Complete desk wild swarm win , rates background plus the retired V3/R1 endpoints take the fresh DeepSeek rates web page. 100 percent free level, operates the modern V4 models, produced by DeepSeek. We hold the design listing, API rates and you can sign on actions searched against DeepSeek’s individual paperwork and you will print the newest review date on every webpage. To have advice with the Anthropic API style, excite consider Anthropic API.
- Sign in from the talk.deepseek.com — the state speak features a free tier and works the present day V4 patterns.
- Keys, endpoints, design IDs, punctual caching and you may code advice — as well as just what changed when the old cam and reasoner endpoints was resigned.
- The brand new blended type in rates brings together for each provider’s cached and you may uncached input rates at your cache struck rate; company as opposed to a published cached-type in rates is energized at the full enter in price.
- The sooner V3.1 and R1 patterns remain popular and you may discover-lbs, plus the old deepseek-cam and deepseek-reasoner API endpoints were retired for the twenty four July 2026.
- Newest DeepSeek patterns and you will API cost per million tokensModelBest forInput (cache miss)OutputGuideDeepSeek-V4.1-FlashFast relaxed model.
This is a separate publication, perhaps not the state provider. Just before powering this type of instances, place the brand new DEEPSEEK_API_Trick ecosystem variable on your own terminal for the API key. If you are using products such as Claude Code, GitHub Copilot, or OpenCode, you need to use DeepSeek since the backend design individually — zero code required. The new DeepSeek API are backed by of numerous common AI agent and you will programming secretary systems.
Wild swarm win: See the DeepSeek Harness Book to possess details
The brand new DeepSeek API spends a keen API structure suitable for OpenAI/Anthropic. The cost comparator in this post shows exactly what the exact same monthly workload will set you back on each provider’s very own published cost, having a relationship to the price credit. For each and every million tokens, V4-Flash try $0.15 input (cache miss) and you can $0.six efficiency; V4-Professional are $0.66 type in and you may $step one.98 productivity. Our very own demo isn’t the authoritative solution and will not focus on V4. You may also is the newest free demonstration chat on this web site, and therefore runs DeepSeek V3.1 to the area pathways (losing back into R1 whenever a road is busy).

Real time changelog out of design launches, prices shifts and you may system transform. DeepSeek is actually a great Chinese AI research whoever high language designs try surprisingly cheap to work on and unlock-weight. This can be a non-load analogy, you might put the newest load factor to genuine to locate weight effect.
The newest fundamental design is DeepSeek-V4.1-Thumb (API label deepseek-flash), generally readily available while the 9 September 2026, with local visualize knowledge built in. DeepSeek AI is a great Chinese AI look organization one produces highest language models. The state site is deepseek.com and also the formal cam try speak.deepseek.com. The line links on the provider’s very own price page. Text as well as an experimental vision variation.$0.15$0.6Read DeepSeek-V4-ProFlagship cause and programming model, to have harder multiple-step performs.$0.66$1.98Read DeepSeek’s own speculative decryption bunch ships resources-aware schedulers and you may dos–3× token throughput to the H800s.
Once you’ve acquired a keen API trick, you can access the brand new DeepSeek design with the pursuing the example scripts on the OpenAI API format.
Tips, endpoints, design IDs, prompt caching and code examples — and just what altered in the event the old talk and reasoner endpoints had been retired. Sign in from the chat.deepseek.com — the state speak provides a free of charge level and you will operates the present day V4 designs. The fresh mixed enter in rate brings together per provider’s cached and you can uncached enter in rate at your cache hit speed; company instead a released cached-type in price are recharged in the complete type in speed. Current DeepSeek models and you may API rates per million tokensModelBest forInput (cache miss)OutputGuideDeepSeek-V4.1-FlashFast casual design. The brand new managed provider procedure encourages on the DeepSeek’s infrastructure inside the Asia, numerous governments limit they on the authoritative products, and you may mind-hosting the fresh open weights hinders giving study anywhere. The sooner V3.step one and you may R1 habits are still commonly used and you can open-lbs, and also the dated deepseek-speak and you will deepseek-reasoner API endpoints have been retired to the twenty-four July 2026.
Distilled versions complement just one 24GB GPU; complete models you desire multiple-GPU resources. On the composed API prices DeepSeek is actually much smaller for every million tokens than the GPT, Claude and you may Gemini flagships, and is aggressive for the coding and you will reason standards. The prices webpage has the full desk, cache-hit rates and you may price records. Open loads, methods you would like, and you will step-by-step setup having Ollama, vLLM and llama.cpp. Come across a good work, next see what a similar volume costs on every provider’s published API prices.
