Cost baseline for one task: output costs 200 times input
Measured on 2026-11-14 — numbers from real runs
The raw data here comes from real runs; token counts and timings were parsed from the session files with tools/usage.py. Sample size and what does not generalise are stated on the page, with no averaging.
Question to answer
What does one real task actually cost? Is the money in input, cache hits, or output? How much does off-peak differ from peak?
Control groups
Take the real records of six completed tasks, price them at both off-peak and peak rates, and keep the three line items separate rather than merged.
Environment & version
- Host version
- 0.1.2-alpha.5
- Model ID
- host default route (no model ID recorded)
- Mode / permissions
- --profile headless / workspace-write / Windows 11 / node 22.22.2 / all runs off-peak
- Input files
- deepseek-flash per 1M: cache hit off-peak $0.003 / peak $0.006; cache miss off-peak $0.15 / peak $0.30; output off-peak $0.60 / peak $1.20. Off-peak is half of peak
Fields to log
| # | Field | Result |
|---|---|---|
| 1 | Price source and capture date | https://api-docs.deepseek.com/quick_start/pricing, captured 2026-11-14 |
| 2 | Off-peak and peak unit prices for all three items | deepseek-flash per 1M: cache hit off-peak $0.003 / peak $0.006; cache miss off-peak $0.15 / peak $0.30; output off-peak $0.60 / peak $1.20. Off-peak is half of peak |
| 3 | Cache-miss, cache-hit, and output tokens per task | cache-miss / cache-hit / output: pdf-index 6,594/305,792/6,699; excel-monthly 40,355/1,912,064/28,786; meeting-docx 31,958/1,731,968/37,642; bug-triage 21,468/188,544/12,160; prd 25,216/86,400/15,221; handout 26,266/89,216/11,069 |
| 4 | Cost per line item and total at off-peak | $0.005926 / $0.029061 / $0.032575 / $0.011082 / $0.013174 / $0.010849 (six-run total $0.102667) |
| 5 | The same numbers repriced at peak | Repriced at peak: $0.011852 / $0.058122 / $0.065150 / $0.022164 / $0.026348 / $0.021698 (total $0.205334) — exactly double |
| 6 | Which line item dominates | Output share: pdf 67.8% / excel 59.4% / meeting 69.3% / bug 65.8% / prd 69.3% / handout 61.2%. Cache-hit share: 15.5% / 19.7% / 15.9% / 5.1% / 2.0% / 2.5% |
| 7 | Order of magnitude per task (cents, dimes, dollars) | All six landed between $0.006 and $0.033, i.e. a few cents; the dearest was 5.5 times the cheapest |
What the runs showed
- All six real tasks landed in single-digit cents: $0.0059 to $0.0326. The dearest was meeting-docx at $0.0326, the cheapest pdf-index at $0.0059. The magnitude is the point: one structured document generation costs roughly twenty US cents' worth of a fraction.
- Output dominates cost, not input. Output is 59.4 to 69.3 percent across the six, median about 66. At official prices output is $0.60 per million — 200 times a cache-hit input token at $0.003, and four times a cache-miss token at $0.15. So asking it to say less saves more than asking it to read less.
- The cache-hit share grows with step count: 2.0 percent for the five-step prd task, 19.7 percent for the 44-step excel task. Every extra step resends the whole context, making step count the direct driver of cache reads.
- Peak hours cover the working day. The docs state: Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday. In UTC+8 that is 09:00-12:00 and 14:00-18:00 — the normal working day in China is entirely peak, so off-peak pricing only helps jobs run late at night or early morning. These runs all happened after 19:00 local and are priced off-peak.
- Peak and off-peak differ by exactly a factor of two (the docs say off-peak rates are half of peak). The same batch costs $0.205 by day and $0.103 at night.
Conclusion
A real task costs a few cents, with output dominating at about 66 percent because output is priced 200 times a cache-hit input token. The cache-hit share grows with step count, and peak hours coincide with the Chinese working day, doubling the bill. To save money, cut output and steps first — not input.
What does not generalise
- The six tasks differ in input size from tens to hundreds of kilobytes, so their costs cannot be compared head to head — only the structure and magnitude.
- All runs used deepseek-flash. On v4-pro at five to six times the price the absolute figures change, but the structural finding that output dominates holds, since all three prices scale together.
- Prices change at any time. Anywhere these numbers are quoted, the capture date must travel with them, or they are wrong within six months.
- These figures were never reconciled against an official invoice (no billing detail was available here). They are theoretical values computed from published prices, not actual charges.
Why we don’t write results yet
Prices change, so every entry must carry a capture date and a source link. A price table without a date is wrong within six months.