Sonnet 5 looks cheaper.
Check the tokenizer.
Claude Sonnet 5 arrived on 30 June 2026 with introductory pricing that reads as a substantial cut. Buried in the migration notes is a change that quietly moves the other way, and the two do not cancel out for long.
Sonnet 5 carries a one million token context window, up to 128,000 tokens of output, and Anthropic’s largest gains in coding and agentic work. It is a drop-in replacement for Sonnet 4.6: change the model ID and it runs. The pricing story is where it gets interesting, and it is worth ten minutes with a calculator before you assume your bill is going down.
The headline numbers
Introductory pricing of $2 per million input tokens and $10 per million output is in effect through 31 August 2026, after which standard pricing of $3 and $15 takes over. Sonnet 4.6 cost $3 and $15. So from September, the per-token price is exactly what it was before, and the introductory period is a migration window rather than a permanent reduction.
The tokenizer changed, and that is the real story
Anthropic states it directly in the migration notes: Sonnet 5 uses a new tokenizer, and the same input text produces approximately 30% more tokens than on Sonnet 4.6. Read that again with a price list in front of you. Identical text, identical per-token price from September, roughly 30% more tokens counted. That is not a saving.
What the arithmetic actually looks like
During the introductory window the per-token price is a third lower while the token count is around 30% higher, so like-for-like cost lands somewhere near where it was. From 1 September the discount ends and the token count does not, which is when the change shows up on the invoice. Anthropic’s own migration guidance says to recount your prompts using token counting rather than assuming, and that is the step to actually do rather than nod at.
Three things that now return an error
Setting temperature, top_p or top_k to a non-default value returns a 400. Manual extended thinking, deprecated on Sonnet 4.6, is removed and also returns a 400. Adaptive thinking is on by default, and turning it off is an explicit parameter. If you maintain code that tunes sampling parameters, that code will break on the day you change the model ID, not a week later when someone notices the output drifted.
Where it lands on capability
At maximum reasoning settings, reporting at launch put Sonnet 5 roughly in line with Opus 4.8 at its medium-to-high setting on the OSWorld-Verified and BrowseComp benchmarks. A clear improvement on Sonnet 4.6, and short of the top of the range. For a lot of teams that is the right trade, provided the cost modelling is honest. The mistake is treating a mid-range model as a cheap frontier model rather than as its own tier with its own jobs.
Take three prompts you send at volume, count their tokens on both models, and multiply by your monthly request count at September prices. You will have a real number instead of an impression, and if it comes out higher you have until 31 August to decide what to do about it. This kind of unglamorous arithmetic is a routine part of Claude Code coaching for teams running work at scale.
Things people
usually ask.
Subscription plans are not billed per token, so there is no invoice to check. Usage limits are still consumed by tokens, though, so the same work can take a slightly larger bite out of your allowance. The effect is real but harder to see than on the API.
Anthropic describes it as one, and for most integrations changing the model ID is genuinely all it takes. The exceptions are code that sets sampling parameters or uses manual extended thinking, both of which now return errors rather than warnings.
The opposite. The introductory window is the cheap time to test it against your real workloads and measure the token difference, which is exactly the information you need before the standard price arrives.