Loading...
Loading...
DeepSeek has permanently cut DeepSeek-V4-Pro API prices to one-quarter of the original list, locking in a 75% reduction after the promotional period ends May 31, 2026. The lower rates — affecting both input and output token billing — aim to reduce inference costs for developers, startups and enterprises using the multimodal model, potentially reshaping competitive API pricing. Separately, DeepSeek addressed reports that a special input string (notably "<think>") triggered occasional hallucinations. The company says the issue is a character-induced glitch with no security or privacy implications, and plans targeted retraining and fixes to improve robustness and prompt handling.
Permanent large price cuts lower inference costs for developers and businesses and can shift competitive API pricing dynamics; fixing a character-induced glitch addresses reliability and prompt-safety concerns for production use.
Dossier last updated: 2026-05-23 08:19:33
DeepSeek announced that its DeepSeek-V4-Pro model API will permanently move to one-quarter of its original price after the current 75% off promotion ends on May 31, 2026. The change was disclosed via the company's official WeChat account and applies to the API access tier for the V4-Pro model. This permanent price cut could lower costs for developers and enterprises using DeepSeek’s LLM services, potentially increasing adoption and competitive pressure in the Chinese AI API market. The move matters for startups, cloud providers, and AI platforms tracking model pricing and go-to-market strategies.
DeepSeek has made a permanent price cut for its V4 Pro model: after a promotional 75% discount ends on May 31, 2026, DeepSeek will set V4 Pro API rates to one-quarter of the original price. The company also reduced input cache-hit pricing to one-tenth of launch rates effective April 26, 2026, and published detailed per-1M-token rates for flash and pro variants, concurrency limits, and feature differences (thinking vs non-thinking modes, JSON output, tool calls). This matters for developers and startups budgeting for large-context, high-throughput LLM use—lower, predictable pricing and explicit billing rules make DeepSeek more competitive for commercial embedding, chat, and reasoning workloads. The firm warns prices may change and recommends monitoring the pricing page.
DeepSeek announced that the DeepSeek-V4-Pro API pricing will be permanently reduced to one-quarter of its original list price, effectively making the current 2.5x discount (25% of original) permanent after the promotional period ends on May 31, 2026. Previously listed rates showed input (cache hit) at ¥0.1 per million tokens, input (cache miss) at ¥12 per million tokens, and output at ¥24 per million tokens; the new permanent pricing will be one-fourth of those original levels. The move matters for developers, startups and enterprises using large multimodal models because it lowers inference costs, potentially accelerating adoption and changing competitive pricing dynamics in the AI API market. DeepSeek is the vendor behind the model.
DeepSeek said on May 19 that a special-character input (notably "<think>") occasionally caused its model to produce unexpected or hallucinatory responses, but after investigation the company concluded this is a character-triggered hallucination and poses no security or privacy breach. DeepSeek’s team plans targeted retraining and model fixes to improve recognition and handling of such special characters and to prevent similar anomalous outputs. The company emphasized its commitment to data security and invited user reports. The clarification aims to calm users who suspected dialogue leakage and outlines remediation steps for robustness against malformed inputs.