Claude Fable
by Anthropic · Model · Language & code models
strong
Top-ranked on most independent leaderboards, but expensive and beaten on some tasks.
#1 agents#1 coding#1 documents#1 image understanding#1 knowledge work#1 robot control#1 web development
Scores 54% on real vs 9% on shuffled protein sequences, a +45.1 point gap, the largest of any LLM tested
AccessAPIPrice$8 in · $40 out per M tokensContext1,000K tokens
Why we say so
Arena puts Fable 5.2 first on agents and web development, Artificial Analysis ranks it first on its Intelligence Index (65 versus Astra's 62), and Vals has it just ahead of Gemini Neon. Xinming Tu found its protein-sequence knowledge far ahead of other LLMs, and Simon Willison built a plugin almost in one shot. Against that, Mercor has Astra well ahead on coding, AI Search saw it fail an image test that Gemini Carbon and DeepSeek Flash passed, and HN commenters call it too pricey for routine coding.
Versions
| Version | Released | Rank |
|---|---|---|
| Claude Fable 5.2LATEST strong Current leader on agents, web development and overall index rankings | Nov 9 | #1 agents#1 coding#1 knowledge work#1 robot control#1 web development#2 3d generation |
| Claude Fable 5.1 | Sep 1 | #1 documents |
| Claude Fable 5 unproven Anthropic's filter blocked 44% of protein questions, so it was excluded from testing | #1 image understanding |
Evidence
- Nov 19Anthropic document says Fable 5.2 and Mythos 5.2 were capability-capped in training for AI R&D domains like kernel optimization; commenters call it anti-competitive gating
- Nov 18Beats every open protein language model tested on next-residue prediction (ProGen3-3B +32.5, ProtGPT3-10B +27.7)
- Nov 18GPT, Gemini and open-weight LLMs score near zero gap; on 56 post-cutoff proteins Fable 5.2 got +22.1, GPT-6.2 Sol -0.4
- Nov 15#1 on Arena webdev (1824)
- Nov 15
Show all 40
- Nov 14
- Nov 14
- Nov 12
- Nov 11On GDPx materials science costs 89% more than Astra with 4.3x output tokens for about the same score; 7.5x Carbon's cost
- Nov 11One lab's 100 unsaturated environments found Astra and GPT-6.1 Sol produced better solutions; Fable paused to ask a nonexistent human
- Nov 11#1 on Arena webdev (1838)
- Nov 11Ranks number one on the Artificial Analysis Intelligence Index, a few points above Gemini 4 Carbon
- Nov 11Ranks number one on Vals Index for knowledge work, well above the best Gemini models and GPT-6 Astra
- Nov 11Ranks first on FrontierSWE, but the gap to GPT and Gemini Carbon is within error bars
- Nov 11Built a detailed third-person Unreal Engine ninja game with a critic-agent loop, taking about seven hours and two five-hour limits; best of frontier models
- Nov 11Wrote a from-scratch bullet-through-water-balloon physics simulation in 35 minutes with adjustable settings; better than Opus 5.5 or GPT-6 Astra
- Nov 11Failed Frogbench: could not find the hidden frog even at extra effort; DeepSeek Flash is the only model that finds it
- Nov 11Identified 2 of 6 brain tumor types correctly, versus 1 of 6 for GPT-6 Astra and Kimi; it answered despite earlier Fable refusing medical questions
- Nov 11Medical deep-research answer on leukemia was too concise and left out detail; reviewer prefers GPT, DeepSeek and Kimi
- Nov 9#1 on Arena document (1515)
- Nov 9#1 on Arena vision (1326)
- Oct 19#1 on Arena vision (1325)
- Sep 27Previous best Claude; more expensive and restrictive than Opus 5.5, and ranked just above it on LiveBench
- Sep 13Work used an unnamed internal research model, described as roughly comparable to Claude Fable 5.1, with dozens of agents in a Claude Code-based harness
- Sep 13#2 on Arena document (1513)
- Sep 6Anthropic's best model, pitched for agentic scientific research, coding and knowledge work; beats Fable 5 and GPT 5.6 Sol in its table
- Sep 6
- Sep 3Anthropic chart: Fable 5.1 beats Fable 5 at every price point on terminal science tasks, clearing 50% versus under 30%
- Sep 3Autonomously researched earnings and produced a motion-graphics financial video; layout looked a bit better than GLM 5.3 and GPT output
- Sep 3Directed a matcha commercial via Higgsfield CLI, six clips; video good but audio overlapped. GPT Sol result shown for comparison
- Sep 3Costs $3.69 per task vs $3.14 for Fable 5, and 4-8x more than GPT, Qwen, Kimi and GLM; evaluating it cost $8,500
- Sep 1Failed a multilingual pick-up-line test: translations were clean and in native scripts but culturally thin, mostly the same sentence translated
- Aug 28Scores 21.4% on Terminal-Bench-Science in Claude Code, below Opus 5 (30.0%) and GPT Sol (22.4%)
- Aug 22Author thinks Fable could finish the watch UI pixel-perfect unsupervised if given a simulator screenshot feedback loop
- Aug 22Fable 5 (Claude Code, high effort) topped the speedrun: record 2,726, 81.7% of the gap to human record closed, 8.7 agent-days
- Aug 20One commenter finds Fable 5 less verbose than Opus 5, or better at following style guidelines
- Jul 27Refused the tumor-identification and atherosclerosis research prompts that Opus 5 answered
- Jul 17Closed weights, no parameter count published, 1M context, 128K max output, thinking always on; $10 in / $50 out per million tokens
- Jul 17On the same prompt, Fable 5 produced the most polished UI with 8/8 green checks and responsive tabs; judged second of three
- Jun 10Took #1 on Arena vision
Cost to run
Alternatives
Made for Claude Fable