Keys, endpoints, model IDs, fast caching and you may password advice — in addition to what altered when the dated cam and you may reasoner endpoints was resigned. Sign in at the speak.deepseek.com — the official cam have a free of charge tier and you can operates the current V4 models. The new blended enter in price brings together for each and every slot starburst slot provider’s cached and uncached input speed at your cache strike speed; company rather than a circulated cached-enter in price try energized during the complete type in rate. Newest DeepSeek patterns and you can API cost for each and every million tokensModelBest forInput (cache skip)OutputGuideDeepSeek-V4.1-FlashFast relaxed design. The new managed services processes encourages for the DeepSeek’s infrastructure inside China, numerous governing bodies limit they on the formal gadgets, and you may self-holding the fresh unlock loads hinders sending investigation anyplace. The earlier V3.1 and R1 patterns are still popular and open-weight, and also the old deepseek-speak and you will deepseek-reasoner API endpoints was resigned on the 24 July 2026.
DeepSeek Harness is inside the developer examine for broker use designers around the world. By the altering the newest setting, you can use the brand new OpenAI/Anthropic SDK or programs compatible with the fresh OpenAI/Anthropic API to access the newest DeepSeek API. Lookup / Fetch API is now available, powered by the within the-home search and you can supporting internet look and you will get.
Chat assistantRAG / document Q&ACoding agentBulk class Legacy deepseek-chat and you may deepseek-reasoner aliases was resigned; migrate in order to deepseek-v4-flash. V4-Flash and V4-Flash-Vision-Exp is retired; V4-Pro needs path to V4.1-Flash from 14 September.
Slot starburst slot | DeepSeek habits and you may whatever they costs
![]()
Once you’ve received a keen API trick, you have access to the brand new DeepSeek model utilizing the following example texts on the OpenAI API style. Distilled variants complement just one 24GB GPU; full patterns you would like multi-GPU methods. For the authored API rates DeepSeek are much lower per million tokens than the GPT, Claude and you will Gemini flagships, and it is competitive to the programming and you will need standards. Our very own cost web page features the full table, cache-hit costs and you can speed background. Discover weights, tools you desire, and you can step-by-step options which have Ollama, vLLM and you can llama.cpp. See a work, next see just what an identical regularity can cost you on every provider’s wrote API cost.
Legacy API aliases retired — surcharge however perhaps not effective
Full desk, rate background as well as the resigned V3/R1 endpoints take the brand new DeepSeek prices web page. Free level, runs the current V4 patterns, made by DeepSeek. I contain the model number, API cost and you will sign on steps searched against DeepSeek’s own records and you may printing the new review day on each page. To have examples using the Anthropic API style, delight consider Anthropic API.
Make use of the official DeepSeek speak
- Find a workload, next see what the same frequency will cost you on each provider’s composed API costs.
- Just before running this type of instances, place the brand new DEEPSEEK_API_Key environment changeable in your terminal for the API trick.
- Open weights, methods you want, and you may action-by-action settings which have Ollama, vLLM and you may llama.cpp.
- To your published API costs DeepSeek is actually much lesser for each and every million tokens than the GPT, Claude and Gemini flagships, and it is aggressive on the programming and you will need benchmarks.
- Our very own cost webpage features an entire table, cache-strike cost and you will speed record.
- This can be another book, not the official services.
DeepSeek publishes unlock loads to the GitHub and Hugging Deal with, plus they work with in your area which have Ollama, vLLM, llama.cpp otherwise LM Studio. This is another publication and you will area funding. Batch deals, long-perspective surcharges a lot more than a provider’s tolerance, 100 percent free levels and you may business selling commonly provided.
Invoke The brand new Speak API
Its newest chief design are DeepSeek-V4.1-Flash (API identity deepseek-flash), fundamentally offered while the 9 Sep 2026, with local image information made in. DeepSeek AI is actually a Chinese AI research team you to definitely produces higher language habits. The official site is deepseek.com and the official talk try talk.deepseek.com. The line website links for the provider’s individual price web page. Text message along with an experimental sight variant.0.150.6Read DeepSeek-V4-ProFlagship cause and you will coding model, for harder multi-action work.0.66step 1.98Read DeepSeek’s individual speculative decoding bunch vessels resources-aware schedulers and dos–3× token throughput for the H800s.
Is our free trial talk
![]()
Alive changelog away from design launches, cost changes and you may infrastructure alter. DeepSeek is a great Chinese AI research whoever highest words patterns try surprisingly inexpensive to work at and discover-weight. This can be a non-weight example, you could set the fresh stream parameter to correct to locate stream response.
DeepSeek publishes unlock loads on the GitHub and Hugging Deal with, and work at locally that have Ollama, vLLM, llama.cpp otherwise LM Facility. That is another book and you can area investment. Group deals, long-perspective surcharges more than a provider’s tolerance, totally free tiers and you may corporation sales aren’t integrated.
DeepSeek Funnel is becoming inside the developer examine to have representative harness developers international. From the altering the newest setting, you should use the brand new OpenAI/Anthropic SDK or programs compatible with the newest OpenAI/Anthropic API to access the fresh DeepSeek API. Look / Get API is now available, run on our inside the-house internet search engine and you will support net lookup and you can get.