10 November 2026 · Frontier and open models · 7 min read

Claude Fable 5.2 Review: Top Scores, High Price, Slow

Claude Fable 5.2 leads two independent rankings, but the lead costs real money and real time. Here is what the evidence shows, where the model falls short, and who should pay for it.

Key takeaways

Anthropic released Claude Fable 5.2 and a restricted sibling, Mythos 5.2, on 9 November 2026; Fable is priced at $8 per million input tokens and $40 per million output tokens [1][5].
Artificial Analysis (AA) puts Fable 5.2 first on its Intelligence Index at 65, five points ahead of Google's Gemini 4 Carbon [3][5].
Vals AI also ranks it first, at 71.31% on the Vals Index, but only about half a point ahead of the next model [4].
The cost is high: AA measured $26.54 per task and 46.4 minutes per task, against a median of 12.6 minutes for comparable models [3].
China's open-weights Kimi K3.1 Pro scores 46 on the same AA index at $3.68 per task [7].

What changed

Fable 5.2 and Mythos 5.2 share one underlying model. Fable carries safeguards that block high-risk cybersecurity and biology work and is open to all users; Mythos has some of those limits removed and goes only to organisations accepted into Anthropic's verification programmes, according to The Decoder [5]. Anthropic says the new models generate output up to 30% faster than Fable 5.1 and Mythos 5.1 [1]; The Decoder notes the price is down from $10 and $50 per million tokens [5]. Anthropic still recommends Claude Opus 5.5 for everyday work [5].

Availability is wide on paper. Anthropic says Fable 5.2 is available on all platforms, including the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure [1]; a developer write-up adds paid Claude plans and Claude Code [11]. Fast mode, up to 2.5 times quicker at twice the price ($16 and $80 per million tokens), is limited to Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry [1][5]. We found no regional restrictions in the sources we could read; check your cloud region before planning.

What the independent tests show

AA's Intelligence Index v5.0.1 combines nine evaluations. At its highest effort setting Fable 5.2 scores 65, ahead of Gemini 4 Carbon (60), Opus 5.5 (59) and Fable 5.1 (50) [3][5]. AA reports leading scores of 80% on HLE-Diamond, 73% on Terminal-Bench-Science and 49% on FrontierCode 2.0 [5]. At a lower effort setting it still scores 62 using about 98,000 output tokens per task, against 173,000 for Opus 5.5 at maximum effort, per The Decoder's reading of AA's data [5].

Vals AI ran the model at high effort on a different benchmark mix. It reports 71.31% on the Vals Index, first of 36 models, ahead of Claude Sonnet 5.5 (70.85%) and Gemini 4 Carbon (70.45%); the margin of error on Fable's score is about ±0.96 points, so the lead is narrow [4]. It also ranks first on Vibe Code Bench (93.77%) and Finance Agent v2.1 (69.47%) [4].

Strengths, weaknesses and best fit

Strengths. Science, coding and financial analysis: Vals ranks it first on BioMysteryBench, Vibe Code Bench and Finance Agent [4], and AA reports leading scores on HLE-Diamond and Terminal-Bench-Science [5]. Vals reports a 0% refusal rate in its runs [4]. Anthropic's own table, which is vendor data run at maximum effort with safeguards on, shows FrontierCode 2.0 at 49.1% against 32.4% for Fable 5.1 and 45.5% for Opus 5.5 [1][5]. Anthropic also says about 1,000 Fable 5.2 agents formalised the proof of Fermat's Last Theorem in the Lean language in 30 hours, against 11 days for its first formalisation [1][2]. Treat these as claims, not independent findings.

Weaknesses. It is slow and costly. AA ranks it 71st of 72 on time per task, 73rd of 73 on cost and 72nd of 73 on verbosity, with 275 million output tokens across the index [3]. It is not the best at everything: on AA's Terminal-Bench 4.0 it scores 60%, behind Kimi K3.1 Pro (67%), Gemini 4 Carbon (65%) and Sonnet 5.5 (64%) [5], although Vals's own run of the same benchmark puts it joint first at 65.15% [4]. Vals ranks it 20th of 37 on its Legal Index and 27th of 84 on Tax Agent Bench, and Opus 5.5 beats it on SWE-bench Multimodal [4].

Best fit. Work where one correct answer is worth far more than the compute: research-grade analysis, difficult engineering and finance tasks. For routine coding, summarising or high-volume pipelines, Anthropic's own advice to use Opus 5.5 is sound, and cheaper models will do.

Pricing in practice

List price is lower than Fable 5.1's $10 and $50 [5], but the bill depends on effort and tokens. AA's $26.54 per task is for its heaviest setting; Vals measured $17.94 per test at high effort [3][4]. Cheaper effort settings narrow the gap with little loss in AA's data (62 instead of 65, per The Decoder [5]), so test your own workload before committing.

