Model positioning
DeepSeek-V3.1 was released on August 21, 2025. It is a hybrid reasoning model for dialogue, reasoning, and agent tasks, with both thinking and non-thinking modes supported by the same model. At release, deepseek-chat in the official API corresponded to non-thinking mode, while deepseek-reasoner corresponded to thinking mode. (deepseek.com)
Capabilities and uses
- Supports a 128K token context, making it suitable for long-text processing, multi-turn dialogue, and complex tasks.
- According to the official release, post-training enhanced its tool use and multi-step agent capabilities, enabling its use for coding agents, terminal tasks, and complex search tasks.
- At release, the API supported the Anthropic API format; strict function calling was available in the Beta API. (api-docs.deepseek.com)
- The model has 671B total parameters, with 37B parameters activated per token. The official release included weights for the base and post-trained models; the post-trained model card specifies the MIT license. (huggingface.co)
Limitations and version notes
deepseek-chatis an API name that changes with service upgrades and does not point permanently to V3.1: the provider upgraded it to V3.1-Terminus on September 22, 2025, and subsequently upgraded it to other versions. An official April 2026 announcement said that the old model name was scheduled to be discontinued on July 24, 2026; the current official pricing page also does not list V3.1 ordeepseek-chat. Therefore, it cannot be assumed that using this ID today will access the original V3.1; actual availability is subject to the official service. (api-docs.deepseek.com)- V3.1's tokenizer and chat template differ significantly from V3's; the weights use the UE8M0 FP8 Scale format, so framework compatibility should be checked before self-deployment. (api-docs.deepseek.com)
- Model outputs may be incorrect; code, search results, and important conclusions should be verified.
API pricing
The following are historical official V3.1 prices, not prices at which it can currently be purchased. Neither the current Chinese nor English official pricing page lists V3.1, and current official API pricing for this version has not been verified. The prices below were all checked on November 18, 2026.
- Official CNY pricing for mainland China: From 00:00 Beijing time on September 6, 2025, input per million tokens cost CNY 0.5 for cache hits and CNY 4 for cache misses; output per million tokens cost CNY 12. The nighttime discount was discontinued at the same time. Official domestic release and pricing announcement (cdn.deepseek.com)
- Official international USD pricing: From 16:00 UTC on September 5, 2025, input per million tokens cost USD 0.07 for cache hits and USD 0.56 for cache misses; output per million tokens cost USD 1.68. The off-peak discount ended at the same time. Official international release and pricing announcement (cdn.deepseek.com)
The two sets of prices apply respectively to the services described on the corresponding official pricing pages and should not be treated as the same plan with directly convertible prices. Current pricing pages: domestic, international. (api-docs.deepseek.com)