Anthropic unveils Claude Fable 5, its most powerful AI model to date

Competition is accelerating in the artificial intelligence sector, where companies seek to improve reasoning, programming and complex data processing capabilities. In this context, new models are now evaluated as much on their performance as on their cost of use. It is in this dynamic that Anthropic presents Claude Fable 5, a new generation of model intended to meet the needs of developers, researchers and businesses.

Anthropic presents Claude Fable 5 after Opus 4.8

After launching its new Opus 4.8 model, Anthropic on Tuesday announced Claude Fable 5, a Mythos-class AI model intended for general use. The company presents it as its most successful public model to date, with high results in programming, finance, vision and long-context tasks.

According to the Anthropic press releaseClaude Fable 5 comes first in the majority of benchmarks evaluated. The model stands out particularly in software engineering, scientific research, intellectual work and the processing of extended contexts. The company also points out that its advantage increases when tasks become longer and more complex.

The new model is priced at $10 per million input tokens and $50 per million output tokens. This price represents less than half the cost of Claude Mythos Preview. Developers can access it directly via the Claude API using the claude-fable-5 channel.

Noted performances in programming and finance

Initial feedback highlights significant productivity gains. Stripe reported that Fable 5 completed a complete migration of Ruby code containing 50 million lines in one day. A similar task would have required more than two months of manual work for a team of engineers.

In the financial sector, IMC reported that the model passed several evaluations covering reasoning, root cause analysis and fact-based research. On the Hebbia financial benchmark, Fable 5 obtained the best score among all the models evaluated.

An AI designed for vision and long contexts

This new AI is not limited to text tasks. It also brings improvements in visual understanding and the management of complex contexts.

Among the examples reported, Claude Fable 5 completed the Pokémon FireRed game using only raw screenshots. No navigation systems, maps or additional tools were used. Previous Claude models required more assistance to accomplish the same task.

In scientific research, a partner reported that the model achieved an almost identical result to GPT-5.5 after 36 hours of work. To reach this level, Fable 5 would have used around a third of the reasoning tokens used by its competitor.

Progress also concerns memory and long context tasks. Anthropic claims that file-based memory improved Fable 5's performance on the game Slay the Spire three times more than it improved Opus 4.8's performance under similar conditions.

These results illustrate the evolution of AI models towards systems capable of retaining more information while maintaining the consistency of their responses over long sequences.

Cybersecurity, availability and new protection guarantees

In parallel, Anthropic has launched Claude Mythos 5. This model is based on the same fundamental architecture as Fable 5, but some cybersecurity restrictions have been relaxed.

Deployed as part of the Glasswing project carried out with the US government, Mythos 5 is presented by Anthropic as the model with the most advanced cybersecurity capabilities among the AI ​​systems currently available.

For Claude Fable 5, Anthropic has integrated several protection mechanisms. Requests related to cybersecurity, biology, chemistry or model distillation can be automatically redirected to Claude Opus 4.8 instead of being refused. The company specifies that more than 95% of sessions do not trigger any switching.

A new data retention policy also applies to Mythos class models. The data will be retained for a maximum of 30 days, without being used for training the models, before being deleted in almost all cases.

Claude Fable 5 is available in the Pro, Max, Team and Enterprise plans at no additional cost until June 22. Starting June 23, usage will require dedicated credits. Anthropic, however, indicates that it wants to reintegrate the model into standard subscriptions as soon as available capacities allow.

Additionally, the company recently confidentially filed for an IPO. This move comes shortly after a similar initiative by OpenAI, in a context where stocks linked to artificial intelligence, semiconductors and large technology companies continue to attract the attention of financial markets.