DeepSeek's API Is Two Endpoints and Two Model Names — Here Is What Each One Accepts

The DeepSeek API looks simpler than it is. There are two base URLs, two model identifiers, and a feature matrix where the two columns are almost identical — except for the one line that decides whether your workload is possible at all. If you are migrating from an OpenAI SDK you need one of them; if you are migrating from an Anthropic-shaped harness you need the other; and if you are sending screenshots, only one of the two models will take them.

The two bases

• OpenAI format — code>https://api.deepseek.com/code>. Chat completions in the OpenAI shape, including the Responses API, JSON output, tool calls and prefix completion in beta.

• Anthropic format — code>https://api.deepseek.com/anthropic/code>. The same two models behind an Anthropic-shaped surface, which is what a coding harness that speaks that protocol needs.

Both bases serve both models. The choice of base is about your client library, not about capability.

The two names, and what is behind each

• deepseek-flash — currently served by DeepSeek-V4.1-Flash-1017, the October checkpoint. Multimodal: text and images in, text out. 1M context, 384K maximum output, 2,500 concurrent requests.

• deepseek-v4-pro — served by DeepSeek-V4-Pro-0813. Text only. Same 1M context and 384K output, 500 concurrent requests.

Two legacy identifiers are still accepted — code>deepseek-v4-flash/code> and code>deepseek-v4-flash-vision-exp/code> — but DeepSeek's documentation is explicit that the corresponding models have been retired, that requests are served by V4.1-Flash-1017, and that they are billed at the Flash price. In practice that means an old integration keeps working and quietly starts returning a different model's answers, which is worth checking before you attribute a change in output to anything else.

The feature matrix, row by row

Almost everything is shared. The exceptions are the whole reason this guide exists.

• Thinking mode — both models support non-thinking and thinking, with thinking as the default. Effort levels are low, high and max; DeepSeek's guidance is low for simple tasks, high for daily agent work, max for harder scenarios.

• JSON output, tool calls, prefix completion — supported identically on both. Prefix completion is in beta.

• Responses API and Anthropic API — both supported on both models, which is unusual and worth knowing: you can drive either model from either client shape without a translation layer.

• FIM completion — beta on both, and non-thinking mode only on both. If your use case is code infill, that restriction applies to the flagship too.

• Vision — a check mark for code>deepseek-flash/code>, "Not supported" for code>deepseek-v4-pro/code>. This is the single fork in the matrix, and it is not something a prompt change can work around.

What the identifiers imply about cost

The names carry the price structure. Flash is $0.15 in and $0.60 out per million tokens, with a cache hit at $0.003; V4 Pro is $0.66 and $1.98, cache hit $0.022. Both double during 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. So the model name you pick sets both the capability and the bill, and the legacy names above set the bill at the Flash rate regardless of which checkpoint answers.

One practical consequence: because code>deepseek-flash/code> now resolves to a checkpoint released in October, a cost model or a quality baseline written against the September model is describing something you no longer receive under that identifier. Pin nothing to a name you have not re-measured this month.

How this compares to a routing layer

DeepSeek's own surface is two models behind two protocols. If your application only ever needs the strongest DeepSeek model, that is enough. The cases where it is not enough are the ones this shape cannot express: choosing a fallback for a provider outage, composing DeepSeek with a different vendor's model in one workflow, or A/B-testing Flash against Pro on live traffic without two integrations.

That is what OrcaRouter's 224-model, OpenAI-compatible endpoint is for, and DeepSeek V4 Pro and V4.1 Flash both sit on it at DeepSeek's published rates with no markup — including the cache-hit row and the peak doubling, because the vendor's price is what passes through. The endpoints above still work exactly as DeepSeek documents them; a router does not replace the vendor's API, it gives you one place to switch between several of them.

The migration checklist, in order

• Pick the base by your client shape, not by the model you want.

• Pick the model by what you send: images mean Flash, text-only work means the price decides.

• Set thinking effort explicitly rather than inheriting the default, because the default is thinking-on.

• Check whether your old identifier has been retired under you — the two legacy names are accepted, rebilled and re-served.

• Measure the cache-hit share before you estimate the monthly bill, since a stable prefix is a 30× discount and an unstable one is nothing.