DeepSeek prices V4 at 97% discount

AI pricing is getting weirdly aggressive. DeepSeek didn’t just launch a new model family and hope people noticed — it used price as the weapon. In late April, the Chinese lab cut API costs across its lineup, pushed cached-input pricing way down, and put a temporary 75% discount on V4-Pro through May 31. That turns the story from “new model release” into “can frontier labs keep charging frontier prices?” (api-docs.deepseek.com) ### What actually changed? DeepSeek’s own pricing page shows two big moves. First, input cache-hit prices across all models were reduced to one-tenth of launch levels, effective April 26, 2026. Second, DeepSeek-V4-Pro got a temporary 75% discount, now extended through May 31, 2026. Those are not tiny tweaks — they directly target the parts of AI bills (api-docs.deepseek.com)ti-step agents. (api-docs.deepseek.com) ### Where does the “97% cheaper” claim come from? It comes from the low end of the comparison. OpenAI’s API page lists GPT-5.5 at $5.00 per 1 million input tokens and $30.00 per 1 million output tokens. DeepSeek’s V4-Flash sits far below that on input pricing, at $0.14 per 1 million input tokens, which is about 97% lower than GPT-5.5’s input rate. That(api-docs.deepseek.com)h OpenAI’s flagship tier, not like-for-like premium models. (api-docs.deepseek.com) ### So is V4-Pro also that cheap? No — and this is the important catch. The 97% figure is mostly about V4-Flash, the cheaper model in the family. V4-Pro is the stronger reasoning model, and even after the temporary discount it is still priced above Flash. The actual move is more nuanced: DeepSeek is building a ladder, with Flash as the “cheap enough to try everywhere” option and Pro as the upsell for harder tasks. (api-docs.deepseek.com) ### Why do cache prices matter so much? Because agents repeat themselves. A normal chatbot answers one prompt and stops. An agent loops through plans, tools, retries, memory, and long context windows. That means token bills pile up fast, especially on repeated context. Cutting cache-hit costs to one-tenth is basically a subsidy for serious production u(api-docs.deepseek.com)nowledge base over and over. (api-docs.deepseek.com) ### Is this about prestige or market share? Market share, basically. DeepSeek is not trying to win the “most advanced model on Earth” argument on price alone. It is trying to become the default choice for developers who care about cost per task, not benchmark bragging rights. That is a smart place to attack, because many enterprise buyers do not need t(api-docs.deepseek.com)nd dramatically cheaper at scale. (scmp.com) ### What does this threaten for rivals? It pressures the pricing umbrella. OpenAI, Anthropic, and others can still argue for better quality, reliability, tooling, and enterprise controls. But once a credible model gets “good enough,” huge price gaps become hard to defend for bulk workloads. That is e(scmp.com)l-selection debate. (openai.com) ### Does open access change the equation? Yes. DeepSeek’s V4 release has also been framed as open-source or open-weights, which lowers adoption friction beyond API price alone. If developers can inspect, fine-tune, or self-host parts of the stack, the switching cost drops. That makes the pricing move more dangerous to closed labs than a simple temporary sale would be. (msn.com)ntelligence/deepseek-v4-released-as-open-source-but-falls-short-of-us-frontier-models-analysts-say/ar-AA22bu4f)) ### Bottom line? This is really a story about AI becoming a utility business faster than frontier labs wanted. DeepSeek is betting that for a lot of real deployments, “cheap, fast, and good enough” beats “best, expensive, and closed.” If that bet works, the next AI price war won’t be about chatbots — it’ll be about agents.