Baidu’s Token Exchange lets buyers compare prices for the same AI model

Baidu formally launched Token Exchange at its November 12 Baidu World event in Beijing, giving developers and businesses a marketplace for buying access to AI models. Customers can compare prices, response times and request success rates across companies offering the same model, then call the service they choose through a common API.

Customers pay for a provider to run the model and return a result. The company that built the model and the company that handles the request can be different. The marketplace lists each combination of provider, region and service terms as a separate offering. Its documentation says buyers can narrow the list by service region and capabilities such as tool calling, structured output and support for long inputs.

Four suppliers for the same GLM model

On November 12, the Guangzhou listing for GLM-5.3-Flash showed four suppliers, including the model’s developer, Zhipu. Zhipu listed prices of 0.8 yuan per million input tokens and 2.8 yuan per million output tokens. The lowest-priced supplier on that page listed 0.4 yuan and 1.4 yuan, respectively. The page says the listed prices include tax.

Next to the prices, the marketplace showed how long each provider took to return the first token, its output speed, its seven-day success rate and its request volume over that period. Zhipu’s listing showed a 99.32 percent success rate across about 253,000 calls, compared with 89.04 percent across 219 calls for the lowest-priced provider. These were marketplace figures based on different numbers of requests. Baidu’s terms say the price in effect when a request runs determines the charge.

A single API can split requests among suppliers

Baidu AI Cloud’s onboarding page lists phone registration, identity verification and a funded balance. The developer portal manages API keys separately by service region. Customers that configure multiple offerings for the same model can give them different priorities. When suppliers share a priority, customers can assign weights to divide requests among them.

The developer portal also breaks out requests and costs by supplier. Its documentation says an individual request record includes the offering used, the time and duration, the input and output token counts and the charge. Teams can export the records and check the past 24 hours of availability for services they use. The terms give customers 30 days from the date a bill is generated to dispute it.

Baidu’s terms say the platform forwards request contents to the provider that runs the model. The exchange says it retains usage details such as request times and token counts for billing and troubleshooting. The provider’s policy governs whether it retains the prompts and responses or uses them for training. The terms say the exchange discloses those policies on model detail pages where it can.

How Baidu rates model suppliers

Baidu says it compares model fingerprints during screening to check whether suppliers deliver the advertised model and version. It also tests performance and availability. Its rating guide describes recurring checks across a supplier’s listings, including whether requests succeed consistently, whether a provider meets its promised capacity for simultaneous requests and how the model performs on representative tasks.

The guide sets out five rating tiers. The two lowest identify suppliers under observation or with frozen ratings. The latter can result from serious concerns about model identity or failure to meet service commitments and is subject to review. The ratings apply to suppliers under Baidu’s own criteria, alongside separate rankings for individual models.

Baidu said that after about a month of operation the exchange had more than 100 offerings and averaged over 50 million calls and 700 billion tokens a day. An offering can represent one of several suppliers or regions for a single model.