Claude Sonnet 5: the price per token is not the price per task

Anthropic has released Claude Sonnet 5. The headline almost writes itself: close to Opus 4.8, at an introductory price of 2 dollars per million input tokens. I still read the announcement twice, because of a different paragraph. Sonnet 5 uses a new tokenizer, and the same text now maps to 1.0 to 1.35 times as many tokens as before. Anyone comparing prices per token right now is comparing two different units of measurement.

What is Claude Sonnet 5?

Sonnet 5 is, per Anthropic, the most agentic Sonnet yet: it makes plans, uses tools like browsers and terminals, and runs autonomously at levels that previously required larger models. Compared to Sonnet 4.6, Anthropic cites substantial improvements in reasoning, tool use, coding, and knowledge work. On the agentic evaluations BrowseComp and OSWorld-Verified, the model covers a wider range of cost-performance points than its predecessor and in some cases matches the capability level of Opus 4.8. On the API it runs as claude-sonnet-5. It is now the default model on the Free and Pro plans, available to Max, Team, and Enterprise users, in Claude Code, and on the Claude Platform, with rate limits raised across Chat, Cowork, Claude Code, and the Platform. On safety, Anthropic reports a lower rate of undesirable behaviors than Sonnet 4.6, substantially weaker performance than Opus 4.8 on exploit development, and the cyber safeguards ship enabled by default, as with Opus 4.7 and 4.8.

What does it really cost?

On paper the math is simple. Through August 31, 2026, the introductory price applies: 2 dollars per million input tokens, 10 dollars for output. After that it is 3 and 15. Opus 4.8 sits at 5 and 25. Except the unit has changed: the new tokenizer splits the same content into 1.0 to 1.35 times as many tokens, depending on the content type. Anthropic itself says the introductory price is set so that the transition is roughly cost-neutral. The reverse conclusion is not in the announcement, but it is obvious: the standard price is not. From September you pay 3 and 15 on a token count that, depending on the content, runs up to 35 percent above the old tally.

Why the 1.35 factor matters for agents in particular

For a single chat prompt the difference is pocket change. For an agent run it is not, because there the same context passes through the model on every loop: system prompt, tool definitions, file contents, intermediate results. A factor on the token count applies to every one of those rounds, not just the first. Hence my practical advice: do not compare price tables, compare cost per task. Take two or three typical runs from your daily work, run them on Sonnet 5 and on your current model, and look at the actual bill. The weeks until the end of August are the right window for that, while the cost-neutral introductory price cushions the switch.

The better tool, honestly priced

When Opus 4.8 came out, I wrote that for my work the benchmarks decide little; the everyday behavior does. The same line applies here, with the sign flipped: a Sonnet that carries agentic tasks at this level is probably the more economical tool for many routine runs. But “cheaper per token” is a claim, “cheaper per task” is a measurement. As everywhere in agentic engineering, it is not the spec sheet that decides but the check against your own runs. You do that check, not the model.

Sources

• Anthropic: Introducing Claude Sonnet 5