DeepSeek

deepseek-chat selects DeepSeek's general-purpose V3 chat line; deepseek-reasoner selects the R1-style reasoning line that thinks before answering. Both are moving aliases pointing at whatever generation is current, not frozen model versions.

deepseek-reasoner puts its chain of thought in message.reasoning_content and the user-facing answer in message.content. Both come back in one ordinary chat-completion response. Show users content; treat reasoning_content as diagnostic text you log, collapse, or discard.

DeepSeek's JSON output mode constrains the model to emit syntactically valid JSON, nothing more. It does not validate against a schema you supply, so field names, types and required keys remain your responsibility to prompt for and then verify.

DeepSeek splits a request's input tokens into two billed buckets: prompt_cache_hit_tokens, served from its context cache at a steeply discounted input rate, and prompt_cache_miss_tokens, billed at the full input rate. The two always sum to prompt_tokens.

DeepSeek rejects a request whose messages array contains reasoning_content, returning a 400. Append only the assistant's content to the conversation. The model re-derives fresh reasoning each turn from the visible messages, so nothing is lost by dropping the previous trace.

Chat completions port cleanly. Everything outside them does not: DeepSeek exposes no embeddings, image, audio or assistants endpoints. Schema-strict JSON output is unavailable, some parameters are accepted but ignored, and billing errors surface as an unfamiliar HTTP status.

Nothing. DeepSeek's context caching is on by default for every account on both deepseek-chat and deepseek-reasoner. There is no cache parameter, marker or header to send, and DeepSeek charges nothing to store the cached content.

It permits commercial use, modification, redistribution and derivative models with no user or revenue threshold and no naming requirement — you only keep the copyright and licence notice. It covers the weights, not your use of the hosted API.

No. The R1 distills are third-party base models — Qwen or Llama checkpoints of 1.5B to 70B — fine-tuned on reasoning traces produced by DeepSeek-R1. They are dense models with a different lineage, not shrunken R1 weights.

DeepSeek documents temperature, top_p, presence_penalty and frequency_penalty as unsupported on deepseek-reasoner: the request succeeds but the values do nothing. logprobs and top_logprobs are rejected with an error instead of being ignored. Verify against the docs for your model release.

Fill-in-the-middle completion and assistant-prefix continuation are opt-in beta features on DeepSeek. Point the client at https://api.deepseek.com/beta to use them; on the standard base URL the requests are rejected or the extra fields ignored.

Whenever the task has no deliberation to do: short classification, extraction, reformatting, routing, and any path with a tight latency budget or many small agent turns. Deliberation is generated text, so it costs time and output tokens on every call.

DeepSeek does not enforce a per-key request or token limit. Under heavy traffic it queues your request and holds the HTTP connection open, sending filler — blank lines when not streaming, SSE comment lines when streaming — so client timeouts, not 429s, are the usual failure mode.

Chunks first carry delta.reasoning_content while the model thinks, then switch to delta.content for the answer. Accumulate the two into separate buffers and treat the first content delta as the phase transition; usage totals only arrive at the end of the stream.

The chain of thought is generated output. It consumes the same token budget as the answer and is counted in completion_tokens, so a long think can exhaust max_tokens before the answer finishes. Check finish_reason for length and size the budget for thinking plus answer.

Treat the alias as a floating dependency: hold a regression evaluation suite, run it continuously against production-shaped prompts, and alert on drift. Where behaviour must actually be frozen, serve a pinned revision of the open weights yourself — the only real version lock available.

Model each call as a three-term sum: cache-hit input tokens, cache-miss input tokens and completion tokens, each at its own rate. Caching discounts only the first term, so an output-heavy or reasoner workload can show a 90% hit rate and barely move the bill.

Partly. Wire compatibility genuinely portes transport, streaming and error plumbing, which is most of the integration code. It does not port behaviour, capabilities or cost profile, so a base_url swap is a good starting point and a poor abstraction boundary.