Overview
- Name: Token Exchange (词元所). In Chinese search and news coverage it is also called 百度词元所.
- Operator: Baidu Intelligent Cloud. The platform service agreement names Beijing Baidu Netcom Science Technology Co., Ltd. (北京百度网讯科技有限公司) as the contracting party.
- Official website: https://cysuo.com/ (English UI: https://cysuo.com/en; Baidu Intelligent Cloud product page: https://cloud.baidu.com/product-s/token_exchange.html)
- What it is: An open, trusted model marketplace that matches model providers with model consumers
Token Exchange is an open, trusted model-service platform operated by Baidu Intelligent Cloud for developers, enterprise buyers, and model suppliers. Individuals who complete identity verification can also register. According to the product overview and the official homepage, it brings together models and provider resources in one place for selection, unified API access, transactions, and usage management, and it uses resource and provider credit evaluations plus blockchain attestation to make supply easier to check. “Open and trusted model service platform” (开放可信模型服务平台) is the wording in the official product overview. The official English name is Token Exchange (the Baike entry writes “token exchange” in lowercase); the official Chinese name is 词元所. The product was formally launched at Baidu World 2026 on 12 November 2026, where Robin Li called it “an exchange for large models” (大模型的交易所).

It is important to distinguish a model from a model resource. The service agreement states that the same model can be offered by multiple providers, and that those resources may differ in region, price, performance, concurrency, availability, data-retention policy, and whether prompts are used for training. What you purchase is a provider's inference service under stated conditions, not ownership of the model.
Token Exchange is operated by Baidu, but it is not a storefront that sells only Baidu's own models, and it is not simply another name for other Baidu Intelligent Cloud model services such as Qianfan (千帆). ERNIE (文心) appears as one of the models listed on the marketplace, alongside third-party models such as DeepSeek, GLM, Kimi, and MiniMax. Listed providers include both model makers and third-party inference services. The platform also accepts provider applications. It is not the same product as “词元无忧” or other services whose names only share the word 词元. This page covers only the Token Exchange service at cysuo.com.
Key uses
- Compare models in the marketplace: On the model market you can filter by model type, service region, capabilities, and context length, then compare price, time to first token, 7-day success rate, and call volume across providers of the same model. The docs list text, image, video, and audio types. On 14 November 2026 the market filter showed “40 models,” with text and video visible as type filters and Beijing and Guangzhou as regions. Check the live catalog for the current list.
- Use rankings as a reference: The rankings page shows call volume, model share, and speed and success rate, with windows for the last 24 hours, 7 days, and 30 days. The figures come from the platform's own traffic and usage data. They can inform a shortlist, but they are not a substitute for your own evaluation.
- Review providers and credit ratings: The providers page lets you filter by type, region, status, and credit grade, and review each provider's models, resource count, and recent call volume and success rate. Credit grades are Excellent (优信), Good, Trustworthy, Watch, and Frozen, based on authenticity, stability, performance/capacity delivery, and capability delivery. The platform says a credit grade is not the same as a model's capability ranking and is not a guarantee of availability.
- Call models through one API: In the Developer Center you create an API key and attach models and provider resources. Docs say API keys are managed separately by region. The API reference supports OpenAI-, Anthropic-, and Gemini-style protocols. Chat and Images use the Guangzhou synchronous base URL `https://api.cysuo.com`; Videos use the Beijing asynchronous base URL `https://vod.bj.baidubce.com`. Chat Completions, for example, is `https://api.cysuo.com/v1/chat/completions`, authenticated with `Authorization: Bearer YOUR_API_KEY`.
- Configure routing and failover: For one model, a single API key can include multiple provider resources with priority and weight. Traffic is split by weight at the same priority and follows the configured order across priorities. The service agreement recommends a multi-resource fallback plan on the application side.
- Track usage, bills, and balance: You can filter by time, model, and API key to see call counts, token usage, cost, and failure rate, and you can export request details. The funds account shows available balance, top-ups, and spending. You need a sufficient balance before calling models.
- List supply on the platform: Model makers and inference providers can apply through “Become a provider.” Official docs describe admission tests, periodic probing, model-fingerprint checks, and blockchain attestation of key evaluation results.
Who it is for
- Teams choosing among several models: If you need to compare DeepSeek, GLM, Kimi, ERNIE, and different providers of the same model on price, latency, and success rate, the marketplace is built for that comparison.
- Teams building agents or apps who want fewer vendor integrations: Official copy emphasizes one API integration, after which you can switch models or provider resources in the console instead of maintaining a separate adapter for each vendor.
