A pricing guide that admits cache hits won't save you

What it does This is a documentation-only repo: no code, just a careful guide to estimating DeepSeek API costs for deepseek-flash and deepseek-v4-pro as of October 27, 2026. It covers the three billing lines (cache-hit input, cache-miss input, output), the peak/off-peak split, and how to sanity-check a budget with low/base/high scenarios. The author explicitly notes no accounts, requests, or bills were actually tested.

The interesting bit The most useful finding is counterintuitive: in the worked example, output costs $300 while all input costs $123.60. Doubling output length adds more to the bill than losing the cache entirely — so obsessing over cache-hit rates can distract from the line that actually dominates.

Key highlights

• Full rate table per 1M tokens, with peak rates exactly double off-peak across every category
• Peak windows defined in UTC (01:00–04:00 and 06:00–10:00, weekdays, excluding Chinese public holidays)
• A shortcut multiplier (1 + p) for when all token categories share the same peak share
• Emphasizes measuring prompt_cache_hit_tokens and prompt_cache_miss_tokens rather than labeling whole requests as hits
• Warns that caching is best-effort and that aggregator billing paths need separate verification

Caveats

• Prices are a snapshot; the README itself says to recheck DeepSeek’s pricing page before approving a budget
• No empirical validation — all numbers are hypothetical arithmetic, and no model quality or latency comparison is offered
• The date (October 2026) and model names should be treated as the repo’s claims, not independently confirmed here

Verdict Worth a read if you’re budgeting DeepSeek usage and want a framework instead of a single blended rate. Skip it if you want code, benchmarks, or anything you can run — this is an article wearing a repo’s clothes.