Claude Sonnet 5 undercuts agents at $2 — but every benchmark is Anthropic's own

Read the launch post and Claude Sonnet 5 sounds like a gift: near-Opus-4.8 quality, built to plan and drive browsers and terminals on its own, at $2/$10 per million tokens instead of Opus's $5/$25. Anthropic shipped it on 30 June 2026, made it the default model for Free and Pro, and pointed it squarely at anyone who winced at last quarter's agent bill. On the substance, it's good. On the framing, it's worth a harder look.

Here's the catch nobody in the launch-day headlines led with: almost every concrete performance number sits inside Anthropic's own blog post. The partner testimonials are early-access, not independent. That doesn't make them wrong — it makes them unverified. And Anthropic itself quietly edited the launch post on 30 June to fix a BrowseComp cost-performance chart that used a simpler methodology and underestimated Sonnet 5. When a vendor is correcting its own charts on day one, the honest read is: wait for independent benchmarking before you rewire your stack. None of that makes Sonnet 5 a bad model — it just means the launch-day numbers deserve the same scepticism you'd apply to any vendor grading its own homework.

Credit where due: the numbers Anthropic does share are strong. On an agentic coding benchmark Sonnet 5 hit 63.2% — behind Opus 4.8's 69.2% but comfortably ahead of Sonnet 4.6's 58.1% — and on a knowledge-work benchmark it slightly outperforms Opus 4.8. An adjustable effort level lets you dial cost against performance. Near-flagship quality at a mid-tier price is a real thing here, not marketing vapour. Just don't confuse a first-party chart with a settled fact.

Sonnet 5 ships with a new tokenizer, and that quietly moves the goalposts. The same text can map to roughly 1.0x to 1.35x more tokens depending on your content. Anthropic set the introductory price so switching from Sonnet 4.6 stays roughly cost-neutral — but 'roughly' and 'depending on your content' are doing heavy lifting. Your actual bill is a function of your prompts, not their average. Model it before you assume the sticker price.

The pricing itself is straightforward — until the intro window closes:

The genuinely sharp decision is the one Anthropic buried in the safety section: it explicitly chose not to train Sonnet 5 on cybersecurity tasks, leaving it substantially weaker at exploit development than Opus 4.8. Overall it shows a lower rate of undesirable behaviours than Sonnet 4.6 and is safer in agentic contexts. In a month when the US government yanked Fable 5 and Mythos 5 over cyber capability, shipping your cheapest, most widely deployed agent model as one that deliberately can't hack well is both good safety and shrewd politics. That's not a hedge — that's Anthropic reading the regulatory room correctly.

So where does that leave the buyer? Sonnet 5 is a strong, well-timed, defensibly-safe model that will cut a lot of agent bills. It's also a first-party story with an asterisk on the price and a chart the vendor had to correct before lunch. Deploy it — but benchmark it yourself, and price it on your own tokens.