Claude Sonnet 5 is Anthropic's lower-priced answer to its Opus models for coding and AI agent work. Anthropic says the new Sonnet can plan, use browsers and terminals, and run on its own at a level that recently needed larger, more expensive models.
Sonnet 5 succeeds Claude Sonnet 4.6 as Anthropic's default Sonnet model for Free and Pro users. Sonnet 4.6 had added a one million token context window in beta. The current model docs list a one million token context window and up to 128,000 output tokens in the synchronous Messages API.
Anthropic Sonnet 5 launch, Claude model docs, Anthropic Sonnet 4.6 launch
What changed from 4.6
Anthropic's comparison says Sonnet 5 improves on 4.6 in coding, search, computer use, visual reasoning, and professional work. Opus 4.8 still looks stronger for higher-accuracy agent search, computer use, and the most complex tasks.
Pricing does the hard work. It launches at $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that, it moves to $3 and $15, the normal Sonnet price. Opus 4.8 costs $5 and $25.
That sounds generous. But Anthropic also changed how Sonnet chops text into billable pieces, called tokens. The same prompt can turn into 1.0 to 1.35 times as many tokens, so Anthropic says the launch discount is meant to make the move from 4.6 roughly cost neutral.
Anthropic Sonnet 5 launch, Claude model docs
Safer, but fussier
Anthropic's system card gives the useful safety details. The model makes fewer unsupported claims, flatters users less than 4.6, and is better at resisting hidden instructions that try to hijack an AI agent. The trade is that it also blocks more harmless cyber requests in those tests.
Anthropic is blunt about security work. Anthropic says it did not set out to train Sonnet 5 for cybersecurity tasks, and the model is much less able than current Opus models to help with dangerous hacking. That trade may make the default model less useful for some security research work.
Worth the switch?
For everyday Claude users who want the default Sonnet lane, yeah, probably. Sonnet 4.6 was already Anthropic's everyday default for Free and Pro users. Sonnet 5 keeps that lane and, according to Anthropic, gets closer to Opus 4.8 at a lower price.
But it isn't a free lunch. Very large prompts may eat some of the launch discount, and Opus 4.8 still looks better for the harder agent jobs that Anthropic names. Anthropic has a sensible new default, not a new ceiling.
