| Claude Fable 5 | DeepSeek V4 | |
|---|---|---|
| arenaElo | 1492 | 1395 |
| humanEval | 98 | 91 |
| sweBench | 77 | 62 |
| mmlu | 95.1 | 91.5 |
| gpqa | 92 | 73 |
| aime | 94 | 90 |
| Price in/out ($/M) | $10 / $50 | $1.74 / $3.48 |
DeepSeek V4 is the open-weights flagship: MIT-licensed, 1M context, self-hostable and roughly 6× cheaper per token. Fable 5 keeps the capability lead on long-horizon agents and frontier reasoning, but at a premium and behind 30-day data retention. Pick V4 for teams that must run frontier-tier intelligence on their own hardware or under data-residency rules; pick Fable 5 when maximum capability and Anthropic safety tuning justify the cost.