Keys, endpoints, design IDs, prompt caching and you may code advice — in addition to exactly what altered when the old speak and you may reasoner endpoints were resigned. Check in at the talk.deepseek.com — the official speak have a free tier and works the current V4 models. The new blended input speed combines for each provider’s cached and uncached enter in rates at the cache strike speed; company as opposed to a published cached-type in rates try billed during the complete type in rate. Latest DeepSeek models and API prices for every million tokensModelBest forInput (cache miss)OutputGuideDeepSeek-V4.1-FlashFast informal model. The newest organized service techniques encourages for the DeepSeek’s structure inside China, numerous governments limitation they for the formal gadgets, and you may thinking-holding the fresh unlock weights hinders sending study anyplace. The earlier V3.step one and R1 habits remain commonly used and you may open-pounds, and the dated deepseek-chat and you can deepseek-reasoner API endpoints had been resigned on the 24 July 2026.
DeepSeek Use is now in the creator examine to own broker funnel designers global. From the changing the brand new arrangement, you can utilize the fresh OpenAI/Anthropic SDK otherwise programs compatible with the fresh OpenAI/Anthropic API to gain access to the new DeepSeek API. Look / Bring API has become offered, running on our inside the-family google and you can supporting online research and you may bring.
Chat assistantRAG / document Q&ACoding casino kitty glitter agentBulk classification Legacy deepseek-speak and deepseek-reasoner aliases have been retired; migrate to help you deepseek-v4-thumb. V4-Thumb and you can V4-Flash-Vision-Exp try retired; V4-Expert desires approach to V4.1-Flash away from 14 September.
DeepSeek patterns and you will whatever they rates: casino kitty glitter
After you’ve gotten an API trick, you have access to the brand new DeepSeek model with the after the example texts from the OpenAI API style. Distilled alternatives match just one 24GB GPU; complete habits you need multiple-GPU methods. On the published API rates DeepSeek is actually much lesser for each million tokens versus GPT, Claude and you can Gemini flagships, and is also competitive for the coding and you may need criteria. The costs page has an entire dining table, cache-hit cost and speed records. Open loads, methods you would like, and you may action-by-action options with Ollama, vLLM and you can llama.cpp. Discover a great workload, up coming see what a similar volume can cost you for each provider’s authored API costs.
History API aliases resigned — surcharge still perhaps not energetic
Full table, speed record and also the retired V3/R1 endpoints take the new DeepSeek rates webpage. Free level, works the present day V4 designs, produced by DeepSeek. I secure the model list, API rates and you will log in actions seemed up against DeepSeek’s own paperwork and you may printing the new opinion go out on each web page. To own instances utilizing the Anthropic API style, excite reference Anthropic API.
Use the official DeepSeek speak
- See a good work, up coming see just what a similar regularity will set you back on each provider’s published API costs.
- Prior to running these examples, put the fresh DEEPSEEK_API_Key environment variable on the critical to your API secret.
- Discover loads, methods you would like, and you can step-by-step configurations which have Ollama, vLLM and llama.cpp.
- On the wrote API costs DeepSeek try much lower for each million tokens than the GPT, Claude and you will Gemini flagships, and is aggressive for the coding and you can need standards.
- Our cost webpage has the full table, cache-hit prices and rate history.
- That is an independent guide, not the state provider.
DeepSeek posts discover weights to your GitHub and you can Hugging Deal with, and work with locally that have Ollama, vLLM, llama.cpp otherwise LM Business. This really is a different guide and you may neighborhood funding. Group deals, long-context surcharges over a good provider’s threshold, free levels and you will company product sales aren’t provided.
Invoke The new Speak API
The most recent chief design is actually DeepSeek-V4.1-Flash (API identity deepseek-flash), essentially available since the 9 September 2026, with local image understanding made in. DeepSeek AI try a Chinese AI research business one to creates higher words patterns. The state web site is actually deepseek.com and also the official speak try talk.deepseek.com. All of the line links on the provider’s very own rates page. Text and an experimental attention variation.0.150.6Read DeepSeek-V4-ProFlagship reason and you may coding design, for harder multiple-action performs.0.66step one.98Read DeepSeek’s very own speculative decoding heap boats methods-aware schedulers and you may dos–3× token throughput to the H800s.
Try all of our free demo talk
Alive changelog from design releases, cost shifts and you will structure changes. DeepSeek try a Chinese AI lab whose large language habits are unusually cost effective to focus on and you may unlock-pounds. This really is a low-stream analogy, you could set the fresh load factor to genuine to find stream effect.
DeepSeek posts unlock weights for the GitHub and Hugging Face, and focus on locally which have Ollama, vLLM, llama.cpp otherwise LM Studio. That is a different guide and you will area funding. Batch discounts, long-context surcharges above a provider’s tolerance, free sections and you can business product sales aren’t incorporated.
DeepSeek Utilize is now within the developer preview for agent use builders worldwide. By modifying the new setup, you can use the newest OpenAI/Anthropic SDK or software programs suitable for the newest OpenAI/Anthropic API to get into the newest DeepSeek API. Lookup / Bring API is becoming offered, powered by our within the-house internet search engine and support internet search and fetch.