Anthropic Relaunches Sonnet 5 and Restores Fable, Mythos

Anthropic Relaunches Sonnet 5 and Restores Fable and Mythos After Restrictions

A New Model for Anthropic: Claude Sonnet 5

On June 30, 2026, Anthropic unveiled its latest artificial intelligence model, Claude Sonnet 5. This model is designed to offer agentic capabilities comparable to those of the most powerful models, but at a reduced cost. On the same day, U.S. authorities lifted the restrictions that had prevented access to the Fable 5 and Mythos 5 models.

Claude Sonnet 5 succeeds Sonnet 4.6 in Anthropic's lineup of intermediate models. This new model promises agentic performance that approaches that of the Opus 4.8 model while being offered at a more affordable price.

This model becomes the default choice for both free and Pro offerings of Claude. It is also accessible to subscribers of the Max, Team, and Enterprise plans, as well as in Claude Code and via the API under the identifier claude-sonnet-5.

Until the end of August, Claude Sonnet 5 is available at a launch price of $2 per million input tokens and $10 per million output tokens. After this period, the standard price will increase to $3 for input and $15 for output. Anthropic has introduced a new tokenizer that increases token consumption by 35% for the same text, but assures that the launch price is adjusted to compensate for this increase.

Agentic Performance of Sonnet 5

Claude Sonnet 5 has been designed to execute a series of actions, navigate the internet, manipulate a terminal, or write code, all while autonomously verifying its own work. On the SWE-bench Pro benchmark, which evaluates task resolution in real-world development conditions, Sonnet 5 achieved a score of 63.2%, surpassing Sonnet 4.6, which scored 58.1%. On Terminal-Bench, which measures mastery of terminal commands, Sonnet 5 reached 80.4%.

Anthropic offers different levels of effort for users to balance cost and performance. When pushed to its maximum, Sonnet 5 manages to match the results of Opus 4.8 on certain tasks, a model previously reserved for the most demanding uses.

In terms of safety, Sonnet 5 exhibits a lower rate of undesirable behaviors compared to Sonnet 4.6, with fewer hallucinations and a reduced tendency to flatter the user. However, it has not been specifically trained for cybersecurity tasks. In a test conducted with Mozilla on Firefox, Sonnet 5 failed to produce a functional exploit.

Anthropic clarifies that Sonnet 5 is capable of performing some common and non-dangerous cybersecurity tasks, but its performance is inferior to that of models like Opus 4.8 and Mythos 5 for potentially dangerous skills such as exploit development.

Lifting of Restrictions on Fable 5 and Mythos 5

Anthropic also announced the lifting of export restrictions imposed by the United States, which had blocked access to Fable 5 and Mythos 5. On June 12, the Department of Commerce had required the immediate suspension of access to these models for foreign users after discovering a method to circumvent Fable 5's cybersecurity safeguards. Unable to verify users' nationality in real-time, Anthropic had disabled these models for all its clients.

Since then, the company has developed a new security classifier, specifically targeting the reported technique and blocking it in over 99% of cases, according to them. Mythos 5 had already been redeployed on June 26 to about a hundred U.S. organizations operating critical infrastructure. Fable 5, its consumer version, has been back since July 1 on Claude, Claude Code, and the API, with a return expected on AWS, Google Cloud, and Microsoft Foundry.

The U.S. government imposed a similar restriction on OpenAI, which had to limit the deployment of GPT-5.6 Sol to a circle of partners approved by the White House.