Baidu Built a Stock Exchange for Models, and It Grades the Sellers

On 12 November, at its annual conference in Beijing, Baidu launched something that does not have an obvious Western equivalent: a regulated trading venue for model tokens. It is called 词元所, which translates roughly as "token exchange", and the analogy Baidu's founder used on stage was explicit. A stock exchange attracts good companies and lets investors find them. This attracts good models and lets buyers find them.
The problem it claims to solve is one that anyone who has bought API access through a reseller will recognise. Robin Li put it in plain terms: in the first nine months of this year, a new model shipped almost every day, sometimes several in a day. A cottage industry of token resellers grew up around that flood, and those resellers, in his description, take no responsibility for output quality, service stability or response time, and cannot even establish that the model behind the endpoint is the model they advertise.
That last clause is the whole business case. A reseller can route your request to a cheaper model, keep the difference, and unless the output degrades noticeably you will never find out.
Two features, both about trust
Baidu's answer has two parts, and the names are worth translating carefully because they carry the argument.
The first is verification. Suppliers are checked through model fingerprint comparison, capability testing, and performance and availability testing. Fingerprinting here means what it sounds like: probe the endpoint, compare the responses against the reference model's signature, and establish that the thing answering is the thing being sold.
The second is credit reporting. Baidu runs periodic spot checks and end-to-end monitoring of live calls, and from that builds a credit rating for each provider. Buyers see the rating. A downgrade triggers a notification.
Neither feature is technically novel. Fingerprinting a model through behavioural probes is a known technique, and monitoring uptime is what every API gateway does. What is new is packaging both as a market primitive and publishing the rating as a product attribute.
The numbers after one month
Baidu says the venue opened a month before the announcement and has more than 100 supply resources listed, with daily calls above 50 million and daily token volume above 700 billion.
Those two figures deserve a second look, because the ratio between them is odd. Fifty million calls and 700 billion tokens works out to roughly 14,000 tokens per call. That is a long average, closer to document processing or agent loops than to chat. Either the mix is dominated by batch workloads, or the call count and the token count are being measured on different bases. Baidu did not break it down on stage.
The 100 supply resources figure is also ambiguous in a useful way. It could mean 100 providers, or 100 model listings, or 100 endpoints. In a marketplace, those are very different numbers, and the difference is exactly what a buyer would want to know.
Baidu also framed the launch against a specific villain: the token relay station, the reseller that marks up a model and takes no responsibility for what comes back. That framing is doing a lot of work. It positions the exchange as the honest broker in a market the audience already distrusts, which is a stronger sales argument than any price comparison.
Why an exchange is the right shape for this problem
Model access has become a financial product. You prepay, you consume, the unit is metered, and the supplier can change the terms of the underlying good without changing the price. That is a commodity market with an information asymmetry in it, which is the historical precondition for an exchange.
Exchanges exist because bilateral trust does not scale. Once there are a hundred sellers of a thing that is hard to inspect before purchase, you need a standard contract, a settlement layer, and somebody whose job is to catch the seller who ships something other than what was graded. Baidu is proposing to be that somebody, for a cut.
The comparison to Western infrastructure is instructive. OpenRouter does something adjacent: one key, many models, live usage rankings, and a per-model price list. What it does not do publicly is verify that the model answering is the model you selected, or grade the provider. Its rankings are revealed preference, not audit.
Both approaches have a failure mode. A pure marketplace optimises for liquidity and lets buyers do their own diligence, which works until the cheap listing wins on price and the buyer cannot tell what they got. A graded exchange optimises for trust, which works until the grader's incentives drift toward the sellers who pay the most.
The part that is hard to copy
Verification at scale needs something most marketplaces do not have: a reference implementation to compare against. Baidu runs its own models, so it has fingerprints for the models it hosts and a baseline for what a given prompt should produce. That is an advantage a neutral third party would struggle to assemble without cooperation from every lab.
It is also why the model is unlikely to be replicated exactly in the US or Europe. A neutral exchange there would have to fingerprint models owned by companies with no interest in being fingerprinted, and would have to publish ratings that those companies would litigate over.
The Chinese version gets to define the reference. That is a governance choice dressed as a technical one.
What to watch
Three things, over the next two quarters.
Whether the credit ratings are public and comparable, or visible only inside the platform. A rating that buyers cannot cite in a procurement document is a loyalty programme.
Whether verification covers the hard case, which is a provider that serves the genuine model for the probe and a cheaper one for production traffic. Fingerprinting at onboarding does not catch that; continuous monitoring might.
And whether the exchange publishes its own take rate. A venue that grades sellers and also earns from their volume has the same conflict as a ratings agency paid by issuers, and the disclosure is the only thing that keeps it honest.
The pitch on stage was about choice. The product is about proof. Anyone who has been served a downgraded model by a reseller already knows which of those two they would pay for.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.