Model Prices Are Falling Fast. Why Overseas AI Bills Are Not

Model prices fell sharply in the last week of September. Anthropic's Claude Opus 5.5 costs about 40% less to use than its predecessor, and OpenAI's low-cost GPT-6 variants Sol and Luna are half the price of the GPT-5.6 series. For companies doing business overseas, the signal is clear: models are becoming cheap public goods. What remains expensive is the step of putting them into everyday workflows.

Pricing logic shifted gear between September 22 and 25

On September 22 US time, Anthropic and OpenAI released new models on the same day, about 90 minutes apart. Claude Opus 5.5 went live at 00:31 Beijing time, with the GPT-6 series following at 02:00, according to Sohu on September 23. On price, Cailianshe gave two figures on September 25: Claude Opus 5.5 costs about 40% less to use than the previous generation, while OpenAI's low-cost GPT-6 versions Sol and Luna are 50% cheaper than the GPT-5.6 series.

Chinese vendors moved earlier. On September 22, Xiaomi released and open-sourced the MiMo-V2.6 series, with MiMo-V2.6-Pro holding a unit task cost of $0.13. According to a Cailianshe comparison, at comparable intelligence that price is one-twentieth to one-sixtieth of overseas models. Since September, DeepSeek has also announced a new round of price adjustments for its Flash series.

The reaction showed up in the capital markets. On September 25, shares of Zhipu (02513.HK) fell to around HK$610.5 at one point, taking its market capitalisation back to around HK$300bn, a new low for the recent correction, according to Cailianshe.

What the market fears is not any single company, but the fact of low pricing itself.

What is getting cheaper is not just the API

This round of cuts is not a promotion. It is the inevitable result of a narrowing capability gap.

Stanford University's 2026 AI Index Report puts the performance gap between top Chinese and American AI models at 2.7% as of March 2026, according to Sing Tao Daily on September 25. When leading models differ by only a few percentage points, "who is strongest" stops being a procurement reason. "Who can finish the same job at lower cost" becomes one.

Analysis cited by Cailianshe is blunter: Chinese models once opened markets on value for money, but once the price war spread, low prices stopped being any one vendor's advantage and became the industry norm. Goldman Sachs analysts expect several more model releases in the second half of 2026, with total parameter scale expanding to 2trn to 5trn, the fiercest competition in coding, and continued pressure on low-end API pricing.

For those who use AI as a production tool, this means two things.

First, do not tie your business to a single vendor's model. One review counted seven frontier-level models released by Anthropic and OpenAI between May 28 and September 24, 2026 alone, with the front-runner changing roughly every four to eight weeks.

Second, stop paying a premium for "the newest and strongest". Today's best becomes tomorrow's standard configuration.

Why the saved token money never shows up on the books

This is the point we most want to flag: model unit prices fell 40% or 50%, yet many companies' AI spending has not fallen by the same proportion.

The reason is not complicated. Tokens are only the most transparent slice of the bill.

What really eats the budget are costs that do not look like AI: a product copy that takes an afternoon to write, a poster that waits for a designer's schedule, a dozen social media posts that need publishing every day. These tasks used to be handled by piling on outsourced labour. Many have now switched to AI, but not completely. Three or four tools get bought, each needing to be learned and switched between. The API fees saved come back as time and communication costs.

The gains from cheaper models only materialise for companies that fit model capability into specific workflows.

How to put it to work

If your business is content acquisition aimed at overseas Chinese markets, this kind of work can be handed to SHEYU AGENT. Its 16 industry strategists each cover a professional area, and 34 zero-threshold tools cover Xiaohongshu image-and-text posts, AI posters and copywriting, producing finished output in one click, with one login working across desktop, mobile and web.

It addresses exactly the problem above: turning content output from "outsourcing plus human relay" into same-day production by your own team, so that the money saved on cheaper models lands in content capacity rather than in extra accounts nobody uses.

The model price war will continue, and APIs will only get cheaper. But your operating costs will not fall automatically with them.

Cutting costs was never about buying more cheaply. It is about putting already-cheaper capability into the work you do every day.