DeepSeek

Route your DeepSeek API calls through AI SpendOps for automatic usage tracking and cost attribution.

Configuration

Setting	Value
Route	/v1/deepseek/*
Upstream	https://api.deepseek.com
Auth header	Authorization: Bearer ...
Streaming usage	Auto-injected (stream_options.include_usage)

SDK base URL

https://proxy.aispendops.com/v1/deepseek/v1

Example

curl https://proxy.aispendops.com/v1/deepseek/v1/chat/completions \
  -H "Authorization: Bearer your-deepseek-key" \
  -H "X-ASO-API-Key: aso_k_yourkey.secret" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-chat","messages":[{"role":"user","content":"Hello"}]}'

Usage fields

Field	Description
prompt_tokens	Input tokens, including any cache hits
completion_tokens	Output tokens
prompt_cache_hit_tokens	Portion of prompt_tokens served from cache
prompt_cache_miss_tokens	Portion of prompt_tokens that was not cached

Cache token accounting

DeepSeek is the only supported provider that reports caching as a hit and miss pair. The two values sum to prompt_tokens, so a cache hit is a breakdown inside the prompt total rather than a separate category alongside it.

AI SpendOps maps prompt_cache_hit_tokens onto its own cache_read_tokens field and subtracts it from prompt_tokens before pricing, so the cached portion is charged once at the cache-read rate rather than at the full input rate. prompt_cache_miss_tokens needs no special handling, since it is simply the remainder.

That puts DeepSeek in the same family as OpenAI and Google Gemini, which also report cached tokens inclusively, and makes it the opposite of Anthropic, where cache tokens are separate non-overlapping categories and nothing is subtracted.

If you are reconciling a DeepSeek invoice by hand, note that adding prompt_tokens and prompt_cache_hit_tokens together double counts the cached input.

What AI SpendOps records

Field	Populated
cache_read_tokens	Yes, from prompt_cache_hit_tokens
cache_write_tokens	No. DeepSeek caching is automatic with no separate write charge
reasoning_tokens	Yes for reasoning models, via completion_tokens_details
provider_cost	No. Cost is calculated from token counts and the current DeepSeek rate card

Notes

• DeepSeek caching is automatic. There is no cache control parameter to set and no separate cache write cost, unlike Anthropic.
• The proxy injects stream_options: { include_usage: true } on streaming requests, so streamed calls report real token counts rather than a character-based estimate. See Streaming.
• Full field definitions are in the usage event field reference.