DeepSeek Shakes Things Up: V4 Pro at a Quarter of the Price, US Competition Under Pressure

In recent days, DeepSeek has made permanent the 75% reduction on the prices of its flagship model, V4 Pro. The promotion, originally set to expire on May 31, 2026, now becomes the new standard list price: the Chinese lab has set the price of the V4 Pro APIs between $0.003625 and $0.87 per million tokens, down from the previous $0.0145 and $3.48. The cut applies to all ranges but particularly focuses on input components with cache hits, where the cost drops to about a tenth of the original price, according to the official pricing page.

The decision comes just over a month after the launch of the V4 family, which took place on April 24 and consists of the Pro variants (totaling 1.6 trillion parameters in Mixture-of-Experts architecture, with 49 billion active for inference) and Flash (284 billion). Both versions support a context of one million tokens and a maximum output of 384,000 tokens per single request, licensed under MIT. At launch, DeepSeek defined the move as the start of the "low-cost 1M token context era," a stance that becomes even more valid after the decision for a permanent cut.

Huawei Ascend Supply Remains the Uninvited Guest

Already at the launch in April, DeepSeek had indicated that the price of the Pro would drop significantly once the Huawei Ascend 950 supernodes become available in significant volumes, a condition expected in the second half of 2026. The Pro specifications, which were up to 12 times more expensive than the Flash at the time of release, were justified by the "high-end computing capacity constraints."

The link to Chinese infrastructure is strategic as well as economic. The V4 family was adapted for Ascend 950 from day zero, and DeepSeek did not grant early access to NVIDIA and AMD during the pre-launch optimization window, a reversal of the industry’s established practice. The permanent discount may somewhat reflect a cost improvement on the inference side stemming from the domestic hardware ecosystem, although the lab has not released any statement in that regard.

The Gap with Western Rivals Widening

The new V4 Pro prices place it in the lower tier of the frontier model market. Claude Opus 4.7 from Anthropic costs $5 per million tokens in input and $25 in output; OpenAI GPT-5 starts from $2.50 for input and $10 for output. On output, V4 Pro at $0.87 is about 28 times cheaper than Claude Opus 4.7 and 11 times cheaper than GPT-5. The gap widens in workloads with high cache hits, where the input price drops to $0.003625 per million tokens, a typical scenario for AI agents with repetitive prompts over extended sessions. Only Google Gemini 3.5 Flash, the low-cost model from Mountain View, remains more affordable in absolute value with $0.15 for input and $0.60 for output, though it is positioned in a distinct tier of performance and purpose.

Imagining the application of new prices to enterprise workloads, the advantage would be significant. For example, Salesforce has projected an annual spending of $300 million on Anthropic tokens for 2026. At the same volume of tokens, applying DeepSeek’s new pricing, that spending would drop to around $20-30 million, between one-tenth and one-fifteenth of the total, depending on the mix of input, output, and cache hits. However, the advantage clashes with compliance and geopolitical risks when transitioning sensitive workloads through a Chinese provider, in a context where Anthropic accused DeepSeek of a distillation attack last February.

The updated list price is already visible on the lab's API page. Verification on the enterprise volume front will come in the next quarters, while the availability timing of Huawei Ascend 950 supernodes in the second half of 2026 remains the variable to observe to measure the sustainability of the cut.