DeepSeek API Documentation
The getting-started reference for deepseek‘s model API — the seed source of this wiki
(routed here from the hub _inbox). Documents the provider’s API surface.
What it covers
- Models:
deepseek-v4-flashanddeepseek-v4-pro(current); deprecateddeepseek-chat/deepseek-reasonerretire 2026-07-24. - Compatibility: OpenAI- and Anthropic-SDK-compatible endpoints (
https://api.deepseek.com); no-code integrations with Claude Code and GitHub Copilot. - Features: configurable thinking mode (reasoning effort), context caching, streaming.
Re-verified 2026-08-04 (staleness queue) — prices now recorded
The June read documented the API surface and no rates, which left this page unable to support the “budget end of llm-api-pricing” claim it makes. Pricing page read directly, per million tokens:
| Model | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
deepseek-v4-flash | $0.0028 | $0.14 | $0.28 |
deepseek-v4-pro | $0.003625 | $0.435 | $0.87 |
Two things worth carrying. The cache-hit rate is 50× cheaper than cache-miss on flash and ~120×
on pro, which makes context caching a pricing lever rather than a convenience — the gap is far wider
than the headline number a comparison table would quote. And the model lineup is unchanged from
June: the two v4 models stand, and the deepseek-chat / deepseek-reasoner retirement date of
2026-07-24 has now passed, so those endpoints should be treated as gone.
Announced, not yet in effect: a peak/off-peak structure where “during peak hours, prices will be 2x the regular prices”, peak being 09:00–12:00 and 14:00–18:00 Beijing time. No effective date given. Recorded because a 2× time-of-day multiplier would be a genuinely unusual move in this market and worth watching, not because it is live.
Why it’s here
A primary-source datapoint on deepseek as a low-cost, OpenAI-compatible provider — the budget end of llm-api-pricing. Its OpenAI/Anthropic-compatible surface is itself a market signal: challengers conform to incumbents’ API shapes to lower switching costs. (Reference docs; specifics are snapshots and will drift — note the deprecation date.)