- Buyers who care about both price and delivery records: You can combine resource details, live operating metrics, and provider credit grades, and you should still read each provider's data policy on the model detail page.
- Model or compute suppliers: Providers can apply to list inference APIs on a public market where buyers can compare offers.
- When comparing similar gateways: For a UCloud-operated aggregation gateway, see AstraFlow on this site. For a broader overseas catalog and cross-cloud routing, see OpenRouter. Token Exchange is operated by Baidu Intelligent Cloud, and its catalog, accounts, and data policies are not the same. If you only want to chat with one model in a browser, that model's official chat product is usually simpler than going through a trading platform.
Pricing and plans
Token Exchange charges for actual usage. There is no separate public page of fixed subscription tiers. Prices vary by model, provider, and billing unit, and they can change. The points below are limited to rules that can be checked in official docs and the service agreement. Always confirm the live model market resource page and your Developer Center bill.
- Billing units: The service agreement says billing dimensions include tokens, request counts, images, audio, and video, subject to each resource's detail page, API responses, and platform billing notices. On the market, text models are commonly priced in yuan per million tokens (`/M`); some video models show yuan per second (`/s`).
- Top up before you call: The homepage's three-step guide is: register with a phone number, complete identity verification, and top up; then pick a model resource; then create an API key and start calling. If the balance is insufficient, the agreement allows the platform to suspend calls. Public docs and the agreement do not describe a standing free quota or trial plan. Any promotional credits should be checked in the signed-in funds account.
- The price at execution time applies: The agreement says resource prices may change when the model maker or provider adjusts them, and that call fees are based on the price in effect when the request is actually executed. When you compare cost, look at input price, output price, or other units — not only the single “lowest available price” figure.
- Bills and invoices: You can download call-detail bills in the console. Billing disputes must be submitted in writing within 30 days of bill generation; otherwise the bill is treated as accepted. The platform issues invoices to eligible users in accordance with the law; users are responsible for applicable taxes.
- Billing when a connection drops: The official FAQ Billing for interrupted HTTP connections explains several cases for synchronous chat APIs. For example, if the client disconnects or times out before the request is sent to the model, there is no charge; if the model returns a clear error, there is usually no charge unless the error includes usage; if a streaming response is cut off, usage is billed or estimated from the content already returned or from usage reported by the model. The same page documents estimation rules, such as counting text with the `o200k_base` tokenizer and using fixed token conversions for images, audio, video, and files. Billing for video and other APIs follows the relevant docs and resource pages.
- Starting prices shown on the homepage on 14 November 2026 (subject to change): deepseek-v4.1-flash, from ¥1.3 input and ¥1.56 output per million tokens; glm-5.3-flash, from ¥0.52 input and ¥1.82 output; kimi-k3, from ¥13.68 input and ¥68.42 output; ERNIE-5.1, from ¥4 input and ¥18 output. These are homepage “from” prices and may not match the market's “lowest available price” or provider counts. Use the current detail page before you spend.
Tips for best results
- Choose resources before you create a key: Compare price, time to first token, 7-day success rate, and provider credit in the marketplace, pick a primary and a backup resource, then create a region-specific API key in the Developer Center and finish the resource configuration.
- Do not choose on the lowest price alone: Providers of the same model can differ substantially in price, latency, and success rate. The agreement also says the platform does not promise that a third-party resource will remain available; if you pin a single resource, an outage there will fail the request.
- Keep regional keys separate: Chat and Images use the Guangzhou synchronous endpoint; Videos use the Beijing asynchronous endpoint. API keys are managed by region. Do not assume one key and one base URL cover every modality.
- Read the data policy before you integrate: Requests are forwarded through the platform gateway to the provider that actually runs inference. The agreement says the platform will not use prompts or outputs to train its own models without the user's explicit permission, but providers may have different retention and training policies. If your business cannot allow training on inputs, select only resources whose detail pages clearly say so.
- Protect keys and plan for failover: The agreement forbids selling, renting, or publishing API keys, including in public repositories. If a key leaks, disable or rotate it in the console immediately. For production traffic, configure more than one resource and keep the `X-Vod-Task-Id` response header for troubleshooting.
Limitations
- The platform is a matchmaker, not the model vendor: Except where a platform-owned model is clearly labeled, the agreement says the platform does not build third-party models and does not take primary responsibility for their capabilities, outputs, or data policies. It is responsible for platform-layer services such as its gateway and authentication. For third-party outages, network issues, model degradation, or a provider taking a model down, the agreement says the platform does not bear corresponding breach-of-contract liability. Total compensation is capped at the service fees the user actually paid in the previous 12 months.
