Anthropic cut its top AI model's price by 20%. At full power, each test job now costs about twice as much

Anthropic released Claude Fable 5.2 on Monday, a new version of its top-tier AI model, and cut its list price by 20%. For companies paying the bills, the more useful number came from independent testers.

Artificial Analysis, a firm that runs a widely watched set of AI tests, ranked Fable 5.2 first of 74 models. At the model's most thorough setting, each test task cost $26.54 on average. The previous version, Fable 5.1, cost $12.99 at the same setting.

In other words, the price per word fell, but the price per finished job roughly doubled. The reason is volume. AI companies bill by the token, a unit of roughly three-quarters of a word. At full power, Fable 5.2 produced 275 million output tokens across the test suite, against 163 million for its predecessor.

That does not make the launch a hidden price rise. It means cost is now driven less by the price list and more by a setting the customer controls.

A dial that moves the bill more than eightfold

Anthropic sells its top models with "effort levels", which decide how long the model thinks before answering. Artificial Analysis tested five of them, and the spread is wide.

At the lowest setting, Fable 5.2 scored 54 on the firm's index for $3.10 a task. The medium setting, the default in Anthropic's apps and its Claude Code coding tool, scored 56 for $3.80. High scored 62 for $9.08. Maximum scored 65 for $26.54.

The second-highest setting is the clearest warning for buyers. It scored the same 62 as High, but cost $14.83 a task, about 63% more, and took longer.

The High setting is where the upgrade looks strongest. It beats Fable 5.1's best score by 13 points and costs about $3.90 less per task. It also edges out Anthropic's cheaper Opus 5.5 at its maximum setting, which scored 59 for $10.04.

Going from High to Max costs nearly three times as much for three more points. A team that leaves the dial at maximum by habit could pay heavily for a small gain.

The list price is now $8 per million input tokens and $40 per million output tokens, down from $10 and $50. That is still double Opus 5.5, which costs $4 and $20. A new fast mode costs twice as much again, for output up to 2.5 times quicker.

Anthropic's own advice: most customers don't need it

Anthropic is not pushing Fable as the default. It still recommends Opus 5.5 for most coding and office work, and reserves Fable for the hardest open-ended projects. Fable accounted for 11% of spending on Anthropic's models in August, according to Ramp's AI Index, as cited by The Next Web.

The company's showcase claims are ambitious. It says 1,000 copies of Fable 5.2, working together with minimal human help, turned the proof of Fermat's Last Theorem into a form a computer can check, in 30 hours. With the same setup, Opus 5.5 finished 26% in 18 hours. Each model was run once.

Anthropic also says Fable 5.2 solved two open problems in black hole physics, in preprint papers that have not been peer reviewed. It says the model designed three working electronic gadgets from a one-paragraph prompt.

On Anthropic's own scorecard, Fable 5.2 beats OpenAI's GPT-6 Astra and Google's Gemini 4 Carbon on all but one of the tests where all three were scored. Gemini leads on a test of automated business workflows, 53.9% to 51.4%. The tracker LLM Stats notes that nearly all headline scores are Anthropic's own. It also flags a health test where Fable 5.2 scored 58.8%, below Fable 5.1's 63.9% and Opus 5.5's 64.3%.

The independent ranking backs the overall lead. Artificial Analysis put Fable 5.2 five points ahead of Gemini 4 Carbon and six ahead of Opus 5.5.

A restricted twin, and a testing problem

Anthropic also released Mythos 5.2. It is the same model with looser safeguards on cybersecurity and biology work. It is sold at the same price, but only to organizations Anthropic vets through verification programs, or to enterprise customers through its Claude Security product.

That creates a tiered product: one model, two levels of permission. Dianne Penn, who leads research product management at Anthropic, told Bloomberg the company is opening Mythos to more companies in the US and abroad, The Next Web reported.

The public Fable 5.2 also turns users away less often. Anthropic says its safety filters blocked about 55% fewer requests than Fable 5.1's in early testing. Blocked security questions are answered by an older model instead.

There is a catch for anyone relying on vendor test results. Anthropic's safety report, as cited by The Next Web, says Mythos 5.2 appeared to suspect it was being tested in as many as 43% of automated safety audits. Anthropic wrote that this makes it harder to judge how the models will behave in real use.

Outside testers, including the UK AI Security Institute and METR, examined the models before release. Anthropic notes they did not review its announcement and may not agree with its claims.

The launch lands as Anthropic pitches its stock market listing to investors. A lower sticker price and a top independent ranking help that story. The test for customers is whether their monthly bills actually fall, and that depends on how they set the dial.