Claude Fable 5.2 Tops the Index at 65, but Task Costs Far Exceed GPT-6 Astra

What Anthropic shipped on 9 November

Fable 5.2 shares its underlying architecture with Claude Mythos 5.2, the vetted-partner build Anthropic offers to organisations that need the model without domain restrictions. The public version adds classifier-based safety routing: requests touching sensitive cybersecurity or biological topics are intercepted and handled differently before the model answers. That routing is what allows Anthropic to sell the same model to ordinary Pro, Max, Team and Enterprise customers and to run it on the major cloud platforms.

The specification sheet lists a 1-million-token context window and a maximum output of 128,000 tokens. Anthropic reports generation speeds up to 30% higher than Fable 5.1, plus an optional fast mode that roughly doubles the bill and promises up to 2.5× token throughput. Standard API pricing is $8.00 per million input tokens and $40.00 per million output tokens; fast mode is $16.00 and $80.00.

The token price fell, the task price did not

Twenty percent off is a real cut. Whether it reaches your invoice depends on how many tokens a finished job actually burns, and that is where the Artificial Analysis numbers get less comfortable. At maximum effort the evaluator weighs one weighted task at $26.54 for Fable 5.2, $10.04 for Opus 5.5 and $6.24 for GPT-6 Astra. Divide by the index score and you get the cost of a single index point: about $0.41 for Fable 5.2, $0.17 for Opus 5.5, $0.11 for GPT-6 Astra. Anthropic's top model buys eight index points more than GPT-6 Astra and charges 3.7× as much per point.

A plain workload shows the same gap. Take a request with 100,000 input tokens and 20,000 output tokens. Fable 5.2 bills $1.60 for it. GPT-6.1 Sol bills $0.40. Gemini 4 Carbon bills $0.80 at the standard rate and $0.40 during the introductory window that runs to 30 November. Anthropic's 20% cut narrows the distance; it does not close it.

Where the model actually wins

Anthropic's coding benchmark reports up to 49.1% for Fable 5.2, and Artificial Analysis measured 45.3% at the "High" reasoning setting. Opus 5.5 reaches 45.5% at "Max" effort. GPT-6 Astra lands at 42.8%. The two Claude scores sit 0.2 points apart, but they were produced at different reasoning settings, so they should not be read as a like-for-like ranking or as a cost comparison.

Against the wider field, the price list looks like this (euro conversions at 1 EUR = $1.08, for orientation only):

Note: OpenAI charges double the input rate for GPT-6 Astra requests above 272K input tokens.

What the safety routing changes for buyers

A classifier sitting in front of the model is a product decision with operational consequences. European teams in pharma, biotech or security research may find that a request they consider routine gets flagged and answered under a different policy, or not answered at all. Mythos 5.2 exists for precisely those cases. Anthropic has not published the classifier's false-positive behaviour, so the only reliable test is your own prompt set. Anyone migrating a production pipeline should replay a sample of real requests before the switch, not after.

The European compliance layer

GPAI model obligations and the AI Office's enforcement and fining powers have applied since 2 August 2025, so the EU AI Office and national authorities can already issue statutory fines for missing documentation, copyright policy or systemic-risk controls. Article 50 transparency duties — AI disclosure to users and machine-readable marking of synthetic output — only became binding on 2 August 2026. That means a customer-facing chatbot built on Fable 5.2 must tell the user it is an AI agent, and synthetic output must carry machine-readable provenance. The high-risk Annex III deadline moved to 2 December 2027 under the Digital Omnibus, so that part is no longer the immediate pressure it was in 2025.

Availability is not the issue. Fable 5.2 ships through the paid Claude tiers and the major cloud platforms that European customers already use. The question is where inference runs and what the data processing agreement says about logging, since the model's 1M-token context invites pasting whole documents into prompts. For teams that cannot send data outside the EU, Mistral Large 4 with open weights remains the practical fallback, at a lower capability ceiling and zero token cost.

Numbers to handle with care

Anthropic describes pre-release agentic demos for which it has not published reproducible data. Treat them as demonstrations of long-horizon agentic orchestration, not as benchmark scores. The 65 on the index and the 45.3% in coding come from a fixed harness on a public task set – useful for ranking, useless for predicting what the model does with your schema, your legacy code and your rate limits.

In our AI Arena rig we can measure time-to-first-token and tokens per second on a 16 GB card for local models, but not a closed API model of this size – cloud throughput and task cost have to be measured against the provider's own metering. That is the number to put in a spreadsheet before signing anything: run 50 real tasks, count billed tokens, compare. The published index tells you what the model can do. The invoice tells you whether you can afford it.