On 30 June 2026 Anthropic released Claude Sonnet 5. The headline is a plain one for anyone running a firm. A model that sits close to the top of the range in what it can do now costs a good deal less to run, and that changes the calculation for putting AI into everyday legal work.
What has changed
Sonnet 5 is the most agentic model in the Sonnet line to date. It can plan a piece of work, use tools such as browsers and terminals, and carry a task through with less hand holding than earlier versions needed. Anthropic reports gains in reasoning, tool use, coding and knowledge work, and the model checks its own output without being asked to. Its capability sits close to the top Opus 4.8 model, which until now carried a higher price for that level of performance.
The pricing is the part a firm should read carefully. Until 31 August 2026 Anthropic is charging two dollars per million input tokens and ten dollars per million output tokens, an introductory rate. After that the price settles at three and fifteen. Sonnet 5 is also the default model on both the free and paid consumer plans, so many of your staff will already be using it whether the firm has decided to or not.
Why the price matters more than it looks
A cheaper model that performs near the top of the range lowers the cost of the work you might route through it, which means tasks that were not worth automating a year ago start to look worth it. Summarising a long bundle, drafting a first attempt at a letter, pulling the key dates out of a file, all of these become cheaper to run at scale. For a small firm watching every licence and every hour, that is the shift that moves AI from a curiosity to a line in the budget.
The catch runs in the opposite direction. A more autonomous tool does more on its own, and that raises the supervision stakes rather than lowering them. When a model plans its own steps and uses tools to carry them out, the distance between the instruction and the finished work grows, and a person still has to own what comes back. Cheaper and sharper does not mean safer to leave alone.
The fact that Sonnet 5 checks its own output without being asked is welcome, and it is not a substitute for a person doing the same. A model reviewing its own work shares the blind spots that produced the work, and it cannot be held to account for what it misses. That self-check is a feature that reduces the number of obvious errors reaching you. It does not shift the responsibility for the ones that get through, which stays where it has always been.
The point for a firm
The direction of travel is clear enough. Capable models are getting cheaper, and they will keep doing so. That makes AI easier to justify and easier to spread across the firm, and it makes the discipline around the tool the thing that decides whether the firm benefits or suffers. A cheaper model does not change your duty to the client or the court.
Match the tool to the task rather than reaching for the most capable option out of habit. Keep a named person accountable for anything the tool produces. Revisit your choices as prices and capability move, because the right answer this quarter will not be the right answer next year. The release of Sonnet 5 is a good moment to look again at what you route through AI and who signs it off before it leaves the firm.
Anthropic’s newsroom has the release details for anyone who wants the technical account.
If cheaper capability changes the sums for your firm, the readiness audit prices the opportunity for your practice rather than in the abstract: see the packages.