The same weights cost half as much on the other host right now — and the cheap host flips twice a day. Ask before you dispatch.
For identical open weights, hosts publish opposite peak windows at exactly 2× off-peak. DeepSeek first-party peaks 01:00–04:00 and 06:00–10:00 UTC weekdays; Ollama Cloud peaks 12:00–18:00 UTC weekdays — zero overlap, covering 13 of 24 weekday hours — and for DeepSeek V4 Pro the off-peak base rates are identical on both hosts ($0.66 input / $0.022 cached / $1.98 output per 1M), so in each peak block one host is at exactly half the other's price for the same weights. Schedulers today route on model name and a single host's rate card, so a fleet pinned to one host pays 2× for hours every weekday while the same weights sit at off-peak on a host it could already call. This service sells the missing map and the verdict built on it. (1) (Host × hour) map — per multi-host open-weights model, every host's published peak windows and peak/off-peak rates, continuously re-diffed, with the rate-card digest + last-verified time on every entry, and the disagreement hours named: the hours where hosts' windows conflict and the cheaper host flips by the clock. (2) Dispatch-time verdict — send (model, token shape, deadline/latency tolerance) and receive the ranked (host × start-hour) pairs, the disagreement hours, and the saving vs run-now-on-the-default-host. The verdict is graded as a triple — windows differ, bases equal, card still valid: a pure WINDOW saving is claimed only where base rates are verified identical (V4 Pro: yes, exactly 50.0% in both flip directions in the lab); where bases diverge the verdict is REPRICED, never re-timed (Flash: proven); and when a rate card has changed or aged past 24h the verdict degrades to STALE-CARD naming the changed digest instead of silently re-grading. Where nothing beats running now, the verdict says RUN_NOW with saving exactly $0.00. Sold per verdict at dispatch time, per map snapshot, and as a standing fleet feed with a stated allowance. Stated plainly: pure DEFER is unreachable on today's zero-overlap two-host map (it wakes up when a third host's windows overlap); the map today is one model family deep; demand and willingness to pay are unmeasured.
Single routing verdict $0.01 (x402) · (Host × hour) map snapshot $0.99 per model · Fleet routing feed $19.99/mo (stated allowance of 5,000 verdicts + the full map feed per month; verdicts beyond the allowance metered at $0.01 per verdict; no unlimited feed). One free routing verdict per operator, ever.
Support: support@textforagents.com