Claude Fable 5.2
Anthropic·
Overview
Successor to Claude Fable 5.1 in the Mythos-class tier, released alongside Claude Mythos 5.2 (the same underlying model with safeguards lifted, tracked separately). Price cut to $8/M input and $40/M output (from $10/$50), output up to 30% faster than Fable 5.1, and an optional fast mode (up to 2.5x token speed) at $16/$80. 1M-token context, 128K max output. Cyber and biology safeguards stay in place; Anthropic reports about 55% fewer unnecessary blocks than Fable 5.1 in early testing. Anthropic positions Opus 5.5 as the daily driver and Fable 5.2 for the most ambitious open-ended work, including as advisor or orchestrator for smaller Claude models. The launch table leads with FrontierCode 2.0 (49.1 at xhigh effort) and reports no GPQA, MMLU-Pro or SWE-bench figure on the announcement page.
Capabilities and innovations
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| Humanity’s Last Exam (no tools) | 74.3% | third party HLE 74.3% no-tools, measured by Artificial Analysis at max effort with default fallback (68.9 xhigh, 66.2 high), rank 1 (accessed 2026-11-09). Same source as the Fable 5.1 (59.1) and Opus 5.5 (61.4) entries. Anthropic reports 74.2 with tools in the launch table and no no-tools figure on the announcement page; the near-match is a coincidence of two different setups, not a confirmation. |
| Terminal-Bench 4.0 | 65.15% | independent Terminal-Bench 4.0 65.15% (±2.31), measured by Vals AI (model page settings: high effort, mini-swe-agent harness), rank 1 of 63, tied with Claude Opus 5.5 (accessed 2026-11-09). Counts tasks completed via Anthropic's server-side safeguard fallback as successes: 17 of 198 tasks (8.59%) fell back. Same source and convention as the Opus 5.5 entry. Not yet on the official tbench.ai 4.0 board. Kingy.ai reports a vendor figure of 67.6; not verified against the system card. 4.0 is its own generation and not comparable with the 2.x or 3.0 fields. |
| ARC-AGI-1 | 98.5% | independent ARC-AGI-1 98.5% semi-private, verified by ARC Prize at XHigh reasoning, best of 10 harness configurations (accessed 2026-11-09). |
| ARC-AGI-2 | 93.8% | independent ARC-AGI-2 93.8% semi-private, verified by ARC Prize at XHigh reasoning (accessed 2026-11-09). ARC-AGI-3: 81.2% Standard harness (XHigh), 97.7% Provider Adapter harness (High); not tracked in the schema. |
- GPQA Diamond: independent evaluation pending — Not in the launch table; Artificial Analysis lists GPQA as null for this model and Vals AI has archived GPQA as saturated (checked 2026-11-09).
- SWE-bench Verified: not reported by the vendor
- SWE-bench Pro: independent evaluation pending — Not on the announcement page. BenchLM lists 93 and attributes it to the system card, but the system card PDF (assets.claude.com) was unreachable from the intake sandbox, so the figure is unverified and not entered. Check the system card directly.
- MMLU: not reported by the vendor
- MMLU-Pro: independent evaluation pending — Not in the launch table; no Vals AI run on the model page yet (checked 2026-11-09).
- Terminal-Bench 2.x: independent evaluation pending — No Terminal-Bench 2.x run on Vals AI or Artificial Analysis; Vals now runs 4.0, recorded in terminal_bench_4. A 4.0 value must not be entered in the 2.x field.
- Terminal-Bench 3.0: not reported by the vendor — 4.0 is the current Terminal-Bench release; no 3.0 run of this model.
- DeepSWE v1.1: independent evaluation pending — Not on the announcement page; system card unreachable from the intake sandbox (Opus 5.5 carried a vendor v1.1 figure in its system card, so check there).
- MMMU-Pro: independent evaluation pending — Artificial Analysis has no MMMU-Pro value for this model yet (checked 2026-11-09).
API pricing
$8 per million input tokens, $40 per million output tokens (USD, cheapest OpenRouter endpoint, checked 2026-11-09)
Reliability
- omniscience_index
- 66.8 ()independent
Links
More from Anthropic
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.