DeepSeek raises prices: the price war has a price after all

On 13 August, DeepSeek did two things at once. It moved its flagship V4 Pro model out of preview, upgrading the version to DeepSeek-V4-Pro-0813 and making it available across the app, the web and the API, and it announced the largest price increase in the company’s short history, taking effect on 17 August. The numbers tell the story: for V4 Pro, a million output tokens at peak hours goes from 6 yuan to 27 yuan, a rise of 350 per cent; input tokens with a cache miss go from 3 to 9 yuan; and cache-hit input, the cheapest tier, rises elevenfold from 0.025 to 0.30 yuan. The lighter V4 Flash model follows the same pattern, with peak output up from 2 to 9 yuan. A new peak-and-off-peak pricing scheme means off-peak calls cost half the peak rate, which is the company’s way of asking developers to shift their work to quieter hours.

The company that built its reputation on undercutting the entire industry now argues that the discount had to end. DeepSeek has said the adjustment is about allocating compute resources more reasonably; Goldman Sachs analysts read the sector-wide price rises from DeepSeek and its Chinese peers this year as a sign that compute is tight and the industry is moving from a price war to rational pricing. The trigger is not hard to find: usage has exploded. One developer platform reported V4-Flash processing 8 trillion tokens in a single day on 1 August, and Chinese government data shows daily token consumption in the country has grown more than a thousandfold since early 2024. A model that cheap attracts that much traffic, and that much traffic costs real money to serve.

The technical case for the new model is strong. V4 Pro holds a million tokens of context, writes up to 384,000 tokens of output, and scores 62.7 on the DeepSWE software-engineering benchmark, up from 12.8 for the preview, ahead of Claude Opus 4.8 and, on some agentic benchmarks, ahead of Anthropic’s flagship Fable 5. The pricing story is therefore not a downgrade in disguise: the product improved, and the price went up anyway, which is what happens when a product is underpriced.

The wider meaning is that the era of free AI is ending by inches. The DeepSeek ‘death zone’ debate of late 2025, which held that the company’s prices made mid-tier models unviable, has been replaced by a different question: how long any lab can sustain frontier models at a loss. OpenAI and Anthropic are both heading toward listings with the market asking whether their margins can survive a price war; DeepSeek is doing its second round of fundraising at a valuation around 500 billion yuan and has reportedly begun preparations for its own IPO. Raising prices is what companies do when they need to show investors a path to profit, and DeepSeek is now showing one.

Even after the increase, the company remains dramatically cheaper than its rivals. V4 Pro’s peak output of 27 yuan per million tokens is roughly a twelfth of what Anthropic charges for Fable 5, and its input price about a seventh. The ‘price cutter’ label survives the increase in absolute terms; what has changed is the direction of travel. The interesting signal is not that DeepSeek raised prices, which was inevitable, but that it chose to announce the increase together with the flagship release rather than slip it in quietly. The company is signalling to its customers, and to the market, that its era of subsidized access is over, and that the next phase of the AI industry will be paid for by the people who use it.

The complaints have already started, in the usual places: developers noting that cache-hit rates have fallen since the formal version replaced the preview, and that a single coding session can now cost a noticeable amount of money. Some of those complaints are about the product getting more expensive while the cache gets less efficient, which is a real change and not just sticker shock. But the deeper point stands. For two years, the price of intelligence was treated as a marketing weapon, with every lab cutting rates to prove its models were undervalued. The industry has now reached the stage where the next model costs more to train than the previous one, and someone has to pay for it. DeepSeek’s increase is that bill arriving, and every other lab that has raised prices this year is being handed the same invoice.