Cost for DeepSeek is under-reported during peak hours (price table only stores the off-peak rate)

Summary

Cost is computed from a static price table (Model.cost.input/output/cache_read/cache_write, sourced from models.dev). DeepSeek bills with peak / off-peak rates (off-peak is half of peak), but the catalog only carries the off-peak numbers. As a result the displayed cost is under-reported ~2x during DeepSeek peak hours.

Data

models.dev deepseek/deepseek-flash currently has:

cost: { input: 0.15, output: 0.6, reasoning: 0.6, cache_read: 0.003 }

DeepSeek official pricing (https://api-docs.deepseek.com/quick_start/pricing):

Peak = UTC Mon–Fri 01:00–04:00 and 06:00–10:00 (= Beijing Mon–Fri 09:00–12:00 / 14:00–18:00); all other hours are off-peak. The table above picks the off-peak column.

Impact (measured locally, opencode 1.18.30)

793 assistant messages with providerID = deepseek:

• displayed total: $0.7590
• time-corrected total: $1.0680
• under-reported: $0.3091 (~41%); 299 / 793 messages fell in peak.

Why it can't be worked around locally

There is no time dimension in the cost schema — Model.cost has input/output/cache_read/cache_write plus context-size tiers (context_over_200k), but no peak/off-peak windows. Searching the shipped binary finds 0 occurrences of off-peak/offpeak. A static override can only ever pick one of the two rates (under-report at peak, or over-report ~2x off-peak).

Proposal

• models.dev: carry both rates for models that have them (e.g. cost.off_peak or a small schedule).
• opencode: when computing cost, select the rate from the request/completion time (UTC, weekday-aware).

If models.dev intentionally standardizes on the lowest (off-peak) rate, please at least document that, so the value is understood to be a lower bound rather than the amount actually charged.

Environment

• opencode 1.18.30, macOS 26.6.2
• model deepseek/deepseek-flash (DeepSeek-V4.1-Flash)