A developer finds a cheap service for a popular model. The price looks right and the documentation looks official. Two weeks later the answers get worse, and nobody can say which model is actually answering.
Baidu thinks that is a business problem worth solving. At Baidu World 2026 on November 12, the company launched Token Exchange (词元所), a marketplace where businesses can compare and buy AI model services from multiple suppliers. Founder and chief executive Robin Li described it as an exchange for large models (大模型).
Key takeaways
- Token Exchange went live at Baidu World 2026 on November 12, 2026. It lets a buyer compare prices, performance and availability, then call services from different providers through one interface.
- Listings are checked with model fingerprints, capability tests and performance and availability tests. Providers receive credit ratings from regular spot checks and real-time monitoring of calls.
- Baidu says more than 100 model resources are listed and the platform averages more than 50 million calls and over 700 billion tokens a day. Every one of those figures is company-reported.
- The marketplace includes services from other providers, not just Baidu's own models. The company is trying to earn revenue as a trusted middle layer even when the winning model is not its own.
- The launch came alongside ERNIE 5.5 Flash Preview, released a day earlier with 280 billion total parameters and a one-million-token context window.
What the exchange actually does
Think of it as a shop window for model services. A supplier lists what it can serve: which model, at what price, from which region, with what uptime. A buyer sees those listings side by side, compares them, and calls whichever fits through a single connection.
The important detail is what Baidu says it does before a listing goes up. It compares model fingerprints and runs capability, performance and availability tests. After that, providers are not left alone: Baidu says it conducts regular spot checks and monitors calls end to end, then publishes a credit rating. Trust is meant to be continuous, not a one-time badge.
That design fits a specific problem. In a market where dozens of models claim similar capabilities and prices change quickly, a buyer cannot easily tell whether a cheap service really delivers the model advertised, or whether it will still be there next month. An independent-looking scorecard is useful even if the scorekeeper has its own interests.
Why verification is the product
Baidu already sells its own models. Launching a marketplace that also lists competitors' services looks, at first, like helping rivals.
The logic is that the middle layer may be more durable than any single model. Models change hands and fall out of favour; a place where buyers and suppliers meet can keep charging attention and trust fees regardless of which model wins this quarter. If the verification is good enough, Baidu becomes the place enterprises go when they are unsure — and uncertainty is common in this market.
There is a second effect. Putting rival offers next to Baidu's own forces a fair comparison on price and performance. That is uncomfortable if your model is expensive, but useful if it is competitive. It also gives Baidu a large, continuous stream of data about what buyers actually choose and what breaks in production.
For developers, the practical gain is switching cost. Connect once and you can try several suppliers without rebuilding the integration each time. Whether that works as smoothly as described is a question the company's own figures cannot answer.
The launch around the launch
Token Exchange was the headline, but it was not the only announcement. Baidu also released ERNIE 5.5 Flash Preview, the first post-trained model built on its earlier Flash base, with 280 billion total parameters, 14 billion activated per token and a one-million-token context window. It also upgraded several agents, announced a 1,024-card cloud supernode and gave an updated chip roadmap.
Those items matter because they show the strategy is full-stack: chips and cloud underneath, models and agents on top, and now a marketplace in the middle. But they also make it harder to judge the exchange on its own. A marketplace launched with a large party around it gets attention whether or not the verification layer holds up.
Baidu also used the event to describe an "intelligent organization" idea, saying its own daily active agents grew from 5,000 to 37,000 over a year. That is a company narrative about internal adoption, not evidence about the exchange, and I have kept the two separate.
Honest limitations
Nearly every number in this article — the 100-plus resources, 50 million calls, 700 billion tokens and agent growth — is reported by Baidu. I have not audited the platform or tested the fingerprint checks. Independent reviews of Token Exchange are not yet available.
The description of verification comes from Baidu and from technology outlets summarising its press materials. No third party has published results showing how often the checks catch a mislabeled model or a degraded service. Until that exists, "verified" is a company claim.
I have left out the API prices quoted at the launch. They are small per-token amounts in local currency, and repeating them without a like-for-like comparison would add noise rather than clarity. Competitors' marketplaces are not compared here either, because a fair comparison would require usage data that is not public.
What readers can do now
- Treat a rating as a starting point. A credit score from the marketplace owner is useful, but check the underlying uptime and latency figures against your own workload before committing.
- Keep the switching path open. If you adopt a multi-provider setup, make sure you can actually change suppliers without rewriting your whole integration. Test that before you need it.
- Watch who owns the scorecard. A marketplace run by one supplier has incentives that a neutral exchange would not. That does not make it useless — it means you should read how disputes and failures are handled.
Baidu is betting that in a crowded model market, the scarce thing is not another model but a reliable way to choose one. That is a clever position to want. Whether buyers will hand the choosing to the seller of one of the options is the part that will take time to answer.
