DeepSeek V3.2 listed at $0.28 per million input tokens and $0.42 per million output tokens on the official DeepSeek API until that access was retired in July 2026. Across third-party hosts, the same model now prices from $0.21 to $0.56 per million input tokens among mainstream providers, with two outliers at $3.00, and the cheapest list price is also, for the first time, the cheapest option once cache discounts enter the math.
Update, September 2026: DeepSeek's official API no longer serves V3.2. The deepseek-chat and deepseek-reasoner names were pointed at V4 Flash when V4 launched in April and retired on July 24, 2026, so the official rates below are historical. Third-party hosts still serve V3.2 weights, and that market is now the only one. Current pricing for every DeepSeek model and host is on DeepSeek API pricing, and the current flagship is on DeepSeek V4 Pro API pricing.
This page is the pricing reference: official rates, the full provider comparison, and how to compute what your actual workload will pay. For why the same model prices this differently at all, see Why Token Prices Differ.
What DeepSeek V3.2 is, briefly
V3.2 replaced both DeepSeek V3 and R1 with a unified model that handles chat and reasoning at a single price point, with a 163K context window. It is one of the most widely hosted open-weight models in production, which is exactly why its pricing is worth a dedicated page: open weights mean many competing sellers, and many sellers mean a price spread worth shopping.
DeepSeek V3.2 official API pricing (retired July 24, 2026)
| Token type | Price per 1M tokens |
|---|---|
| Input | $0.28 |
| Input, cache hit | $0.028 |
| Output | $0.42 |
These are the last official rates before DeepSeek retired V3.2 from its API on July 24, 2026. Baidu Qianfan currently serves V3.2 at exactly these rates, cache line included.
The number that used to separate the official API from most hosts was the cache line: a 90 percent discount on cached input tokens. For workloads with heavy prompt reuse, this single term changes the provider ranking, as shown below.
DeepSeek V3.2 pricing by provider
Published rate cards across major hosts, from OpenRouter listings on September 2, 2026:
| Provider | Input $/1M | Output $/1M | Cached input $/1M |
|---|---|---|---|
| GMI Cloud | $0.2088 | $0.3096 | $0.0216 |
| StreamLake | $0.2145 | $0.3218 | $0.0215 |
| DigitalOcean | $0.25 | $0.80 | $0.075 |
| SiliconFlow | $0.259 | $0.42 | $0.135 |
| AtlasCloud | $0.26 | $0.38 | $0.13 |
| DeepInfra | $0.26 | $0.38 | $0.13 |
| Novita | $0.269 | $0.40 | $0.1345 |
| Baidu Qianfan (old official rates) | $0.28 | $0.42 | $0.028 |
| Venice | $0.33 | $0.48 | $0.16 |
| Alibaba Cloud | $0.3705 | $1.112 | $0.0741 |
| Friendli | $0.50 | $1.50 | $0.25 |
| Google Vertex | $0.56 | $1.68 | n/a |
| Phala | $1.00 | $1.00 | $0.50 |
| SambaNova | $3.00 | $4.50 | n/a |
| MARA | $3.00 | $4.50 | n/a |
The mainstream range is $0.21 to $0.56 input, a 2.7x spread; SambaNova and MARA at $3.00 are premium or low-uptime outliers, not the market.
Notes that matter as much as the prices: the SiliconFlow budget endpoint serves an FP8-quantized version, not full precision. Alibaba and ByteDance-hosted endpoints cap context at 131K rather than the full 163K. A cheaper endpoint that serves a different precision or a smaller context is not the same product at a lower price.
What a real workload actually pays
List prices mislead because workloads are a mix of input, output, and cached tokens. Take a production assistant doing 10M input and 1M output tokens per day. The official column is now a reference point rather than an option.
