Just moments ago, Anthropic announced the official release of Claude Fable 5.
Fable can be translated as "fable," derived from the Latin word fabula, meaning "story" or "legend."
According to Anthropic, Fable 5's performance surpasses all its previously publicly released models. It leads in nearly all AI performance benchmarks, demonstrating outstanding performance in software engineering, knowledge work, vision, scientific research, and many other fields.
Furthermore, Claude 5 Fable is more token-efficient than previous Claude models, capable of handling millions of tokens in long-running tasks without losing focus and improving output through self-notes. The longer and more complex the task, the more Fable 5's advantage over other models becomes apparent.
Karpathy also shared his firsthand experience, calling it a "very exciting" version.
"From a practical experience perspective, this is also a leap in progress worthy of a 'major version upgrade.' In my view, it is on the same level as the improvement brought by Claude 4.5 last November."
Members of the Claude Code team state that Fable 5 will give you a quality of experience like never before:
Many users have already reported their real-world usage experiences:
However, due to the model's high performance and concerns about security risks, Anthropic has designed security measures so that when Claude faces certain queries or believes someone is "distilling" it, Claude will respond with Claude Opus 4.8, with about 95% of conversations not triggering model downgrades.
The highly anticipated "cutting-edge model" Claude Mythos 5 has also been released to a small group of cybersecurity defenders and infrastructure providers, with Mythos 5 set to be deployed initially through the "Project Glasswing" initiative.
The model bases of Fable 5 and Mythos 5 are the same; however, Mythos 5 has removed some security restrictions.
Both models are priced at $10 per million input tokens and $50 per million output tokens, which is less than half the price of Claude Mythos Preview. However, this price is double that of Opus 4.8, one and a half times more expensive than GPT-5.5 for input, and 2/3 more expensive for output.
The official announcement states that starting today, Claude Fable 5 is available to all users through all channels, while Claude Mythos 5 is currently only available to Glasswing partners and will later be rolled out to the public through an expanded Trusted Access Program.
However, there is a specific time window to note: from today until June 22, Pro, Max, Team, and seat-based Enterprise subscribers can use Fable 5 for free without any additional charges.
Starting June 23, Anthropic will remove Fable 5 from the above subscription plans. Users who wish to continue using it after that date will need to consume usage credits. If capacity allows, we will extend the window of free usage.
Of course, Anthropic also stated that in the future, “when we have sufficient capacity, our goal is to reintegrate Fable 5 into the standard entitlements of the subscription plan. We will move as quickly as possible on this.”
Evaluation of the Capabilities of Claude Fable 5 and Claude Mythos 5
Anthropic claims that Fable 5 is its most powerful model to date, and from the benchmark table below, it is evident that this description is not an exaggeration; it is in “crush” mode in almost all dimensions...
In software engineering, Fable 5 can compress engineering work that would typically take months into a few days. In a Ruby codebase with 50 million lines of code, the model completed a migration covering the entire codebase in a day, a task that would require over two months for a full team to accomplish manually.
Moreover, the token efficiency of Fable 5 is higher than that of the Claude series models: in Cognition's FrontierCode evaluation, Fable 5 achieved the highest score under only a moderate effort setting.
In knowledge work, Fable 5 excels in complex analytical tasks. In Hebbia's Finance Benchmark focused on advanced reasoning abilities, Fable 5 obtained the highest score among all models, showing significant improvements in document-based reasoning, chart and table interpretation, problem-solving, and more.
Furthermore, Fable 5 has become a new SOTA model in visual tasks, capable of extracting precise values from detailed scientific charts, performing complex visual tasks, such as reconstructing the source code of a web application from a screenshot alone.
It also exhibits lower reliance on scaffolding mechanisms: for example, the previous Claude model, even with an "external" system providing additional assistance tools, struggled to play Pokémon FireRed, while Fable 5 successfully completed the game with minimal visual aids.
Moreover, Anthropic stated that in areas such as memory and context, drug design, molecular biology, genomics, and more, Fable 5 has demonstrated powerful capabilities.
Fable 5 can handle millions of tokens in long-running tasks and improve its output using its own notes. When Anthropic had the model play the card-building game "Slay the Spire," granting it access to persistent file-level memory, its performance improved by three times compared to Opus 4.8. Fable also more frequently reached the game's final levels.
Utilizing Mythos 5, Anthropic's in-house protein design expert improved the efficiency of certain stages in the drug design process by about 10 times.
Mythos 5 is Anthropic's first model capable of continuously proposing novel and compelling scientific hypotheses. In blind tests compared to Opus-level models, the company's scientists favored Mythos' molecular biology hypotheses about 80% of the time and have advanced several of these hypotheses to the experimental validation stage.
Meanwhile, a hypothesis proposed by Mythos regarding a new mechanism of Escherichia coli (E. coli) protein has been validated in a separate study by another laboratory investigating the same issue.
Recently, Anthropic released a report stating that large model self-play speed is increasing, prompting a call for a global pause on AI technology development. The latest Fable 5 release also came with a corresponding restriction: you cannot use Fable 5 to develop new large models.
“Restricting Claude's effectiveness in requests for advanced LLM development” is how they plan to further fortify their moat.
Finally, rumor has it that OpenAI's GPT-5.6 will also be released this week. Looking forward to Claude's next showdown with ChatGPT.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
