Anthropic Unveils Fable 5 and Mythos 5: New Claude Models with Benchmarks

Anthropic released Fable 5 and Mythos 5 models, highlighting new benchmarks.

On June 9, Anthropic released two versions of the Claude model family. The company describes Fable 5 as a Mythos-class solution, but safe for general use. Claude Mythos 5 is a “private” base model with relaxed restrictions in certain areas.

The key difference between them lies in the operation of protective classifiers. The public model monitors sensitive queries in cybersecurity, biology, chemistry, and attempts to distill the model. In such cases, the request is redirected to Claude Opus 4.8.

“To release the model safely and quickly, we have configured these protective measures conservatively — sometimes they react to harmless requests, although on average they trigger in less than 5% of sessions,” the company emphasized.

Fable 5 is available through a classic subscription on the web interface and Claude API.

Both models are priced the same: $10 per 1 million input tokens and $50 per 1 million output tokens. The company noted that this is less than half the cost of Claude Mythos Preview.

More on the Closed Flagship

At launch, Claude Mythos 5 was opened to some trusted participants, including members of Project Glasswing. Within the program, the new model replaced Claude Mythos Preview.

Anthropic plans to expand access for cybersecurity tasks and specific biomedical scenarios by the end of 2026.

Like Fable, Mythos has restrictions, but it offers expanded access in certain areas. 

Anthropic called the closed solution its strongest model for cybersecurity tasks, which “excels at detecting and exploiting software vulnerabilities.” Results in biology also show acceleration in certain stages of drug development by about 10 times. 

“Mythos 5 uses protein design and bioinformatics tools without human involvement, matching or often surpassing experienced human operators. […] Nine out of 14 protein targets in the study yielded promising compounds for drug development, which we are currently exploring,” the company emphasized. 

For biomedical researchers, Anthropic is preparing a separate access circuit to the Fable 5 version without restrictions on biology and chemistry, but retaining censorship in cybersecurity.

First Publicly Available Mythos-Class Model

Anthropic calls Fable 5 the strongest of the Claude models released to the public. The focus is on long and complex tasks. The company highlights programming, analytical work with documents, vision, memory, and scientific research as key areas.

The startup has not disclosed engineering details of the model’s architecture. Information on the number of parameters, training scheme, and stack description is also unavailable. An open license has not been published either.

During preliminary testing, Stripe reported that Fable 5 compressed months of development into a few days. The solution transferred a Ruby codebase of 50 million lines in just one day. Analysts noted that such a task would have taken a team of programmers two months. 

Fable 5 is also more cost-effective in terms of token expenditure than previous Claude versions: in the FrontierCode by Cognition evaluation, the model shows the best results among advanced AI solutions.

In the Hebbia Finance Benchmark test for logical reasoning analysis, Fable 5 scores the highest among all models.

One of the model’s main features is “vision.” The AI can extract precise numbers from detailed scientific diagrams and perform complex tasks, such as reconstructing the source code of a web application from screenshots alone. 

Additionally, Fable 5 requires fewer auxiliary tools. For example, previous Claude models struggled with playing Pokémon FireRed even with additional tools, but the new version completed it with minimal assistance.

Fable 5 has improved memory, maintaining focus on millions of tokens in long tasks and enhancing results using its own notes.

Earlier in June, Anthropic representatives warned about the risks of AI self-improvement. According to internal data, more than 80% of the code for the company’s current products was written by Claude.