Complete table, rates history as well as the retired V3/R1 endpoints are on the new DeepSeek costs webpage. Totally free level, runs the modern V4 patterns, produced by DeepSeek. I hold the design list, API prices and you will log on actions appeared up against DeepSeek’s individual paperwork and you may print the new review go out on each page. For instances utilizing the Anthropic API format, excite consider Anthropic API.
- Check in in the chat.deepseek.com — the state chat features a free of charge tier and you will works the modern V4 patterns.
- Secrets, endpoints, design IDs, prompt caching and you can password examples — and what altered in the event the dated talk and you will reasoner endpoints have been retired.
- The new combined input price brings together for each and every provider’s cached and you can uncached type in rate at the cache strike price; organization rather than a circulated cached-enter in price are billed in the full input rate.
- The earlier V3.step one and you may R1 models are still commonly used and you will open-weight, and the dated deepseek-talk and you may deepseek-reasoner API endpoints were retired to the 24 July 2026.
- Latest DeepSeek designs and you can API prices for each and every million tokensModelBest forInput (cache miss)OutputGuideDeepSeek-V4.1-FlashFast relaxed design.
This can be a separate publication, maybe not the state provider. Just before powering these types of examples, lay the newest DEEPSEEK_API_Key ecosystem changeable on your own terminal to the API secret. If you use products such Claude Code, GitHub Copilot, or OpenCode, you can utilize DeepSeek as the backend model personally — zero password necessary. The fresh DeepSeek API try supported by of many preferred AI broker and programming assistant equipment.
Versailles gold slot machine: See the DeepSeek Harness Book to have information
The newest DeepSeek API uses an API structure appropriate for OpenAI/Anthropic. The cost comparator in this article versailles gold slot machine reveals exactly what the exact same monthly workload will set you back for each provider’s very own published costs, which have a relationship to all of the rate card. For each million tokens, V4-Flash are $0.15 type in (cache skip) and $0.6 productivity; V4-Professional try $0.66 enter in and you will $step one.98 efficiency. The demonstration isn’t the authoritative solution and won’t work at V4. You may also try the fresh 100 percent free demo speak on this site, and that works DeepSeek V3.1 on the community paths (losing back into R1 whenever a path is busy).
Real time changelog out of model releases, cost changes and you can system change. DeepSeek are a Chinese AI research whose large code models are surprisingly cost effective to work at and you may discover-lbs. This really is a low-weight analogy, you can place the fresh stream parameter to correct discover load reaction.
The latest chief design try DeepSeek-V4.1-Thumb (API term deepseek-flash), essentially readily available because the 9 Sep 2026, that have native image expertise made in. DeepSeek AI try an excellent Chinese AI research business one generates high vocabulary models. The official web site is deepseek.com as well as the formal talk try cam.deepseek.com. The row backlinks on the provider’s own speed webpage. Text in addition to a fresh attention version.$0.15$0.6Read DeepSeek-V4-ProFlagship reason and you may programming design, to have more difficult multiple-step performs.$0.66$step 1.98Read DeepSeek’s individual speculative decryption bunch ships resources-aware schedulers and you can dos–3× token throughput for the H800s.
Once you’ve received a keen API trick, you can access the brand new DeepSeek model with the following example texts in the OpenAI API style.
Tips, endpoints, model IDs, quick caching and you will password examples — along with exactly what altered if the old chat and you will reasoner endpoints have been resigned. Check in in the cam.deepseek.com — the official talk have a totally free level and you can operates the modern V4 models. The fresh mixed enter in speed mixes for each and every provider’s cached and you will uncached input rates at the cache hit rate; team as opposed to a published cached-input speed is charged from the full input price. Most recent DeepSeek patterns and you may API costs for every million tokensModelBest forInput (cache skip)OutputGuideDeepSeek-V4.1-FlashFast everyday model. The new hosted services processes prompts for the DeepSeek’s infrastructure inside China, numerous governments restrict they to the formal devices, and you can self-holding the newest discover weights avoids giving analysis anyplace. The earlier V3.1 and R1 designs are nevertheless popular and open-pounds, and the dated deepseek-talk and you can deepseek-reasoner API endpoints had been resigned to the twenty-four July 2026.
Distilled variations complement just one 24GB GPU; complete models you want multiple-GPU tools. To the composed API cost DeepSeek is far less for each million tokens compared to GPT, Claude and you may Gemini flagships, and it is aggressive to your coding and you can reasoning standards. All of our cost web page features an entire table, cache-struck prices and you will rate records. Discover loads, equipment you need, and you may step-by-action configurations having Ollama, vLLM and you may llama.cpp. See a good workload, next see what a comparable frequency will set you back for each provider’s published API prices.