Baidu's Token Exchange Turns AI Model Shopping Into a Price War

Baidu has launched Token Exchange , a marketplace where customers can compare and buy access to AI models from multiple providers. CEO Robin Li describes it as an exchange for foundation models, with pricing, performance data, verification checks, and provider records collected in one catalog.

In this market, a token is a unit of model input or output, usually part of a word or character. Standardizing access around token consumption gives buyers a common pricing unit, while providers gain a channel for selling inference capacity. The exchange arrives as Chinese model vendors cut API prices and compete on latency, reliability, context length, and model quality.

The launch in five numbers

One month after launch, Baidu reported the following operating figures and controls:

Those figures come from Baidu, and no independent audit accompanies them. The published data also omit unique customer counts, failed-call rates, input and output token splits, and traffic concentration by model. The 700 billion-token figure establishes scale, but those missing details matter when assessing sustained demand and reliability.

A fingerprint against model swapping

Token Exchange addresses a recurring problem in model resale: a provider can advertise one model while quietly routing requests to a cheaper or smaller substitute. Customers may notice weaker answers or changed latency, yet proving the substitution is difficult without access to the underlying weights and infrastructure.

Behavioral fingerprinting sends known prompts to a model and compares its response patterns with a reference profile. Capability tests can check reasoning, coding, language support, context handling, and other advertised features, while real-time monitoring tracks availability and performance after listing.

Fingerprinting provides probabilistic evidence about a model’s identity. Stronger assurance also requires repeated tests, transparent thresholds, version tracking, and a process for investigating disputes. Baidu’s public provider records could reinforce those controls if they disclose outages, failed checks, substitutions, and enforcement actions consistently.

ERNIE puts a price on the strategy

Baidu launched the exchange alongside ERNIE 5.5 Flash Preview , giving buyers a low-cost first-party option within the broader catalog. Baidu lists the following specifications:

• Architecture: A sparse mixture-of-experts model with 280 billion total parameters.
• Active capacity: 14 billion parameters used for each token.
• Context window: Up to 1 million tokens.
• Input price: RMB 1 per million tokens, approximately US$0.14.
• Output price: RMB 4 per million tokens, approximately US$0.56.

A mixture-of-experts model activates only part of its total network for each token, which can reduce inference costs while preserving a larger pool of specialized parameters. The one-million-token context window targets workloads involving large codebases, long documents, and extended conversations, although usable context depends on retrieval quality, latency, and accuracy across long inputs.

Baidu says it will publish the model weights by November 25, extending its recent move toward open-weight ERNIE releases. Pairing a competitively priced model with a multi-provider marketplace gives Baidu two routes to inference revenue: direct ERNIE usage and transactions involving third-party models.

What changes for API buyers

• Price discovery: Side-by-side listings make token prices and measured capabilities easier to compare.
• Lower switching costs: A shared integration can simplify routing, fallback, and model evaluation across providers.
• Quality pressure: Fingerprint checks and public provider records raise the cost of degraded service or undisclosed substitutions.
• Capacity sales: Providers with spare GPU capacity gain another route to customers.
• Cloud competition: Alibaba, Tencent, and smaller vendors must decide whether to participate in Baidu’s venue or develop competing catalogs.

A common endpoint can simplify multi-model applications, but model behavior remains uneven. Context limits, tokenizer rules, tool-call formats, safety filters, multimodal inputs, and structured-output support can still require model-specific code and tests.

Questions developers still need answered

The launch information leaves several implementation details open, and each can affect production costs or migration effort:

• API compatibility: Supported request schemas, SDKs, streaming behavior, tool calls, embeddings, batch jobs, and error formats.
• Billing rules: Tokenizer choice, cached-input pricing, reasoning tokens, retries, failed requests, and minimum charges.
• Routing: Whether customers can pin model versions, select providers, configure fallbacks, and prevent automatic substitutions.
• Reliability: Rate limits, latency percentiles, regional availability, service-level agreements, and incident reporting.
• Data governance: Prompt logging, retention periods, training use, encryption, access controls, and data residency.
• Verification: Fingerprint methodology, test frequency, benchmark disclosure, appeal procedures, and penalties for failed checks.

Baidu must earn marketplace trust

Because Baidu operates Token Exchange while selling ERNIE on it, marketplace governance will shape adoption. Developers and competing providers will watch search rankings, benchmark selection, sponsored placement, access to usage data, and the handling of incidents involving Baidu’s own models.

Consistent rules would make Token Exchange useful infrastructure for teams buying inference in China. Preferential treatment would narrow its role to another Baidu distribution channel. Transparent measurements, versioned test results, clear billing, and equal enforcement will determine which path the platform takes.