--- title: DeepSeek's Coming Price Rise Ends The Idea That Chinese AI Labs Play By Different Rules description: DeepSeek says a significant API price increase is coming. Its own published figures show how much margin it was leaving on the table to win developers. author: Darie Nani (Editor-in-Chief) updated: 2026-08-06T23:16:14.056Z canonical: https://www.sovereignmagazine.com/article/deepseek-api-price-rise-market-share image: https://cdn.nanimediahouse.com/deepseek-api-price-rise-114324.webp categories: Artificial Intelligence content_type: Analysis region: Global publication: Sovereign Magazine schema_type: Article --- DeepSeek's own [pricing page](https://api-docs.deepseek.com/quick_start/pricing) now carries a line telling developers to expect a significant increase in what its API costs, and to plan their usage accordingly. It does not say when. It does not say how much. The specific plan, the note says, will be subject to official notice. The cheapest serious API on the market getting more expensive is not a twist if you have read what DeepSeek has published about its own economics. The company has been open about charging a fraction of what its tokens were worth, and it has been taking the discounts back one at a time since last summer. What the price rise does end is the story that grew up around DeepSeek in the West: that a Chinese lab was handing out frontier models cheaply because it thought about its customers differently from the American ones. Undercutting on price is how a new entrant takes a market, wherever it is based. Inference costs money, and a lab that sells it below cost is buying something with the difference. ## DeepSeek Published The Margin Behind Its Own Cheap Prices In a technical post on GitHub in February 2025, DeepSeek gave a detailed account of what it cost to run V3 and R1 for a single day. Over the 24 hours to 28 February 2025, the services ran on an average of 226.75 nodes of eight H800 GPUs each. At an assumed leasing cost of $2 per GPU hour, that put the day's infrastructure bill at $87,072. In the same period the system took in 608 billion input tokens, more than half of them served from a disk cache, and produced 168 billion output tokens. DeepSeek then did the sum a customer would recognise. If every one of those tokens had been billed at R1's prices at the time, $0.14 per million input tokens on a cache hit and $2.19 per million output tokens, the day's revenue would have come to $562,027 against that $87,072 of cost, which the company described in its own words as “a cost profit margin of 545%”. Actual revenue, it went on, was substantially lower, and it listed the three reasons: V3 was priced well below R1, only a subset of services were charged for while web and app access stayed free, and nighttime discounts applied automatically off peak. Those are the choices of a company that had worked out exactly what it could charge and decided not to charge it yet. Free consumer access, an underpriced flagship and an automatic discount overnight all buy the same thing, which is [developers who build on you](https://www.sovereignmagazine.com/article/microsoft-ai-token-budget-copilot-engineers). ## The Discounts Have Already Come Off One At A Time The pricing has been moving for two years, and not only downward. In August 2024 DeepSeek added context caching to disk and said the change was “reducing prices by another order of magnitude”. Then in August 2025, buried in the release notes for V3.1, came a line that got far less attention than the model did: new pricing would start and off-peak discounts would end at 16:00 UTC on 5 September. One of the three reasons DeepSeek had given for its own thin revenue was gone. Three weeks later the headline number fell again. The V3.2-Exp notes in September 2025 announced API prices “cut by 50%+, effective immediately”. In April 2026 the company released V4-Pro and V4-Flash with [open weights](https://www.sovereignmagazine.com/article/white-house-open-weight-ai-models-security-testing) and million-token context windows, under the banner of “the era of cost-effective 1M context length”. Three months after that, the same pricing page is telling developers a significant rise is on the way. Nothing in this is a broken promise. That page has always carried the line that prices may vary and that DeepSeek reserves the right to adjust them. The increase was an option the company kept open in the small print, including in the months it was cutting. ## Anthropic Prints The Same Structure On Its Own Page Anthropic lists Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens, and [prints the expiry date beside them](https://www.anthropic.com/pricing): introductory pricing through 31 August 2026, then $3 and $15 as standard. Its other current models sit on the same kind of scale, Opus 5 at $5 and $25, Haiku 4.5 at $1 and $5, Fable 5 at $10 and $50. OpenAI's flagship rates run from gpt-5.6-luna at $0.20 and $1.20 up to gpt-5.6-sol at $5 and $30. An introductory price with a date on it, and a higher standard price behind it, is a structure printed in plain dollars on a San Francisco company's own website. DeepSeek's version is less precise and much cheaper. deepseek-v4-flash currently lists at $0.14 per million input tokens on a cache miss and $0.28 on output, and deepseek-v4-pro at $0.435 and $0.87, both with million-token context. Anthropic has already fixed its new numbers and its date; DeepSeek has fixed neither, and developers get the warning without either figure. ## Anyone Budgeting On DeepSeek's Current Prices Should Treat Them As Temporary Teams that moved workloads to DeepSeek because the tokens were cheap picked a number the company has now said will change. The only guidance available is DeepSeek's own: watch for the official notice, and plan usage accordingly. Anyone modelling costs against the current range of $0.14 to $0.87 per million tokens on V4 should treat that range the way Anthropic's own customers have to treat $2 and $10 on Sonnet 5, as a price with a shelf life. The open weights are the hedge, and DeepSeek has kept publishing them. A team that can serve V4 itself has somewhere to go when the new numbers arrive. A team that wired its product to the API on the strength of the price is waiting on a notice with no date. ## FAQ **Q: How much does the DeepSeek API cost right now?** Per million tokens, deepseek-v4-flash is $0.14 for input on a cache miss, $0.0028 on a cache hit and $0.28 for output. deepseek-v4-pro is $0.435 for input on a cache miss, $0.003625 on a cache hit and $0.87 for output. Both models carry a million-token context window, and DeepSeek's page states that prices may vary and it reserves the right to adjust them. **Q: How does that compare with Claude and GPT prices?** Anthropic lists Claude Sonnet 5 at $2 input and $10 output per million tokens through 31 August 2026, rising to $3 and $15 after that date, with Opus 5 at $5 and $25, Haiku 4.5 at $1 and $5 and Fable 5 at $10 and $50. OpenAI's flagship rates run from gpt-5.6-luna at $0.20 and $1.20 to gpt-5.6-sol at $5 and $30. DeepSeek's current list prices sit below all of them. **Q: When does the DeepSeek API price increase take effect?** DeepSeek has not said. Its pricing page states only that a significant increase is planned in the near future and that the specific pricing plan will be subject to official notice. **Q: Is DeepSeek still free to use on the web and in the app?** In its February 2025 account of its own service, DeepSeek described web and app access as unmonetised while API usage was billed. The notice on the pricing page concerns the API. **Q: Why are AI inference prices going up across the industry?** Serving a large model costs GPU time whoever runs it, and the current generation of API prices was set to win developers rather than to cover that cost with a margin. DeepSeek's own figures show it collecting a fraction of what its tokens were worth at list prices while it grew, and Anthropic's page shows the same mechanism with the step-up already dated.