Fable 20% cheaper |
| Output | $40 | $50 | Fable 20% cheaper |
| Cached input | $0.20 | $1 | Fable 80% cheaper |
| Cache writes | $105 minutes. $16 for 1 hour | $12.50 | Fable 20% cheaper |
| Long prompts | Same priceAcross the full 1M context | 2x input, 1.5x outputPrompts over 272K tokens | Fable has no surcharge |
| Batch API | $4 and $20 | $5 and $25 | Fable 20% cheaper |
| Fast mode | 2x priceUp to 2.5x faster | 2x price | Both offer one |
| Context window | 1M tokens | 1.05M tokens | Astra slightly larger |
| Max output | 128K tokens | 128K tokens | Same |
| Reasoning effort | Low to MaxThinking always on, default High | Low to Max | Same five levels |
| Input and output | Text and images in, text out | Text and images in, text out | Same |
| Knowledge cutoff | June 2026 | April 30, 2026 | Fable newer |
| Released | November 9, 2026 | September 3, 2026 | Fable 67 days newer |
| API model ID | claude-fable-5-2 | gpt-6-astra | Different |