Anthropic shipped Claude Sonnet 5 today. It calls the model the most agentic Sonnet so far. That sounds like marketing copy, but there is something concrete underneath it: this model makes its own plan, reaches for tools like a browser or a terminal, and works a task to the end without you sitting over its shoulder at every step.
The interesting part is the position Sonnet 5 takes. Its performance lands close to Opus 4.8, Anthropic’s heaviest model, but the price sits a lot lower. So the work you used to send to the expensive model now moves down to the middle tier.
What is new
As of today, Sonnet 5 replaces Sonnet 4.6 as the default model in the free and Pro plans. It is available across every plan, from Free to Enterprise, and it becomes the new default in Claude Code for Pro users too. Through the API you get a context window of one million tokens, enough to hand it a hefty document or a whole codebase in one go. The model ID is claude-sonnet-5.
What it actually does better
Against Sonnet 4.6, the clearest gain shows up in independent work. Where 4.6 could stall halfway through a complex task, Sonnet 5 walks the whole route. It also checks its own output without being asked. And it holds up on bigger jobs, whether that is a large codebase or a long debugging session.
The numbers back that up. On SWE-bench Verified, a coding test, Sonnet 5 scores 92.4 percent. For comparison: GPT-5.4 reaches 57.7 percent there. On OSWorld-Verified, which measures how well a model operates a computer, it goes from 78.5 percent for Sonnet 4.6 to 88.3 percent. A human expert sits at 72.4 percent on that test, so the model is comfortably above that.
There are safety gains as well. Sonnet 5 makes up fewer facts than 4.6, is less inclined to tell you what you want to hear, and is harder to mislead through hidden instructions. The built-in protection against misuse is on by default.
How to steer it
Anthropic published a prompting guide alongside the release, and a few practical points stand out. Sonnet 5 takes your instructions more literally than its predecessor. It does not fill gaps on its own and does not guess what else you might have meant, so say exactly what you want and you get exactly that. It also decides how long an answer should be by itself, short for a simple question and longer for an open-ended one, which means you no longer have to steer on length. And it reaches for a tool on its own sooner and checks its own work more often. Prompts you wrote for Sonnet 4.6 keep working.
When to pick which
Anthropic now has three models with a clear role each. For a small business owner it comes down to this:
| Model | Best for | Cost |
|---|---|---|
| Haiku | Fast and simple work, short questions, high volume | Low |
| Sonnet 5 | Daily professional work, coding, tasks the model finishes on its own | Mid |
| Opus 4.8 | The heaviest, most critical jobs where you cannot afford a mistake | High |
For most work, Sonnet 5 is the obvious choice from now on. Only when accuracy matters above everything, and the higher price earns its keep, do you move up to Opus 4.8.
What it costs
Through 31 August 2026 there is an introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens. After that it becomes 3 and 15 dollars. Opus 4.8 runs at 5 and 25 dollars per million tokens, so for Sonnet 5 you pay roughly a half to a third of that. And if you just want to try it first, you can: Sonnet 5 is in the free plan too.
One partner who tested it early, Parashift, put it this way: they ran Sonnet 5 against dozens of their hardest real pull requests, and the model carried every one through to a tested, working result on its own.
The full announcement with all the benchmarks and details is on the official Anthropic blog.
This article, by the way, was written with Claude Sonnet 5. Felt fitting.
Want to know how to put models like this to work in your own day? Read along at sparkone.nl or email me at jeroen@sparkone.nl.