Chinese companies buying model access now get a Baidu-run credit rating on model makers and resellers, while Baidu's own Ernie competes on the same platform with DeepSeek, Kimi and GLM.
Baidu has launched a marketplace where companies buy access to large AI models from many suppliers, with Baidu testing and rating those suppliers. Founder Robin Li presented the platform, which Baidu’s English press release calls Token Exchange (词元所), at Baidu World 2026 in Beijing on November 12. Baidu said that one month after launch the exchange had more than 100 model resources on board and handled more than 50 million calls and over 700 billion tokens a day. A token is the small unit of text a model reads or writes, and the unit in which most model usage is billed.
Li framed the problem as one of trust. As the number of models grows, prices vary, sources are mixed and quality is uneven, which makes buyers hesitate, IT Home reported from the venue. He compared the platform to a stock exchange, which draws good companies to list so that investors can find and back them, and Baidu calls it neutral, transparent and trusted.
The new element is the rating layer that sits on top of model access. Baidu has disclosed usage but no revenue, fees or rating methodology, and it has not said how it will keep scores neutral when its own Ernie models are sold on the same platform.
Baidu tests each model and gives each provider a credit rating
Before a resource is listed, the exchange runs admission tests: it verifies the model, compares its fingerprint against the original, and checks capability, performance and availability, according to IT Home. A model fingerprint is a set of telltale outputs that shows which model is actually answering a request. Monitoring continues after listing. The stated aim, in IT Home’s account, is to keep out sellers that quietly swap in a different model or overstate a model’s parameters.
Providers are scored on service quality, delivery on commitments, security and compliance, and commercial contract performance, and receive credit rating labels. When a provider’s rating drops, the exchange warns users promptly and records the data in public credit files, IT Home reported. Baidu’s release says the ratings draw on regular spot checks and real-time monitoring of calls from end to end.
A 36Kr report on November 5, a week before the conference, showed what a buyer sees. The platform then listed 42 models, including DeepSeek, Kimi and Zhipu’s GLM, behind one programming interface that accepts both OpenAI and Anthropic request formats, so a developer can change vendors without rewriting code. Each model can have several suppliers, compared on price, time to first response, throughput, seven-day success rate and call history; DeepSeek-v4.1-flash had six. 36Kr added that supplier credit records are stored on a blockchain.
The exchange puts Ernie on an open shelf beside its rivals
36Kr’s screenshot of the seven-day leaderboard showed 398 million calls, 5,133.6 billion tokens and 68 ranked model resources, with DeepSeek-v4.1-flash first by both calls and tokens. That works out to roughly 733 billion tokens a day, in line with the figure Baidu gave on stage. Suppliers included Zhipu as the original developer alongside third-party resellers, and a sign-up page invites more providers to apply.
The setup means Ernie is sold in an open market next to DeepSeek, GLM and Kimi rather than only through Baidu’s cloud, 36Kr noted. It set the exchange against Alibaba’s Bailian, Volcano Engine’s Ark and JellyToken, which it described as Alibaba-affiliated, and named three difficulties: connecting very different third-party interfaces, keeping evaluations neutral, and handling Ernie’s place on a platform Baidu runs. The ERNIE 5.5 Flash Preview promoted at the same event costs RMB1 per million input tokens and RMB4 per million output tokens on Baidu’s Qianfan platform, according to the release.
The exchange also serves Li’s broader pitch. He presented the "intelligent organization" as the next stage of AI adoption and set out three principles, TechNode reported: organize work around tasks that cross job boundaries, record employees’ know-how so people and agents can reuse it, and measure each workflow so it keeps improving. Baidu offered itself as the case study. Daily active agents inside the company, a yardstick Li promoted in May, rose from about 5,000 to 37,000 in a year, and weekly tasks completed by agents climbed from 84,000 to more than 1 million, according to IT Home’s report on the speech. Baidu has built 4,300 workspaces and stored more than 20,000 organizational skills.
Third-quarter results are the first chance to see the money
Baidu disclosed call volume for the exchange but nothing on what it earns from it, what share of each sale it keeps, or how many providers have been downgraded. Buyers and model makers will want the scoring rules in writing before they treat a Baidu-run rating as independent. The hardware shown at the event, including the Tianchi 1024-card supernode and Kunlunxin’s M100 inference chip, which has started shipping, points at the same goal of selling more tokens.
Baidu reports third-quarter results on November 16, with a call at 8:30 p.m. Beijing time, according to its results notice. A revenue figure for Token Exchange, or a count of providers that have received credit warnings, would show whether the exchange is turning into a business line of its own.