Anthropic launches Claude Sonnet 5 AI model with coding, safety upgrades as Fable and Mythos controls lifted

Anthropic PBC today debuted Claude Sonnet 5, a midrange large language model that outperforms its predecessor in several areas.

The LLM will be the default option in the consumer tiers of the company’s Claude chatbot service.

Anthropic’s commercially available LLMs are organized into three product families: the entry-level Haiku series, Sonnet and the high-end Opus lineup. In April, the company debuted two LLMs called Mythos 5 and Fable 5 that are even more capable than Opus. However, they’re not yet broadly accessible.

Anthropic measured Sonnet 5’s coding capabilities using two benchmarks called SWE-Bench Pro and Terminal-Bench 2.1. The model improved upon the scores of its predecessor by 5.1% and 13.4%, respectively. GPT-5.6 Terra, OpenAI Group PBC’s competing midrange LLM, outperformed Sonnet 5 on Terminal-Bench 2.1 by about 4%.

Anthropic also tested how its new LLM fares in other areas. The model set a score of 1,618 on GDPval-AA v2, a benchmark that includes knowledge work tasks spanning 44 professionals. Sonnet 4.5 earned 1,395 points.

One of the contributors to Sonnet 5’s increased output quality is that it’s more autonomous. According to Anthropic, users who tested the model before its release reported that it sometimes double-checks its output without instructions to do so. Additionally, Sonnet 5 can perform tasks that are too difficult for its predecessor.

Increased LLM autonomy can create cybersecurity risks in some cases. According to Anthropic, Sonnet 5 is better than Sonnet 4.6 at fending off such risks. In particular, it’s more adept at blocking malicious requests and prompt injection attacks. A prompt injection attack is a malicious instruction hidden in the data analyzed by an LLM.

The new model includes guardrails that prevent hackers from using it to launch cyberattacks. According to Anthropic, the model poses a limited cybersecurity risk because it can’t develop working exploits.

Sonnet 5 is the new default model in the free and consumer-focused Pro tiers of the Claude chatbot service. It will also become available in the Max, Team and Enterprise plans. Developers, meanwhile, can access the model through Anthropic’s application programming interface. Sonnet will be priced at $3 per million input tokens and $15 per million output tokens starting in September, slightly more than OpenAI’s midrange Terra.

Anthropic customers will soon also gain access to two other new models. Earlier this month, the company debuted a flagship LLM called Mythos 5 and a scaled-down version known as Fable 5. The latter model includes stricter guardrails that block potentially risky prompts.

The U.S. government imposed export controls on Mythos 5 and Fable 5 a few days after their introduction. In response, Anthropic paused the models’ rollout. The company disclosed today that the controls have been lifted, which will allow it to start restoring access on Wednesday. Anthropic plans to make Fable 5 broadly available, while Mythos 5 will be accessible only to a limited number of trusted organizations.

Image: Anthropic