Long-context jobs that used to blow our budget now run on the DeepSeek V3.2 API for a fraction of the cost.
Loading model information...
Input
DeepSeek V3.2
Hi! I'm a helpful AI assistant. What can I do for you?
DeepSeek V3.2 API — Long Context, Sparse Attention, Radically Cheaper
The DeepSeek V3.2 API runs DeepSeek V3.2, which uses sparse attention to hold long context at a fraction of the old cost — $0.028 per 1M input.
DeepSeek V3.2 rebuilds the attention layer around DeepSeek Sparse Attention, so long-context work no longer scales the way it used to. The DeepSeek V3.2 API passes that efficiency straight to you: $0.028 per 1M input tokens and $0.42 per 1M output — roughly half of the previous generation — while keeping the hybrid design that lets you think step by step or answer directly per request. Long files, long transcripts, and long tool traces stay affordable at scale.
You call the DeepSeek V3.2 API through RouterBase on the OpenAI chat-completions protocol; set the model to deepseek-v3-2 and your client is unchanged. The DeepSeek V3.2 API bills $0.028 per 1M input, $0.42 per 1M output, and $0.028 for cached input — 5% under list, on one key that reaches the whole catalog.
What the DeepSeek V3.2 API gives you
Sparse attention brings the long-context bill down.
Sparse attention
DeepSeek Sparse Attention lets the DeepSeek V3.2 API handle long context without the usual cost curve.
Radically cheaper
At $0.028 in and $0.42 out, the DeepSeek V3.2 API runs at roughly half the previous generation — 5% below list.
Long context, low cost
Long files and transcripts stay affordable, so the DeepSeek V3.2 API fits high-volume, long-input work.
Hybrid thinking kept
Think step by step or answer directly per request; the DeepSeek V3.2 API carries both modes in one model.
Coding and tools
Strong code plus reliable function calling make the DeepSeek V3.2 API a solid backbone for assistants and agents.
OpenAI-shaped
One chat-completions endpoint, one RouterBase key — point at routerbase.com/v1, name the model, skip the DeepSeek SDK.
Get going
First call to the DeepSeek V3.2 API in three steps
Wire it once, then send it long inputs.
Create a key
One RouterBase key reaches the DeepSeek V3.2 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-v3.2, and toggle thinking on or off per request.
Watch usage
The DeepSeek V3.2 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 V3.2 API
Long context that finally fits the budget.
At $0.028 per 1M input, the DeepSeek V3.2 API let us stop trimming prompts to save money.
Sparse attention is why our long transcripts finally fit the budget on the DeepSeek V3.2 API.
We kept the hybrid modes and roughly halved our bill by moving to the DeepSeek V3.2 API.
The DeepSeek V3.2 API stayed cheap even as our context windows grew, which older models never did.
Function calling held up, so our agents moved to the DeepSeek V3.2 API without a rewrite.
One string changed from our old endpoint; the cost drop on the DeepSeek V3.2 API was immediate.
We feed whole documents to the DeepSeek V3.2 API now because the long-context price finally makes sense.
Same key, cheaper long context — the DeepSeek V3.2 API was an easy call for us.
DeepSeek V3.2 API — common questions
What to know before you route traffic to it.
RouterBase's hosted access to DeepSeek V3.2, which uses sparse attention for cheaper long-context inference, over an OpenAI-compatible endpoint.