iFlytek makes Spark Lite API free in China's model price war

Background

Most companies that use a large language model do not train one themselves. They call a model over the internet through an application programming interface (API) and pay for what they consume. The unit of billing is the token, a fragment of text roughly the size of a word or part of a word, and Chinese providers typically quote a price per 10,000 tokens. Because a single chat product can send and receive very large volumes of text, the quoted unit price decides whether an application built on top of a model is affordable at all, and it also shapes which provider a developer tries first.

What iFlytek announced

On 22 May 2024, iFlytek announced that Spark API capabilities were open for free, and that the Spark Lite API would be free permanently ^[1]. In the same announcement the company put a price on its strongest model: the top-tier Spark3.5 Max API was priced at RMB 0.21 per 10,000 tokens ^[1]. The two parts fit together. The entry-level service is given away, and the most capable model in the family is still sold by the token.

The cuts around it

The iFlytek move followed a run of price cuts by other Chinese providers in the same weeks ^[1]. Alibaba Cloud cut the input price of its Qwen-Long model from 0.02 to 0.0005 yuan per 1,000 tokens, a 97% cut ^[1]. Tencent Cloud cut the input price of Hunyuan-pro from 0.1 to 0.03 yuan per 1,000 tokens, a 70% cut ^[1]. Baidu Intelligent Cloud went further for part of its range and made two of its main Wenxin models fully free ^[1]. ByteDance, whose Doubao model was one of the services involved, also cut prices during the same period ^[1].

Reading the numbers

Two details matter for anyone comparing these figures. First, providers usually charge separately for input and for output — the text a developer sends in and the text the model generates — and the reductions described here were on input prices ^[1]. Second, model families are commonly sold in tiers, with a smaller member and a flagship; a free tier at the bottom of a family and a per-token price at the top are two halves of the same menu. A cut of 97% does not mean the service became almost free in absolute terms; it means the starting price was already small, and the new price is a fraction of that small number ^[1].

On that scale, the reveal that the strongest Spark model was priced at RMB 0.21 per 10,000 tokens gives a useful reference point for what a flagship model cost at that moment ^[1].

Why it matters

The significance of the day lies in the combination. A permanently free API removes the budget decision from a developer's first experiment: a student, a small team or an internal prototype can call the model and see whether it is good enough before any money is involved. Sharp cuts on paid models push the same decision in the other direction for heavier users, who care about the cost of running a product at volume rather than the cost of a test.

Taken together, the announcements show where competition among Chinese model providers had moved. Rather than arguing only about benchmark scores, the companies were publishing per-token prices and free tiers within weeks of one another, and iFlytek joined that sequence by making a free tier permanent rather than temporary ^[1]. For readers following the field, the episode is a clear record of how quickly list prices for model APIs were rewritten in that period, and of the units — tokens, thousands of tokens, tens of thousands of tokens — in which the argument was conducted.