Member-only story
Anthropic’s rate card says $2 per million. Artificial Analysis measured $2.29 per completed task — more than Opus 4.8. The gap is one setting, and the migration docs quietly remapped it. #
On June 30, Anthropic cut Sonnet’s price to $2 per million tokens. Three days later, Artificial Analysis published the number that actually decides your bill: $2.29 per completed task — roughly double Sonnet 4.6, and about 15% more than Opus 4.8, the flagship that costs over twice as much per token.
Both numbers are true. Only one of them shows up on your invoice.
I’ve written before about auditing a dual-agent stack over 90 days, and the lesson that survived that audit applies here with interest: sticker prices tell you what a token costs, and almost nothing about what your work costs. Sonnet 5 is the sharpest version of that gap I’ve seen a lab ship. The per-token rate went down. The per-task cost, at the settings most teams will inherit without noticing, went up past the flagship.
Here’s the claim, early, so you can disagree efficiently: which model you pick is now the secondary cost decision. The effort dial is the primary one — and if you…