News · 2026-08-05

DeepSeek warns of a significant API price rise, five days after being called 100 times cheaper

DeepSeek has added a footnote to its official pricing page warning that it plans to raise prices across its API services, with, in its own words, a significant increase expected. No number, no date, and no named models accompany the warning. It landed five days after the benchmarking firm Artificial Analysis published a widely covered study pricing DeepSeek's newest model at roughly 3 cents per test where a leading Western model cost more than 3 dollars, a comparison built entirely on the rates DeepSeek has now said it intends to retire.

Key facts

• The notice is a single footnote on DeepSeek's official pricing page, added between August 3 and August 6, 2026.
• Current published rates remain 14 cents per million input tokens and 28 cents per million output tokens, with cache hits at 0.28 cents.
• An earlier, more specific plan to double prices during Beijing business hours appeared around August 1 and was removed without ever taking effect.
• The affordability study it undercuts was published July 31 by Artificial Analysis.

The footnote reads, verbatim: "We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly. The specific pricing plan will be subject to official notice." That is the complete announcement. It names no models, though DeepSeek's API currently covers deepseek-v4-flash and deepseek-v4-pro.

The trail behind the footnote

The interesting part is what the footnote replaced. Internet Archive snapshots of the same page tell a three-step story. Through July 28, there was no pricing-change notice of any kind. By August 1, the page carried a specific plan: the API would soon adopt peak and off-peak pricing, with peak prices at double the regular rate, applicable to all billing items, during two windows totaling seven hours of Beijing business time each day. By August 6, that mechanism was gone, replaced by the vaguer warning above.

So DeepSeek floated a concrete pricing structure, then withdrew it in favor of an undated warning, and the flat rate is still what you are charged today. Whether that reflects internal disagreement, a scheduling change, or a decision that hourly pricing was too hostile to developers, DeepSeek has not said. It has given no reason for the increase at all. Coverage attributing it to accelerating demand or an industry retreat from unsustainable pricing is inference by outlets, including TechNode, not anything DeepSeek stated.

Why the timing stings

On July 31, Artificial Analysis published its evaluation of the newest DeepSeek release, scoring 50 on its Intelligence Index, ten points above the previous version. The index blends terminal and agentic tasks, a graduate-level science exam, a long-context reasoning test, and several coding and tool-use benchmarks, then plots each model on a cost-versus-intelligence frontier. Priced at DeepSeek's published rates, the model averaged roughly 3 cents per test in the index, against roughly 86 cents for Kimi K3, about 1.86 dollars for GPT-5.6 Sol, and about 3.15 dollars for Claude Fable 5. Forbes ran it as 105 times cheaper. That framing traveled everywhere over the following weekend.

The study is not wrong. Its denominator is simply provisional in a way nobody knew when it was published. Every claim currently circulating that open Chinese models are two orders of magnitude cheaper than Western frontier APIs is measured against a price the vendor has now formally warned it will move, with no floor disclosed.

This is a recurring hazard in AI cost comparisons. Cost-per-task numbers look like measurements of a model, and they are partly measurements of a business decision. The model did not get more expensive to run this week. Somebody decided the price was too low, which is a different fact about a different thing. Anyone building on cheap inference should read the footnote as a planning input rather than a headline, which is the same lesson that keeps recurring around the frontier price reversal and DeepSeek's rapid model iterations.

The honest caveat

Nothing has actually changed price yet. The current rates are live and unchanged, the peak-hour plan never took effect, and it is entirely possible the eventual increase is modest. DeepSeek's weights for prior generations remain openly available, so self-hosting caps the downside for anyone with the hardware, which is the practical argument for caring about open-weight models even when a hosted API is cheap. Bloomberg also reported the plan on August 6, though its article could not be independently read from here. What is confirmed is narrow and sufficient: DeepSeek said, on its own site, that a significant increase is coming, and declined to say how much or when.

Primary source, verified: read the paper →

Ground Truth. (2026, August 5). DeepSeek warns of a significant API price rise, five days after being called 100 times cheaper. Ground Truth. https://groundtruth.day/news/deepseek-says-a-significant-price-rise-is-coming-and-names-no-number.html

@misc{groundtruth:deepseek-says-a-significant-price-rise-is-coming-and-names-no-number,
  title  = {DeepSeek warns of a significant API price rise, five days after being called 100 times cheaper},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/deepseek-says-a-significant-price-rise-is-coming-and-names-no-number.html}
}

Topics: industry · deepseek · pricing · api · open-weight-models · china