DeepSeek has launched its latest model, DeepSeek V3.2. The experimental model builds upon V3.1 Terminus and introduces DeepSeek Sparse Attention (DSA). With it comes a 50% drop in the official API pricing.
The open-source model is available for download on HuggingFace, and DeepSeek has updated its official API to use the new V3.2 model, replacing the previous V3.1 Terminus model. You can also try the model on DeepSeek’s chat interface.
DeepSeek V3.2 – An Experimental Model
DeepSeek V3.2 builds upon V3.1 Terminus, a model released a week ago that addressed issues reported by users in V3.1 and improved the model’s language consistency and agent capabilities.
V3.2 is an experimental model that introduces DeepSeek Sparse Attention (DSA). DSA significantly improves long-context training and inference efficiency with “minimal impact on output quality,” according to DeepSeek’s official announcement. It’s a step ahead towards the company’s next-gen architecture.
DeepSeek V3.2 And AI Roleplay
Please Read: DeepSeek V3.2’s Performance In AI Roleplay
New API Pricing Effective 29th September
The launch of DeepSeek V3.2 also brings an update to DeepSeek’s official API pricing. DeepSeek has slashed prices by 50% effective immediately.
The new pricing is $0.028 per 1 million input tokens that hit their input tokens cache (repeated tokens), $0.28 per 1 million tokens that miss their input tokens cache (new tokens), and $0.42 per 1 million output tokens.
What Is DeepSeek Input Tokens Cache?
DeepSeek’s Input Tokens Cache is a feature available through the first-party API that reduces the cost of processing duplicate Input Tokens, such as repeated instructions and chat history.
Learn More: DeepSeek’s Input Tokens Cache And AI Roleplay
Cost Of Using DeepSeek Official API
In September 2025, we used 798,963 tokens with V3.1 non-thinking mode and 1,440,232 Tokens with V3.1 thinking mode and sent a total of 398 messages to our AI characters. It cost us $0.61. With the price dropping by 50%, if you had 10 times our usage, your total wouldn’t exceed $3.50.
DeepSeek’s Pricing Compared To Others
| Model | Input Tokens (Per Million) | Output Tokens (Per Million) | |
| Cache Hit | Cache Miss | ||
| deepseek-chat deepseek-reasoner | $0.028 | $0.28 | $0.42 |
| GPT-5 | $0.125 | $1.25 | $10 |
| GPT-5 mini | $0.025 | $0.25 | $2 |
| Claude Opus 4.1 | $1.50 | $15 | $75 |
| Claude Sonnet 4 | $0.30 | $3 | $3 |
| Grok 4 | $0.75 | $3.00 | $15 |
| Grok 3 | $0.75 | $3.00 | $15 |
| Grok 3 mini | $0.075 | $0.30 | $0.50 |
| Gemini 2.5 Pro | $0.31 | $1.25 | $10 |
| Gemini 2.5 Flash | $0.075 | $0.30 | $2.50 |
| Mistral Large | $2 | $2 | $6 |
| Mistral Medium 3 | $0.4 | $0.4 | $2 |
| Mistral Small 3.2 | $0.1 | $0.1 | $0.3 |
DeepSeek V3.2 – Testing The Waters
DeepSeek’s latest model, V3.2, is an experimental step forward toward the company’s next-generation architecture. The introduction of DSA significantly improves the efficiency of long-context training and inference while maintaining output quality.
Since DeepSeek V3.2 builds upon V3.1 Terminus, its performance in AI roleplay remains similar to V3.1 Terminus. Along with the new model, DeepSeek also reduced the official API pricing by 50%. It continues to be one of the most affordable options for AI roleplay compared to other similarly capable models.