| Provider | Daily | Monthly | Annual |
|---|---|---|---|
| GMI (cheapest list) | $2.40 | $71.93 | $875 |
| StreamLake | $2.47 | $74.02 | $900 |
| Baidu Qianfan (old official rates) | $3.22 | $96.60 | $1,175 |
| Google Vertex | $7.28 | $218.40 | $2,657 |
On raw list price, the cheapest host saves about 25 percent against the old official rate. Now give the same workload a 60 percent cache hit rate on input, common for agents and RAG systems with stable system prompts:
| Provider | Daily cost with 60% cached input |
|---|---|
| GMI Cloud ($0.0216 cached) | $1.27 |
| StreamLake ($0.0215 cached) | $1.31 |
| Baidu Qianfan ($0.028 cached) | $1.71 |
| DeepInfra ($0.13 cached) | $2.20 |
| Novita ($0.1345 cached) | $2.28 |
A month ago the ranking flipped here: the cheapest list host had no cache discount, and the official API's $0.028 cached rate made it 26 percent cheaper on this workload. That is no longer the case. GMI and StreamLake now carry cache-read rates near the old official line, so the cheapest list price is also the cheapest blended price. The lesson survives, though, in the middle of the table: DeepInfra and Novita match GMI on list within 30 percent but cost 75 percent more once caching is in the mix, because their cached rate is 6x higher. Comparing input list prices when cache terms dominate the bill is still the single most common mistake in provider selection.
The general rule: compute your blended price, the weighted cost across your real input, output, and cache mix. This is the logic behind the Standard Blended Price on the Mercatus Token Index, which tracks per-model blended pricing across providers daily.
How to choose a V3.2 provider
- Heavy prompt reuse (agents, RAG, chat with long system prompts): GMI Cloud or StreamLake, which pair the lowest list price with a cache-read rate around $0.02. Baidu Qianfan is the fallback at the old official terms.
- One-shot workloads with little repeated context (batch processing, document extraction): cheapest list price wins, and it is the same two hosts at $0.21 input.
- Quality-sensitive workloads: confirm full-precision serving; skip FP8 endpoints or benchmark them against your evals first.
- Long-context workloads: confirm the host serves the full 163K window; several cap at 131K.
- Latency-sensitive: the fastest hosts charge a premium; whether it's worth it is a throughput question, not a price question.
Where V3.2 pricing goes from here
Open-weight model pricing has mostly moved one direction: down, as hosting competition compounds, though DeepSeek's August 2026 V4 increase showed the developer can move the other way.
But the spread between hosts persists, because token prices are posted rather than traded; no venue forces the gaps closed. That structure, and what changes it, is covered in How a Token Exchange Works. Until then, the spread is the buyer's problem and the shopper's opportunity: on this model it's a 2.7x spread on list price and an 80 percent difference on a cached workload between the best and worst mainstream hosts.
Serving costs also set a floor. Every V3.2 token price traces back to GPU-hour economics, the cost-per-token chain from hardware cost through throughput to price. Hosts running paid-off or cheap hardware can go lower than hosts on rented hyperscaler capacity, which is a large part of why the table above looks the way it does.
Frequently asked questions
How much does DeepSeek V3.2 cost per million tokens?
On third-party hosts, $0.21 to $0.56 per million input tokens and $0.31 to $1.68 output among mainstream providers, as of September 2026. The official API listed $0.28 / $0.42 with cached input at $0.028 until DeepSeek retired V3.2 access on July 24, 2026.
Who is the cheapest DeepSeek V3.2 provider?
GMI Cloud at $0.2088 input and $0.3096 output, with cached input at $0.0216, followed closely by StreamLake. Because both now offer strong cache terms, they are the cheapest on list price and on cached workloads alike.
Is DeepSeek V3.2 still available?
Not from DeepSeek. The official API retired the legacy model names on July 24, 2026 and lists only V4 models. Fifteen third-party hosts serve V3.2 as of September 2026; confirm precision and context window, since some serve FP8 or cap at 131K.
Is the cheapest V3.2 endpoint the same model?
Not always. Some budget endpoints serve FP8-quantized versions, and some cap context at 131K instead of 163K. Confirm precision and context before comparing prices.
Why does V3.2 pricing vary so much between providers?
Different hardware costs, utilization, quantization choices, and margin strategies, with no market mechanism forcing convergence. The full breakdown is in Why Token Prices Differ.
What is a blended price?
Your weighted cost across input, output, and cached tokens in the proportions your workload actually uses. It is the only number that predicts your bill, and the basis of the Token Index's Standard Blended Price.
Methodology
Provider pricing from OpenRouter's DeepSeek V3.2 listings on September 2, 2026, including promotional discounts in effect that day. Official V3.2 access retired July 24, 2026. Workload examples use stated token volumes and cache assumptions; recompute with your own mix before choosing a provider. Blended pricing methodology follows the Mercatus Token Index Standard Blended Price. Last verified: 2026-09-02.