LLMHOSTING.AI the compute price terminal

DeepSeek V3.2 API pricing

9 providers serve DeepSeek V3.2. DeepSeek V3.2 is an open-weight model (MIT) with 685B total parameters, 37B active per token (mixture-of-experts), up to 164K context. Prices below are per 1M tokens, cheapest input first (self-hosting math further down).

Refreshed 2026-08-23 · list prices from provider APIs
Input floor $0.269/1M at Novita AI Output floor $0.400/1M at Novita AI, DeepSeek, GMI Cloud and OpenRouter Find your workload winner →
Provider Input $/1M Output $/1M Cache read $/1M Context
Novita AI $0.269 $0.400output floor $0.135 164K Use
DeepSeek $0.280 $0.400 - 164K Use
GMI Cloud $0.280 $0.400 - 164K Use
OpenRouter $0.280 $0.400 - 164K Use
Fireworks AI $0.560 $1.68 - 164K Use
Google Vertex AI $0.560 $1.68 - 164K Use
Azure AI Foundry $0.580 $1.68 - 164K Use
AWS Bedrock $0.620 $1.85 - 164K Use
SambaNova $3.00 $4.50 - 33K Use

The cheapest way to run DeepSeek V3.2

On an illustrative 70M-input / 30M-output monthly workload, today's lowest listed cost is roughly $30.83/month, at Novita AI. Input-heavy, output-heavy and cache-heavy workloads can produce different winners. Use the live mix calculator below rather than combining floors from two different providers. On output-token cost alone, the API floor remains below our self-hosting estimate of ~$1.27/1M. Details below. Run your own numbers in the breakeven calculator, or read the breakeven math.

Your actual monthly cost

Uses today's list prices

Cache read uses the listed cache rate where available; otherwise it falls back to normal input price. Batch, write-cache, volume and negotiated discounts are excluded.

Or run it yourself: DeepSeek V3.2 self-hosting economics

DeepSeek V3.2 is open-weight (MIT), so the API price above competes with the GPU-hour market. 685B total parameters (MoE, ~37B active per token) needs roughly 942 GB VRAM at FP8 or 471 GB at INT4, KV-cache headroom included.

No single GPU fits DeepSeek V3.2 at FP8 (942 GB needed), so it needs a multi-GPU node. Cheapest tracked option: 5× MI300X (192 GB each) at roughly $2.50/hr total (RunPod). Rent 5× MI300X at RunPod → INT4 quantization drops the requirement to 471 GB. Tune it in the calculator.
Breakeven estimate: a well-batched vLLM deployment on 5× MI300X ($2.50/hr) at ~1094 aggregate tok/s and 50% utilization works out to roughly $1.27 per 1M output tokens, versus $0.400 via the cheapest output-token API (Novita AI, DeepSeek, GMI Cloud and OpenRouter). The API is cheaper unless your utilization or throughput beats these assumptions. Planning estimate: throughput varies with hardware, quantization, batch size and context. Tune it in the calculator.

Related models

ModelCheapest in $/1MCheapest out $/1MProviders
DeepSeek R1 $0.200 $0.400 17
DeepSeek V3 $0.200 $0.200 16
DeepSeek R1 Distill Llama 70B $0.200 $0.375 10
DeepSeek-V3.1 $0.270 $1.00 9
DeepSeek V4 Flash $0.050 $0.101 9
DeepSeek V4 Pro $0.435 $0.870 6

FAQ

What is the cheapest API for DeepSeek V3.2?

As of 2026-08-23, the lowest input price for DeepSeek V3.2 is Novita AI at $0.269 per 1M input tokens. The lowest output price, $0.400 per 1M tokens, is likewise shared by Novita AI, DeepSeek, GMI Cloud and OpenRouter. The cheapest provider for a real workload depends on its input/output mix. The next-lowest input price is DeepSeek at $0.280, a 4% difference.

How much VRAM do you need to self-host DeepSeek V3.2?

DeepSeek V3.2 has 685B parameters, so plan for roughly 942 GB of VRAM at FP8 or 471 GB at INT4/AWQ, KV-cache headroom included. That exceeds a single GPU. A typical node is 5× MI300X.

← All models · GPU rental prices · Breakeven calculator · Get this data as JSON