Harsh Govind ran the numbers on a model switch the way finance teams actually evaluate one — projected operating cost, not sticker price. Claude Sonnet 5's published token prices sit roughly a third below Sonnet 4.6's. The projected bill fell about 12%. The missing two-thirds of the discount went somewhere specific: Anthropic's own documentation says the new tokenizer produces approximately 30% more tokens for the same input text.
The arithmetic is unforgiving. Your bill is tokens multiplied by price per token. Cut the price by a third and raise the token count by 30%:
| Token price | Tokens per task | Bill per task | |
|---|---|---|---|
| Sonnet 4.6 | 1.00× | 1.00× | 1.00× |
| Sonnet 5 | 0.67× | 1.30× | 0.87× |
0.67 × 1.30 = 0.87. Thirteen percent off — right where Govind's 12% landed. The discount leaked through the tokenizer before it ever reached the bill.
This is the general rule, and it is worth stating plainly: never price the token; price the task. Tokenizers differ between models and vendors. Output verbosity differs. Retry rates differ. The only number that decides anything is what a completed unit of work costs end to end — which is why the Price Index tracks the output premium and cache behavior alongside the headline rates, not instead of them.
Govind's teardown ends with three checks worth stealing before any switch. First, agree on what "better" means — finance wants a smaller bill, product wants a better experience, engineering wants new capabilities, and a switch everyone approved can still disappoint everyone. Second, compare the migration against improving what you have: caching repeated content and routing routine work to lighter models can beat a migration without one. Third, test the demo against ordinary, awkward, real work — and measure response time next to quality, because a smaller bill that makes people slower is not a saving.
The cheapest model on paper can leave the most expensive work to your people. That sentence is the whole essay.