- Prices and the catalog change: Homepage starting prices, “lowest available price,” provider counts, and model counts are snapshots. On 14 November 2026 the homepage showed 200B+ daily tokens, 60M+ daily calls, and 70+ model resources, while the market listed 40 models the same day (Baike says “40+,” which matches that day’s market count). Figures reported from Baidu World 2026 (12 November 2026), and the trial-run figures repeated on Baike (100+ resources, 50 million daily calls, about 700 billion daily tokens), do not match those homepage numbers exactly. These are platform-reported statistics or press restatements, not a durable proof of scale.
- Public pages do not show a full rate card or platform take rate: Unit prices appear on resource pages, but official public docs do not spell out whether the platform also charges a markup, commission, or service fee. Complete top-up methods and refund terms were not visible on the unsigned-in public pages; rely on the funds account and current notices.
- Similar names are easy to mix up: Autocomplete for “词元所官网” and “词元所 API” can also surface “词元无忧,” “词元锁,” the linguistic term 词元, and unrelated tools such as Baidu word-cloud generators. The English name Token Exchange is also generic. To find this product, add “Baidu” or open cysuo.com directly.
- The product is still new: Official docs were updated on 11 November 2026. Press coverage around 5 November 2026 already described the site as open for registration, and Robin Li formally announced it at Baidu World on 12 November 2026. The Baike entry lists the launch date as 12 November 2026, and some reports and the Baike article also call the pre-launch period a trial or “试推版.” The official homepage does not use those names. Resources, models, and rules may continue to change.
Frequently asked questions
What is Token Exchange, and is it a Baidu product?
Yes. Token Exchange (词元所) is an open, trusted model marketplace operated by Baidu Intelligent Cloud. The official site is https://cysuo.com/. The contracting party in the service agreement is Beijing Baidu Netcom Science Technology Co., Ltd. It was formally launched at Baidu World on 12 November 2026. It places different models, and multiple providers of the same model, in one market so you can compare price and service quality and call them through a unified API.
What is the official site, and how do I sign in?
The official site is https://cysuo.com/. The English UI is https://cysuo.com/en. Sign-in redirects to the Baidu Intelligent Cloud account system (`login.bce.baidu.com`). The homepage says you register with a phone number, complete identity verification, top up, and then create an API key. The site's “Contact us” link currently points to a Baidu Intelligent Cloud survey form.
How does Token Exchange relate to Qianfan and ERNIE?
ERNIE (文心) is Baidu's own model family and is listed on the Token Exchange market; the homepage has shown ERNIE-5.1. Qianfan (千帆) is Baidu Intelligent Cloud's existing model-service offering. Official Token Exchange docs do not describe the product as a rename of Qianfan. Press coverage of Baidu World 2026 quoted Shen Dou as saying Baidu Intelligent Cloud offers two kinds of model service: MaaS, where Baidu itself is the model service provider, and Token Exchange, positioned as a neutral, trusted trading platform. Public docs do not fully explain how accounts, billing, and catalogs are connected, so follow the official sites' current descriptions.
How does pricing work, and is there a free quota?
You pay for actual usage. Common dimensions are tokens, but billing may also use request counts, images, audio, or video, depending on the resource page. Text models are often priced per million tokens; some video models are priced per second. The official flow is to register, complete identity verification, top up, and then call. Public docs do not state a standing free quota; check your signed-in funds account for any promotional credits. Prices can change, and fees are calculated at the price in effect when the request is executed.
How do I call the API, and can I use an OpenAI-style client?
Yes, follow the official API guide. Create a region-specific API key in the Developer Center and send `Authorization: Bearer YOUR_API_KEY`. The Chat and Images base URL is `https://api.cysuo.com`, for example `/v1/chat/completions`. Video APIs use `https://vod.bj.baidubce.com` in Beijing. The error-handling docs say upstream model errors are returned in the OpenAI, Anthropic, or Gemini protocol you are calling. Use the API reference for each path and parameter list.
Is Token Exchange the same as “词元无忧”?
No. This page covers only Baidu Intelligent Cloud's Token Exchange at cysuo.com. Search results may also show third-party API services with similar names, such as 词元无忧, or unrelated uses of 词元 as a linguistic term, or tools such as Baidu word clouds. Those are not this product.
Will my prompts be used for training?
For the platform itself, the service agreement says Token Exchange will not use prompts or outputs to train its own models without the user's explicit permission. It stores call metadata such as time, token usage, model identifier, latency, and status for billing, reporting, and troubleshooting. Requests are forwarded to the provider that runs inference, and providers may differ in retention and training policy. Check each model resource's detail page.
