The DeepSeek V3.1 API let us drop two separate models — thinking for the hard tickets, direct for the rest, all behind one call.
Loading model information...
Input
DeepSeek V3.1
Hi! I'm a helpful AI assistant. What can I do for you?
DeepSeek V3.1 API — One Model, Thinking and Direct
The DeepSeek V3.1 API runs DeepSeek V3.1, a hybrid model that can reason step by step or answer directly — you pick the mode per request.
DeepSeek V3.1 folds two behaviors into one model: a thinking mode that reasons step by step, and a non-thinking mode that answers directly. The DeepSeek V3.1 API lets you switch between them per request, so hard problems get deliberation and simple ones get a fast, cheap reply — no second model, no second endpoint. It keeps V3 strengths in coding and tool use, with a 128K context window for long files and transcripts.
You call the DeepSeek V3.1 API through RouterBase on the OpenAI chat-completions protocol; set the model to deepseek-v3-1 and your existing client works unchanged. The DeepSeek V3.1 API bills $0.27 per 1M input tokens, $1.10 per 1M output, and $0.07 for cached input — 5% under list, on one key that reaches the whole catalog.
What the DeepSeek V3.1 API gives you
Deliberation when it helps, a direct answer when it does not.
Two modes, one model
The DeepSeek V3.1 API switches between thinking and direct answers per request, so you match effort to the task without swapping models.
Reason when it helps
Turn thinking on and the DeepSeek V3.1 API works through math, code, and multi-step problems before it answers.
Answer when it does not
In non-thinking mode the DeepSeek V3.1 API replies straight, keeping latency and cost low for routine calls.
Coding and tools
Strong code plus reliable function calling make the DeepSeek V3.1 API a solid backbone for IDE assistants and agents.
128K context
A 128K window lets the DeepSeek V3.1 API hold long files, transcripts, and tool traces in a single call.
Cheap, with caching
At $0.27 in and $1.10 out — plus $0.07 cached input — the DeepSeek V3.1 API stays affordable at volume, 5% below list.
Get going
First call to the DeepSeek V3.1 API in three steps
Wire it once, then flip modes as you like.
Create a key
One RouterBase key reaches the DeepSeek V3.1 API and every other model in the catalog.
Point and pick a mode
Send any OpenAI client to routerbase.com/v1 with model=deepseek/deepseek-v3.1, and toggle thinking on or off per request.
Watch usage
The DeepSeek V3.1 API returns token usage per response, so cost and latency stay visible on every call.
Pay only for what you use
RouterBase passes through partner-tier pricing. Compared against the model's official published API rate.
Teams building on the DeepSeek V3.1 API
Reasoning and direct answers behind a single call.
Toggling thinking off for lookups cut our latency and bill without touching the hard-path quality.
One model with two modes meant the DeepSeek V3.1 API replaced a router we used to maintain by hand.
At $0.27 in with cached reads at $0.07, the DeepSeek V3.1 API kept our costs flat as traffic grew.
Function calling is reliable, so our agent stack runs entirely on the DeepSeek V3.1 API.
We turn thinking on for evals and off in production — same model, same key.
The 128K window let the DeepSeek V3.1 API read our whole repo file in one pass.
Switching from our old endpoint to the DeepSeek V3.1 API was one string; the hybrid mode was a bonus.
We route hard reasoning and quick chat to the same DeepSeek V3.1 API and just flip the mode.
DeepSeek V3.1 API — common questions
What to know before you route traffic to it.
RouterBase's hosted access to DeepSeek V3.1, a hybrid model that can reason step by step or answer directly, over an OpenAI-compatible endpoint.