Claude Sonnet 5 did not go up today. But Opus 5 costs 30% more than Opus 4.6 at the same price.
Anthropic cancelled the increase scheduled for 1 September. It also changed the tokenizer, which makes two rows of its own rate card no longer comparable.
By Parminder Kumar Sharma · · 6 min read
The increase that did not happen
Sonnet 5 launched at $2 per million input tokens and $10 per million output, described at the time as introductory pricing running to 31 August 2026. A rise to $3 and $15 was scheduled for today. Several AI newsletters and trackers flagged it over the weekend.
It was cancelled. Anthropic’s pricing documentation, read this morning, carries the note plainly:
That is worth knowing on its own if you budgeted for it. But it is the smaller of the two things on that page, and the other one is not being reported at all.
Every price is per million tokens, and the tokens changed
Four notes below the pricing table, Anthropic documents a tokenizer change:
Both facts are on the same page. They are simply never read together.
A rate card denominated in tokens is only comparable between two models if a token means the same thing in both. From Claude 4.7 onward it does not. The same document, unchanged, becomes about 30% more tokens, and the headline rate is charged against the larger count.
The headline price, and the price for the same document
Sonnet looks a third cheaper and is an eighth cheaper
Sonnet 4.6 was $3 per million input. Sonnet 5 is $2. The rate card says a 33% cut.
Normalise to one million tokens of text under the old tokenizer. On Sonnet 4.6 that costs $3.00. The same text on Sonnet 5 becomes about 1.3 million tokens, at $2 per million, so $2.60. The actual saving is 13%, and the same ratio holds on output, $15.00 against $13.00.
Still a genuine reduction. Just under half the one the rate card advertises.
Opus looks unchanged and is 30% more expensive
This is the one worth carrying away.
Opus 4.6 is $5 per million input and $25 per million output. Opus 5 is $5 and $25. The rate card is identical, so a straight comparison says the price did not move.
The same million tokens of text costs $5.00 on Opus 4.6. On Opus 5 it becomes roughly 1.3 million tokens, so $6.50. Output goes $25.00 to $32.50.
An unchanged rate card, and a 30% increase in what the same work costs.
The same document, priced four ways
| Model | Rate per MTok | Tokenizer | Cost for the same text | Against its predecessor |
|---|---|---|---|---|
| Sonnet 4.6 | $3 | previous | $3.00 | - |
| Sonnet 5 | $2 | newer | $2.60 | 13% cheaper, not 33% |
| Opus 4.6 | $5 | previous | $5.00 | - |
| Opus 5 | $5 | newer | $6.50 | 30% more expensive |
None of this is hidden. Anthropic states the tokenizer change in its own documentation, gives the approximate figure itself, and notes that the real increase "depends on the content and workload shape". What is absent is any indication on the pricing table that two rows of it are denominated in different units.
Why your own measurements will not match
This is the same failure this site has documented twice before: a number that is accurate about one thing being read as a number about a different thing. Row counts read as people. Bytes read as records. Now tokens read as text.
What to do this week
Take this with you
If you are budgeting Claude usage or comparing models
- Recompute any model comparison you made on rate cards alone. A per-token price is only comparable between models sharing a tokenizer, and Claude 4.7 is the boundary.
- Measure your own ratio with the token counting endpoint. Run a representative sample of real traffic through both tokenizers and derive your own multiplier rather than using 30%.
- Re-examine any Opus 4.6 to Opus 5 migration you costed as neutral. The rate is identical and the spend is not, so a migration signed off as price-flat may be a 30% increase.
- Remove the Sonnet 5 price rise from your forecast. It was scheduled for today and Anthropic has confirmed it will not occur.
- Watch caching and batch discounts, which are multipliers on the base rate. They scale the new rate, so they do not offset the token inflation, they reduce it proportionally.
The position
Nothing here is concealed and nothing here is wrong. Anthropic publishes the rates, publishes the tokenizer change, publishes its own estimate of the size, and even flags that the estimate is workload dependent. Every fact in this piece came off one documentation page in a single reading.
The problem is that a rate card is a comparison instrument, and this one silently stopped being comparable at Claude 4.7. A buyer reading down the table sees Opus 4.6 and Opus 5 at the same price and concludes the price held. It did not, for any given piece of work, and only a footnote four notes below the table would tell them.
The practitioner lesson is older than any of this. When a price is quoted per unit, the first question is whether the unit changed. Here it did, by about 30%, and it is documented in a place nobody comparing prices would think to look.
