Anthropic announced yesterday the release of Claude Sonnet 5a more “normal” and conservative model than what we were used to. It is also cheaper, which makes it interesting to use in those AI agents that consume a lot of tokens. However, the most striking thing about Sonnet 5 is that it is a model that is especially capable in a key section: cybersecurity.
Sonnet 5 is designed to behave well. At Anthropic they define it as “the most agentic Sonnet model to date” and highlight its improvements in the use of tools such as browsers or the terminal. But above all they highlight how Sonnet 5 behaves better than Sonnet 4.6 in terms of “unwanted behavior” and is safer to use. The “System Card” of the model also confirms that we are dealing with a model that does not want to cross red lines.
Not very useful for finding vulnerabilities. The most striking thing is that Anthropic makes it clear that “it has a much lower capacity than our current Opus models to perform cybersecurity tasks.” That’s no accident: Anthropic has had so many problems with Mythos and Fable 5 who have probably wanted to bring a “risk-free” model to the market, even if that means making it more limited.
Benchmarks promise, but do not amaze. Internal tests show how Sonnet 5 represents an important qualitative leap compared to Sonnet 4.6 and is quite close to Opus 4.8 in agentic programming or in the use of the aforementioned tools (terminal, browser), something that undoubtedly makes it very interesting for AI agents that automate many tasks, such as Claude Cowork.
Fable 5 returns. The model comes just at a time when Anthropic was waiting for approval from the US government to be able to give access to its most powerful AI models:
A compelling model for companies. Sonnet 5 is an interesting model for use in companies. Claude Code is the company’s flagship product and its adoption is enhanced with models that can perform well in agentic tasks that complete objectives autonomously. That requires intensive token consumption, and Sonnet 5, close to the performance of Opus 4.8, has a fundamental advantage over it.
cheaper. While Fable 5 costs $10/$50 per million input/output tokens, Opus 4.8 costs $5/$25. That’s where things get interesting, because Sonnet 5 costs 3/15 dollarssignificantly cheaper, although during the first two months the cost will be 2/10 dollars as a promotional offer.
An old school model. The launch of Sonnet 5 is striking because Anthropic had left its “modest” models well behind. Haiku has not been updated since version 4.5, in October 2025. The last version of Sonnet was 4.6, in February 2026. Both seemed to have been relegated to the background and Anthropic had focused on Opus and especially on the new models of the Mythos family. That changes now, probably because the move with these latest models has gone wrong: they were so powerful that the US government ended up vetoing them and taking control.
GIPHY App Key not set. Please check settings