Secrets, endpoints, design IDs, prompt caching and you can password instances — as well as exactly what changed when the old speak and you can reasoner endpoints had been resigned. Check in at the chat.deepseek.com — the state chat features a totally free level and works the current V4 designs. The new combined enter in rates brings together per provider’s cached and you can uncached type in price at your cache strike price; organization instead of a printed cached-type in price are recharged during the complete enter in speed. Newest DeepSeek designs and you will API cost for every million tokensModelBest forInput (cache skip)OutputGuideDeepSeek-V4.1-FlashFast casual model. The new hosted service processes encourages for the DeepSeek’s infrastructure within the China, numerous governing bodies limit it for the authoritative products, and self-holding the new unlock weights prevents giving analysis everywhere. The sooner V3.1 and R1 designs are nevertheless commonly used and you may unlock-weight, and also the dated deepseek-speak and deepseek-reasoner API endpoints had been retired for the twenty four July 2026.
DeepSeek Harness is becoming inside designer preview to have representative use developers global. By changing the newest configuration, you should use the fresh OpenAI/Anthropic SDK or programs suitable for the newest OpenAI/Anthropic API to access the fresh DeepSeek API. Lookup / Bring API is becoming readily available, powered by the inside the-household internet search engine and help net search and fetch.
Cam assistantRAG / file Q&ACoding agentBulk group Heritage deepseek-talk and you may deepseek-reasoner aliases have been resigned; migrate in order to deepseek-v4-thumb. V4-Thumb and you will V4-Flash-Vision-Exp are retired; V4-Specialist desires approach to V4.1-Thumb from 14 September.
DeepSeek patterns and you will what they rates – sakura dragon slot
Once you have gotten an API key, you can access the brand new DeepSeek model by using the pursuing the analogy texts from the OpenAI API format. Distilled versions match an individual 24GB GPU; complete models you desire multi-GPU equipment. To the wrote API costs DeepSeek are much less per million tokens compared to GPT, Claude and Gemini flagships, and is sakura dragon slot also aggressive to your programming and you will cause criteria. All of our rates web page have a complete desk, cache-struck cost and you will speed background. Discover loads, equipment you would like, and you may step-by-action setup that have Ollama, vLLM and you may llama.cpp. Discover a great work, next see what a comparable regularity will cost you on every provider’s published API prices.
History API aliases resigned — surcharge however maybe not energetic

Full table, speed background as well as the retired V3/R1 endpoints are on the fresh DeepSeek prices web page. Totally free tier, runs the current V4 habits, produced by DeepSeek. I contain the model listing, API rates and you may log on actions seemed facing DeepSeek’s individual paperwork and you can printing the new review go out for each page. To possess examples by using the Anthropic API style, excite make reference to Anthropic API.
Utilize the official DeepSeek talk
- Discover a good work, up coming see just what a comparable regularity can cost you on each provider’s authored API costs.
- Before running these examples, lay the newest DEEPSEEK_API_Trick environment changeable in your critical on the API trick.
- Discover weights, tools you desire, and you may step-by-step setup with Ollama, vLLM and you may llama.cpp.
- To the wrote API cost DeepSeek are much less for each million tokens versus GPT, Claude and Gemini flagships, and is aggressive on the programming and cause standards.
- Our costs page features a full dining table, cache-strike costs and you can rate record.
- This can be a different guide, not the state provider.
DeepSeek publishes unlock loads on the GitHub and Hugging Face, and so they work on locally having Ollama, vLLM, llama.cpp or LM Facility. This is a different publication and you will neighborhood funding. Group savings, long-perspective surcharges over a great provider’s threshold, free sections and you may company sale are not included.
Invoke The new Speak API
Its most recent head model are DeepSeek-V4.1-Thumb (API name deepseek-flash), essentially offered while the 9 Sep 2026, having local picture information made in. DeepSeek AI try a Chinese AI look business one to creates highest code patterns. The official site try deepseek.com plus the certified talk try speak.deepseek.com. The row links to your provider’s individual speed webpage. Text and an experimental sight version.0.150.6Read DeepSeek-V4-ProFlagship reasoning and programming model, to possess more complicated multiple-action functions.0.661.98Read DeepSeek’s own speculative decryption bunch boats resources-alert schedulers and you will 2–3× token throughput to your H800s.
Are our totally free trial chat
Live changelog from design releases, rates changes and you will infrastructure change. DeepSeek is actually a good Chinese AI laboratory whoever large vocabulary habits is actually strangely cost effective to work with and you may open-lbs. That is a non-weight analogy, you could potentially place the fresh weight parameter in order to genuine to locate load reaction.

DeepSeek publishes discover loads for the GitHub and you can Hugging Deal with, and so they work on in your area with Ollama, vLLM, llama.cpp otherwise LM Business. This can be a different book and you may community financing. Group discounts, long-context surcharges over a good provider’s endurance, 100 percent free sections and you can firm product sales aren’t included.
DeepSeek Use has become inside the designer preview to have representative funnel builders global. Because of the changing the new setup, you can use the fresh OpenAI/Anthropic SDK or softwares suitable for the fresh OpenAI/Anthropic API to access the fresh DeepSeek API. Look / Fetch API has become readily available, powered by all of our inside the-home internet search engine and you can help net research and you will fetch.