As transport agentic capabilities turns into desk stakes amongst basis mannequin firms, Anthropic is releasing Claude Sonnet 5, a extra highly effective and agentic model of the lab’s midsize mannequin.
“It could possibly make plans, use instruments like browsers and terminals, and run autonomously at a degree that, just some months in the past, required bigger and dearer fashions,” Anthropic stated in a blog post.
That framing mirrors what OpenAI and Google have stated about their very own latest releases. OpenAI’s GPT-5.6 Sol was launched in preview final week, and it is usually the agency’s most agentic mannequin but, permitting customers to separate work throughout subagents for longer autonomous duties. Google’s Gemini 3.5 Flash, which launched in Could, was pitched as a shift from a conversational chatbot to an agentic device that plans, builds, and iterates on actual work with minimal human enter.
Sonnet 5’s pitch is affirmation that agentic functionality is the brand new baseline expectation at each value tier. Now the differentiator isn’t going to be who can do agentic work greatest, however how cheaply they’ll do it and the way reliably with out human oversight.
Sonnet 5 guarantees efficiency near that of Opus 4.8, however for a lot decrease prices. Beginning Tuesday, Claude Sonnet 5 would be the default mannequin at no cost and Professional plans and is out there for each subscription.
At launch, Sonnet 5 is priced at $2 per million enter tokens and $10 per million output tokens via August 31, after which the worth will soar to $3 per million enter tokens and $10 per million output tokens. That makes Sonnet 5 cheaper than Opus 4.8, in addition to OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Professional. (It’s nonetheless dearer than Gemini 3.5 Flash.)
The brand new mannequin additionally demonstrates important enhancements over its predecessor Sonnet 4.6, released in February, on agentic efficiency like reasoning, device use, software program coding, and data work, based on Anthropic.
For instance, on one benchmark, Sonnet 5 scores a 63.2% on agentic coding, in comparison with Opus 4.8’s 69.2% and Sonnet 4.6’s 58.1%. On a data work benchmark, Sonnet 5 really barely outperforms Opus 4.8, which is thought for profitable on fixing the toughest issues like making refined judgment calls and deep analysis.
“Opus 4.8 continues to be the mannequin of alternative for larger accuracy on these duties, however Sonnet 5 supplies builders with lower-priced choices which might be of a lot larger high quality than what was beforehand obtainable,” Anthropic says. “Between Sonnet 5 and Opus 4.8, customers can modify the hassle degree to seek out the proper stability of price and efficiency.”
In line with testers cited within the weblog submit, Sonnet 5 additionally excels at ending advanced duties the place earlier mannequin variations would have stopped brief and “checks its personal output with out explicitly being requested.”
“We handed Claude Sonnet 5 a two-part job — replace Salesforce account tiers, ship a launch announcement to enterprise contacts — and it completed finish to finish,” Daniel Shepard, a senior engineer at Zapier, stated in an announcement. “That used to stall midway. For day-to-day automation, it’s a no brainer. ”
On security, Sonnet 5 additionally demonstrates a decrease fee of “undesirable behaviors” like cooperation with misuse and deception than its predecessor, making it safer to make use of in agentic contexts. It’s higher at refusing malicious requests and sidestepping hijack makes an attempt in prompt-injection assaults. It additionally hallucinates and engages in sycophantic conduct at a decrease fee than Sonnet 4.6.
That stated, it’s not on the identical degree as Opus 4.8 and Claude Mythos Preview on the subject of misaligned conduct. “Evaluations additionally present that it has a a lot decrease means to carry out harmful cybersecurity duties than our present Opus fashions,” reads the weblog submit.
Lovable co-founder Fabian Hedin stated in an announcement that Claude Sonnet 5 “refuses unsafe requests cleanly and persistently.”
“At Lovable, we’re placing highly effective instruments within the fingers of thousands and thousands of builders,” Hedin stated. “A mannequin that is aware of when to say no is simply as essential as one which is aware of find out how to construct.”
Whenever you buy via hyperlinks in our articles, we may earn a small commission. This doesn’t have an effect on our editorial independence.