According to an official StepFun announcement, as of 00:30 Beijing time on October 14, 2026, the company's open platform has doubled the cache-hit input price for the Step 5 Preview pay-as-you-go API, from 0.35 yuan to 0.70 yuan per million tokens. Cache-miss input stays at 7 yuan and output at 20 yuan. The notice also spells out two carve-outs: users who had already called the model through the pay-as-you-go API before 00:30 on October 14 will continue to be billed at the original 0.35 yuan rate, and cache pricing on the Step Plan subscription tier is untouched.

China's large-model API price war has been running for three years. This is a move in the opposite direction, and it landed one day before the model's weights went open source.

What the Price Notice Says

The "Step 5 Preview API Cache Price Adjustment Notice" on StepFun's open platform site is a four-line table: billing item, price before, price after. Cache-hit input goes from 0.35 yuan to 0.70 yuan; the other two items are marked "unchanged." At the bottom, the notice draws the billing line between new and existing users.

Step 5 Preview, the model being repriced, is StepFun's flagship foundation model, released in the second half of September. It uses a sparse MoE architecture with 600 billion total parameters, roughly 27 billion activated per token, and is tuned for AI coding, software engineering, professional knowledge work and finance. It supports a 1 million token context window.

The size and scope of the increase are what matter. Raising only cache-hit input, while leaving cache-miss input and output alone, is the technical detail most likely to be missed. Cache hits are what you get when an agent reads the same stretch of context over and over: multi-turn tool calls, codebase comprehension, long-document analysis. In other words, the increase lands squarely on how heavy agent users actually consume tokens, rather than lifting the price of entry across the board. The company has not explained why this line item and not the other two.

The Rationale Is Outside Speculation; the Company Isn't Commenting

Beijing Business Today approached people connected to StepFun about the mixed raise-some, hold-some strategy; as of publication, there had been no response. The only public explanation so far comes from third parties.

Wang Chao, founder of the think tank Wenyuan Think Tank, offers a two-part reading: "The immediate reason for StepFun's price increase is that computing power across the industry is tight right now. Raising prices is a forced choice and has a certain logic to it, but from a business standpoint it's hard to understand, because large-model API prices in this industry are most likely headed into a declining range." He adds that the current round of price cuts traces back to Anthropic's Haiku 5.5: "It hasn't filtered through to China yet, but if the top-tier models hold this price cut, China's leading models will very likely follow. StepFun isn't in the first tier. Purely on business logic, raising prices now is unwise."

The pricing disadvantage shows up clearly in the data. Per OpenRouter figures cited by Beijing Business Today on October 12, DeepSeek-V4.1-Flash, the world's most-called model that day, charged 0.02 yuan per million tokens for cache-hit input during off-peak hours, 1 yuan for cache misses and 4 yuan for output. Even at its peak-hour rates, those three figures come to 11%, 29% and 40% of Step 5 Preview's new prices. Wang's "not in the first tier" has a direct price-sheet equivalent.

Repricing a Day Before Open Source Sets a Price Floor Against Self-Hosting

StepFun stacked three moves into three days: an agent-native phone on October 13, the API price change on October 14, and the full open-sourcing of Step 5 Preview's weights on October 15, spanning devices, pricing and models.

The timing, one day before the weights dropped, suggests a direct connection. Once weights are downloadable, customers gain a self-hosting option, so the API price has to justify itself first. Without the increase, customers could simply run the weights themselves from day one, on an entirely different cost structure.

The other signal is call volume. OpenRouter data put Step 5 Preview second globally in calls on October 12 Beijing time and fourth for the week, but calls that day were down 22% from the day before. The same day, Tencent's WorkBuddy integrated the model and opened it free for two weeks. Beijing Business Today's read is that an agent-native phone could push the model out of developer circles and into the mass market, while "the swings in call volume are an indirect confirmation of how fierce competition in this industry has become."

The Capital Picture: Valuation Rank and an Asset-Heavy Profile

Among China's "six little dragons" of large models, StepFun sits a rung below the leaders on valuation, but it is unmatched in how asset-heavy it is. Beijing Business Today reports that founder Yin Qi chairs both StepFun and Qianli Technology, and Qianli is the listed-company window he can lean on. In the first half of 2026, Qianli posted revenue of about 5.12 billion yuan, up 22.36% year on year, and non-GAAP net profit of roughly 55.7 million yuan, swinging from a loss.

On the financing side, reports in May said StepFun was set to close a round of nearly $2.5 billion, had dismantled its red-chip structure and was accelerating preparations for a Hong Kong IPO. The latest round reportedly drew a cluster of supply-chain investors including Huaqin, Longcheer, OmniVision and ZTE, spanning device assembly through upstream core components. StepFun has not responded to these reports.

Zhang Xinyuan, head of the consultancy Kefangde, told Beijing Business Today that the appeal of this path lies in its "greater emphasis on agents, device integration and open source … especially the story it can tell around entry points like phones and cars." He also drew two lines to watch: "First, whether the revenue elasticity after the price change is sustainable. Second, whether the agent and device push delivers real paying users and profit. If there are plenty of launches but a weak commercial loop, the appeal gets discounted."