DeepSeek

Direct access versus this gateway

Direct DeepSeek access uses https://api.deepseek.com and a DeepSeek API key. This gateway uses its own base URL and Bearer token. This page currently declares only /v1/chat/completions; do not treat DeepSeek's official Responses, Anthropic-compatible, or beta paths as gateway endpoints.

Current models and thinking

As of this review, the official IDs are deepseek-v4-flash and deepseek-v4-pro, both with thinking enabled by default. This replaces the old mental model that chat always means non-reasoning and reasoner always means reasoning. Direct calls can switch thinking.type and tune reasoning_effort; gateway calls must use a name actually enabled in the gateway model list because channels can map model IDs.

Streaming, JSON, and tools

Official Chat Completions streams SSE and terminates with [DONE]; stream_options.include_usage requests aggregate usage. JSON mode also needs an explicit JSON instruction or it may emit whitespace until the token limit. A tool call only proposes a function and arguments: the application must validate and execute it, never passing model output directly into payments, destructive database actions, or shell commands.

Errors and cost safety

401 usually indicates a bad key, 402 insufficient balance, 429 a rate limit, and 500/503 a potentially retryable service failure. Thinking can increase output tokens and latency. Set max_tokens, request timeouts, concurrency caps, and balance alerts, and read live pricing rather than hard-coding rates.

Production checklist

Verify the exact model ID, thinking mode, context, output cap, tools, and JSON support. Record request IDs, returned model, usage, and review date. Models and prices can change, so reproducible workloads should retain full requests and regression samples.