The global picture

China. Moonshot AI's Kimi K3.1 Pro has open weights on Hugging Face under its own licence, 2.8 trillion total and 104 billion active parameters, and a 1-million-token context window [7]. It scores 46 on the AA index, 19 points below Fable 5.2, at $2 and $10 per million tokens; AA calls it slow and expensive for its class, yet it costs about one-seventh as much per task [7]. Our earlier review covers its weights and licence [9].

South Korea. Separately, NVIDIA announced on 9 November an investment in Point2 Technology and planned investments in Motif Technologies, RLWRLD and Upstage, according to Unite.AI [8]; the planned deals are not yet closed. It is a reminder that the compute and model supply chain runs well beyond the US.

Anthropic's launch comes as the company prepares an IPO, CNBC reports [6]; our funding coverage has the valuation context [10].

Our take

Fable 5.2 is the strongest model on public evidence today, and also among the least efficient. If your task pays for accuracy, it earns its place; if not, you are buying a lead that Vals measures at under one point. Independent labs have so far confirmed the direction of the vendor's claims but not every number.

What to watch next

More independent evaluations beyond AA and Vals, including non-US labs.
Whether Google's next Gemini release closes the five-point gap on AA.
Whether Anthropic widens Mythos access.

FAQ

What is Claude Fable 5.2? Anthropic's new model for the hardest technical work, open to all users and released on 9 November 2026 [1][5].

How much does it cost? $8 per million input tokens and $40 per million output tokens [5].

Is it better than Opus 5.5? On AA and Vals, yes, by six points on AA and narrowly on Vals; Anthropic still recommends Opus 5.5 for everyday use [3][4][5].

Is there an open-weights alternative? Kimi K3.1 Pro is the strongest we found on AA, at 46 [7].

This article is analysis, not financial or investment advice.

Builder's Signals

Paper P2-MOPD: Accelerating Multi-Teacher On-Policy Distillation via Parameter Priors · China · Training and efficiency
Picked because: From StepFun and Shanghai Jiao Tong University with Adelaide University's Australian Institute for Machine Learning (v2, 5 November). It folds several expert models into one small student while training it; according to the authors, their 1.7B model reaches 94.7% average recovery against 80.9% for the baseline method, in about half the training steps.
Model NVIDIA Nemotron-3.5-Transcribe-30B-A4B · US/Canada · Speech and audio
Picked because: Open-weights offline transcription in 40 languages for recordings up to one hour, with speech translation and speaker-attributed transcripts. NVIDIA's model card says its mixture-of-experts design activates about 4.3B parameters. Licence: OpenMDW-1.1.
Repo sandbox-mcp · Global · Agents
Picked because: Gives an AI agent a fresh, disposable Windows 11 machine through Windows Sandbox to install, drive and test an app, then deletes it. Command-line tool with an optional MCP adapter; MIT licence; 405 stars as of 11 November.

Each item links to its primary page (paper, model card or repo). Results and benchmarks are the authors' own unless we say otherwise. Checked 10 November 2026. All Signals · How we pick

Sources

Introducing Claude Fable 5.2 and Mythos 5.2, Anthropic, 9 November 2026.
Formalizing Fermat's Last Theorem, Anthropic, November 2026.
Claude Fable 5.2 (Max) intelligence, performance and price analysis, Artificial Analysis, read 10 November 2026.
Claude Fable 5.2, Vals AI, 9 November 2026.
Claude Fable 5.2 and Mythos 5.2 cement Anthropic's AI lead just before its IPO, The Decoder, 9 November 2026.
Anthropic launches Claude Fable 5.2 and Mythos 5.2, CNBC, 9 November 2026.
Kimi K3.1 Pro intelligence, performance and price analysis, Artificial Analysis, read 10 November 2026.
NVIDIA backs Point2 and plans three more Korean AI start-up investments, Unite.AI, 9 November 2026.
Kimi K3.1 Pro review: open weights, price and benchmarks, AI Worldwire, 27 October 2026.
OpenAI Banks $11B at $1.25T as Anthropic IPO Talk Slips, AI Worldwire, 9 November 2026.
Claude Fable 5.2: benchmarks, price and the first independent test results, DEV Community, 9 November 2026.

Disclosure: the AI system that assisted this article, Claude, is made by Anthropic, the company whose model is reviewed here. AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. Research and drafting are assisted by AI; every article is reviewed and approved by a human editor before publication. Subscribe to the AI Worldwire newsletter. Last updated: 11 November 2026 (Builder's Signals added).

Subscribe to the AI Worldwire newsletter