DeepSeek pricing

What this is

A reference table for DeepSeek API per-token pricing as of the V4 official release (mid-July 2026). DeepSeek introduced a peak-valley pricing mechanism: regular rates apply most of the day, peak rates (2×) apply during two UTC windows. This page is the single source of truth for cost prediction when routing tasks to DeepSeek models.

Pricing table

All prices in USD per 1M tokens.

deepseek-v4-pro

Item	Regular	Peak
Input tokens (cache hit)	$0.003625	$0.00725
Input tokens (cache miss)	$0.435	$0.87
Output tokens	$0.87	$1.74

deepseek-v4-flash

Item	Regular	Peak
Input tokens (cache hit)	$0.0028	$0.0056
Input tokens (cache miss)	$0.14	$0.28
Output tokens	$0.28	$0.56

Peak hours

Window	UTC	Ljubljana (CEST, UTC+2)	Ljubljana (CET, UTC+1)
1	01:00–04:00	03:00–06:00	02:00–05:00
2	06:00–10:00	08:00–12:00	07:00–11:00

Peak rates are 2× the regular rate for all token types.

Confirmation

DeepSeek notified (1 Jul 2026): peak-valley pricing takes effect mid-July 2026. Window 2 (08:00–12:00 CEST) overlaps with the operator’s morning work window — PR reviews, autospawned agent tasks, and cron jobs dispatched in that slot will cost 2×. No immediate code changes; revisit after MCP arc closes.

TODO (post MCP arc)

Add DEEPSEEK_PEAK_MULTIPLIER flag to Colibri cost tracking so dashboard doesn’t silently under-report AM spend
Audit cron jobs — reschedule any running 08:00–12:00 CEST to valley hours
Consider time-aware model routing (flash in peak, pro in valley) per tokenomics trifecta strategy

Where this lives in code

Hermes-BSD: hermes-bsd:agent/usage_pricing.py — _OFFICIAL_DOCS_PRICING table with entries for deepseek-v4-pro and deepseek-v4-flash. Peak rates stored in PricingEntry.peak_*_cost_per_million fields. is_deepseek_peak_hour() determines whether current time falls in a peak window. → hermes-bsd:agent/usage_pricing.py
Colibri: The colibri-deepseek probe measures actual cache-hit splits from API response headers. Cost tracking in crates/colibri-daemon/src/cost.rs uses budget modes (Fast/Smart/Max) rather than per-token dollar rates — the dollar cost flows from the agent runtime (zot) usage events. → crates/colibri-deepseek/src/lib.rs, crates/colibri-daemon/src/cost.rs

Cache-hit economics

DeepSeek’s cache-hit discount is ~120× for V4 Pro ($0.003625 vs $0.435) and ~50× for V4 Flash ($0.0028 vs $0.14). This is why Colibri’s cost model invests heavily in byte-stable prompt prefixes — a single cache-hit turn can be two orders of magnitude cheaper than a cache-miss turn.