DeepSeek

What DeepSeek actually meters

DeepSeek meters tokens, and it's the only vendor in this index that prices the outcome of a cache lookup rather than the work of building the cache. OpenAI and Anthropic both charge a premium to write a cache entry and a discount to read one. DeepSeek charges nothing to write and applies one of two input rates depending on whether your prefix matched.

Every call returns prompt_cache_hit_tokens and prompt_cache_miss_tokens in its usage block, so the split is auditable per request rather than inferred from the invoice.

The gap is severe enough to dominate a bill. On deepseek-flash a cache hit costs 2% of a cache miss. On deepseek-v4-pro it costs roughly 3.3%. Output is the milder multiple: 4x cache-miss input on Flash and 3x on Pro, a narrower spread than the 5x most Western providers run.

Caching itself is automatic, enabled for all users by default with no parameter to toggle. Hits require a full match against a persisted prefix unit, and DeepSeek persists those at request boundaries, at common prefixes detected across requests, and at fixed token intervals inside long inputs. The docs are explicit that this is best effort with no guaranteed hit rate, and that an unused cache clears within hours to days.

How DeepSeek's balance works

DeepSeek runs prepaid, with two balances per account. A granted balance covers promotional credit and a topped-up balance covers what you've paid in. Charges draw from the granted balance first when both exist, which is the deduction order you'd want.

Granted credit expires. The balance endpoint returns granted_balance as "the total not expired granted balance", which confirms expiry exists, but the docs never publish the window or what triggers a grant. That's the one real gap in an otherwise complete rate card.

There are no credit packs, rollover rules or pooling to document, because DeepSeek doesn't sell credits as a product. The balance is a cash float against a meter. A user_id parameter isolates content safety, cache and scheduling per end user, but it doesn't split the balance.

What happens when you hit the limit

Nothing throttles and nothing auto-charges. At zero balance the API returns HTTP 402 Insufficient Balance and you top up to resume, which is a hard stop rather than an invoice surprise.

The separate ceiling is concurrency: 2,500 in-flight requests on Flash and 500 on Pro, counted per account across all API keys. Exceeding it returns HTTP 429. DeepSeek grants expansion on request and states plainly that it costs nothing extra, which is unusual. Most providers attach higher concurrency to a spend tier.