Claude Fable 5
Last updated: June 10, 2026
Specs
| Vendor | Anthropic |
|---|
| Released | June 9, 2026 |
|---|
| Context window | 1M tokens |
|---|
| Input | $10 / M tokens |
|---|
| Output | $50 / M tokens |
|---|
Benchmarks
| mmlu | 95.1 |
|---|
| gpqa | 92 |
|---|
| humanEval | 98 |
|---|
| arenaElo | 1492 |
|---|
| sweBench | 77 |
|---|
| sweBenchPro | 72 |
|---|
| liveCodeBench | 84 |
|---|
| aiderPolyglot | 86 |
|---|
| terminalBench | 52 |
|---|
| mmluPro | 89 |
|---|
| hle | 30 |
|---|
| aime | 94 |
|---|
| math500 | 98.5 |
|---|
| mmmu | 83 |
|---|
Strengths
- Anthropic's most capable widely released model: state-of-the-art long-horizon autonomous agentic work and frontier reasoning
- 1M-token context window (the default) with 128K max output and always-on extended thinking
- Generally available on the Claude API at $10 / $50 per M tokens: no Project Glasswing barrier
Weaknesses
- New tokenizer counts ~30% more tokens than the Opus tier for the same content: re-baseline cost and context budgets
- Premium pricing above the Opus tier: overkill for high-volume, latency-sensitive traffic
- Requires 30-day data retention (not available under zero-data-retention); safety classifiers may refuse some bio/cyber-adjacent requests
Best for
- Long-horizon autonomous coding agents and overnight refactor runs
- The hardest reasoning, mathematics and scientific research problems
- End-to-end enterprise deliverables: financial analysis, spreadsheets, slides and documents
Compare with other AI models