Correctness drift is interesting. Billing drift is expensive.
Every reading records the tokens each model returned for the same frozen prompts and how long it took. Priced against a dated public table, that is a daily cost for an identical run: the number that moves when a model starts answering longer, or slower, at the same list price. Latest reading 2026-09-04.
Billing drift: the same battery, priced every day.
| Model | Run cost | vs previous | 30 runs | Per passed task | Tokens in / out | p50 latency | Trend |
|---|---|---|---|---|---|---|---|
| haiku-4-5 | $0.0016 | 0.0% | after 30 runs | $0.0001 | 546 / 202 | 1.21s | |
| sonnet-5 | $0.0072 | -26.0% | after 30 runs | $0.0006 | 692 / 584 | 1.95s | |
| opus-5 | $0.0241 | -0.1% | after 30 runs | $0.0020 | 692 / 826 | 2.62s | |
| fable-5 | $0.0338 | +3.2% | after 30 runs | $0.0028 | 692 / 538 | 3.71s | |
| fable-5-1 | $0.0298 | +18.0% | after 30 runs | $0.0025 | 716 / 453 | 4.72s |
This measures the battery, not your workload. Winthrop's Token Audit does the same arithmetic on your actual usage export. Prices are transcribed by hand and dated; the full table and the per-category record are on /costs.
| Model | math | instruction | extraction | reasoning | classification | json | code |
|---|---|---|---|---|---|---|---|
| haiku-4-5 | 10/10 | 15/20 | 5/5 | 5/5 | 5/5 | 5/5 | 10/10 |
| sonnet-5 | 10/10 | 19/20 | 5/5 | 5/5 | 5/5 | 5/5 | 10/10 |
| opus-5 | 10/10 | 18/20 | 5/5 | 5/5 | 5/5 | 4/5 | 10/10 |
| fable-5 | 4/4 | 8/8 | 2/2 | 2/2 | 2/2 | 2/2 | 4/4 |
| fable-5-1 | 4/4 | 8/8 | 2/2 | 2/2 | 2/2 | 2/2 | 4/4 |
Passed over attempted, summed across the window. This is the evidence base a downgrade recommendation can cite: which model has actually been getting which kind of task right, and for how many days.
| Family | Matches | Input | Output |
|---|---|---|---|
| Claude Fable/Mythos 5 | fable-5|mythos-5 | $10 | $50 |
| Claude Opus (4.5+) | opus-(5|4-8|4-7|4-6|4-5) | $5 | $25 |
| Claude Opus (4.1 and earlier) | opus | $15 | $75 |
| Claude Sonnet 5 | sonnet-5 | $2 | $10 |
| Claude Sonnet (4.x) | sonnet | $3 | $15 |
| Claude Haiku 4.5 | haiku-4-5 | $1 | $5 |
| Claude Haiku 3.5 | haiku-3-5|3-5-haiku | $0.8 | $4 |
| Claude Haiku 3 | haiku | $0.25 | $1.25 |
Transcribed by hand from Anthropic's published pricing and dated. This is Winthropic's transcription, not Anthropic's document; it may lag a change. Corrections are appended, never edited in place: /prices.json carries every card. Sonnet 5's launch intro price ($2/$10) was made permanent 2026-08-10; the scheduled 09-01 rise to $3/$15 was cancelled. 4.x Sonnets stay $3/$15.
Data: /api/costs.json. Same page as markdown: /costs.md. Winthropic is not affiliated with Anthropic.