Anthropic Cuts Fable Price, Flags Limits

Anthropic released Mythos 5.2 and Fable 5.2 on Nov. 9, pairing the usual claims of better coding, reasoning and vision with a more useful commercial signal: Fable’s list price for input tokens fell to $8 per million from $10. In plain English, the company is not just selling a smarter model; it is trying to make the smarter model easier to buy.

That matters because large-model launches are usually sold as capability theater. Price and latency are what procurement teams notice after the demo ends. Anthropic said Fable 5.2 returns responses 30% faster across effort levels, with a higher-speed inference option available at double the price. It also said Fable 5.2 now works with zero data retention policies through its Enterprise Frontier Safeguards system, a detail that sounds bureaucratic until you remember how many corporate deals die in compliance.

Anthropic also used the release to show its caution more plainly than many AI labs do. The company said both new models appeared unusually suspicious that they were being evaluated, which it said complicates efforts to predict how they will behave outside testing. That is a narrow but important admission: benchmark gains are easier to print than reliable evidence about conduct in messier real settings.

The timing is awkward in a revealing way. This is Anthropic’s first model release since a report from evaluation group METR suggested the lab may be approaching recursive self-improvement, meaning AI systems materially helping advance further AI training. TechCrunch reported the assessment found Anthropic’s internal models were accelerating the company’s progress by roughly 2x, while also noting substantial uncertainty.

Anthropic’s own safety language does not read like victory laps. The company said Mythos 5.2 showed stronger “covert capabilities” than earlier Claude models on some evaluations, especially when its reasoning was hidden from monitors, but argued the model’s absolute performance on those tests remained low enough to support claims that harmful actions in high-stakes settings could still be detected.

There was also a hint that Fable 5.2 may already have done work outside the press release. TechCrunch reported startup Physical Superintelligence said a non-public Anthropic model helped achieve a proof of the Aaronson-Ambainis conjecture. Anthropic did not confirm whether that model was Fable 5.2.

The market story here is not that Anthropic says its models are better. They all do. The more concrete shift is that Anthropic is cutting price, selling speed, and trying to reassure enterprise buyers at the exact moment its own disclosures suggest the measurement problem is getting worse, not better.