This is an independent fan site.Visit official DeepSeek.com
DeepSeek AI: Chat, Models & Independent Guides
DeepSeek is a Chinese AI lab whose large language models are unusually cheap to run and open-weight. Its current models are DeepSeek-V4-Pro for hard reasoning and coding and DeepSeek-V4-Flash for fast everyday use; the earlier V3.1 and R1 models are still open and popular but no longer the flagship line.
This is an independent guide, not the official service. We keep the model list, API rates and login steps checked against DeepSeek's own documentation and print the review date on every page.
Use the official DeepSeek chat
Free tier, runs the current V4 models, made by DeepSeek. You sign in on their site.
Try our free demo chat
No signup, but it is not the official service: it runs DeepSeek V3.1 on free community routes (R1 fallback) and does not run V4.
Open the demo chatDeepSeek V4 Tracker
Live changelog of model releases, pricing shifts and infrastructure changes.
DeepSeek-V4-Flash official API in public beta
Same model string, same architecture, new post-training: Terminal Bench 2.1 82.7 and DeepSWE 54.4 beat V4-Pro-Preview. Native Responses API and Codex support. V4-Pro and the app are unchanged.
Read updateLegacy API aliases retired — surcharge still not active
Legacy deepseek-chat and deepseek-reasoner aliases were retired; migrate to deepseek-v4-flash. A peak-hour surcharge (UTC+8) has been announced but is not active yet — the official rate card still shows one flat tier.
Read updateInside DSpark: semi-autoregressive drafters
DeepSeek's own speculative decoding stack ships hardware-aware schedulers and 2–3× token throughput on H800s.
Read updateDeepSeek raises $7.4B under a founder-locked LP vehicle
50 billion yuan round with 5-year lock-up. Structure keeps investors passive while Liang Wenfeng retains control.
Read updateDeepSeek models and what they cost
Rates are per million tokens, last verified against DeepSeek's official rate card on 2026-07-25. Official pricing page.
| Model | Best for | Input (cache miss) | Output | Guide |
|---|---|---|---|---|
| DeepSeek-V4-Flash | Fast everyday model. Text plus an experimental vision variant. | $0.14 | $0.28 | Read |
| DeepSeek-V4-Pro | Flagship reasoning and coding model, for harder multi-step work. | $0.435 | $0.87 | Read |
Cache hits cost a fraction of that. Full table, price history and the retired V3/R1 endpoints are on the DeepSeek pricing page.
Compare what your workload really costs
Pick a workload and see the monthly bill on DeepSeek V4-Flash and V4-Pro next to the GPT, Claude and Gemini models people actually weigh them against — all from published rate cards, each row linking to the provider's own price page.
DeepSeek vs GPT, Claude and Gemini: monthly cost comparator
Pick a workload, then see what the same volume costs on each provider's published API rates. DeepSeek rates verified 2026-07-25; competitor rates verified 2026-07-25. Every row links to the provider's own price page.
| Model | Input / 1M | Output / 1M | Monthly cost | vs cheapest | Rate card |
|---|---|---|---|---|---|
| DeepSeek-V4-FlashDeepSeek | $0.14 | $0.28 | $9 | cheapest | source |
| DeepSeek-V4-ProDeepSeek | $0.435 | $0.87 | $27 | 3.1× more | source |
| Gemini 3.1 ProGoogle | $2.00 | $12.00 | $300 | 34.4× more | source |
| GPT-5.4OpenAI | $2.50 | $15.00 | $335 | 38.3× more | source |
| Claude Opus 4.8Anthropic | $5.00 | $25.00 | $618 | 70.7× more | source |
| GPT-5.5OpenAI | $5.00 | $30.00 | $669 | 76.6× more | source |
Method: monthly cost = (input tokens × calls × blended input rate) + (output tokens × calls × output rate). The blended input rate mixes each provider's cached and uncached input price at your cache hit rate; providers without a published cached-input rate are billed at full input price. Batch discounts, long-context surcharges above a provider's threshold, free tiers and enterprise deals are not included. DeepSeek has announced a peak-hour surcharge but has not confirmed the percentage or start date, so it is not applied here.
Independent DeepSeek guides
Source-checked walkthroughs for the questions people actually arrive with.
DeepSeek API guide
Keys, endpoints, model IDs, prompt caching and code examples — including what changed when the old chat and reasoner endpoints were retired.
Read the API guideRun DeepSeek locally
Open weights, hardware you need, and step-by-step setup with Ollama, vLLM and llama.cpp.
Local deployment guideIs DeepSeek safe?
Where your prompts go, which governments restrict it, and how self-hosting changes the answer.
Privacy & safetyDeepSeek: frequently asked questions
Prefer a sidebar in your browser?
Our free Chrome extension puts an AI sidebar next to any page. Built by us, not by DeepSeek.