Baidu's Token Exchange Hits 700 Billion Tokens a Day in First Month

Baidu says its new Token Exchange marketplace is already averaging more than 50 million calls and over 700 billion tokens a day, one month after launch. The company unveiled the exchange at Baidu World 2026 on November 12, 2026, alongside a wave of upgrades spanning models, agents, cloud infrastructure and chips. The scale matters because it signals early demand for a neutral channel where enterprises can compare and buy AI model services from multiple providers with verified reliability data.

What Is Token Exchange?

Token Exchange gives businesses and individual users one place to compare prices and performance, then access services from multiple model providers. Baidu says more than 100 model resources have been onboarded so far.

The verification process works in two layers:

• Entry check: Baidu compares model fingerprints and tests capability, performance and availability before a listing goes live.
• Ongoing oversight: Providers receive credit ratings based on regular spot checks and real-time, end-to-end call monitoring.

The design deliberately treats a listing as a continuing commitment rather than a one-time check. That framing matters for enterprise buyers, who routinely cite inconsistent model availability and unverifiable performance claims as barriers to adopting third-party AI services.

What's New in ERNIE 5.5 Flash Preview?

One day before Baidu World, on November 11, 2026, Baidu released ERNIE 5.5 Flash Preview. It is the first post-trained model built on ERNIE 5.5 Flash Base, with 280 billion total parameters, 14 billion activated per token and a 1-million-token context window. Baidu positions it for coding, 3D modeling, knowledge work and content creation.

The company says the Preview balances capability with inference efficiency, citing competitive benchmark scores and fast decoding. Its most striking demonstration ran inside an agent harness: over an 11-hour Unreal Engine task, the text-only model wrote a third-person game in C++ and converted playtest problems, including a hovering character and a misplaced camera, into measurable software tests, because it could not see the screen.

The API is available through Qianfan at RMB 1 per million input tokens and RMB 4 per million output tokens, with cached input at RMB 0.04 per million. Baidu plans to release the model weights under Apache 2.0 by November 25, 2026, while cautioning that this intermediate post-training checkpoint may still fall short in some capabilities as training continues.

Which Application Models Shipped Alongside It?

Two application-specific models accompanied the release:

• ERNIE Wizshot Micro-Expression combines text, images and speech for expressive video creation.
• ERNIE Light Deep Thinking handles complex search requests and tool calls by pairing a fast-thinking model that serves users with a slow-thinking model that continuously refines it.

Baidu says its Sparse Attention technology compresses historical memory during decoding, reducing inference costs without compromising performance.

How Does Robin Li's Agent Strategy Fit In?

Robin Li, Baidu's co-founder and CEO, framed the portfolio around a shift from fixed hierarchies to organizations built around tasks. His three principles: break down job boundaries, convert tacit knowledge into shared organizational assets, and create measurable closed-loop processes that support continuous learning.

Baidu's internal figures provide the adoption case:

• Average daily active agents rose from 5,000 to 37,000 over the past year.
• Weekly completed tasks increased from 84,000 to more than 1 million.

DuMate is now the unified gateway to Baidu's specialized agents, and its Harness engine and Agent API have been opened to developers. The DuMate team said it iterated more than 160 times in six months and used the agent to generate 90% of the team's code.

KOOKO, upgraded from GenFlow 3.0, combines public information, professional databases and users' own content to produce research reports, slide decks and websites. Famou 3.0 addresses production operations by combining an agent platform with forward-deployed engineering services, with Baidu reporting customers across 25 of China's 41 industrial categories.

What Did Baidu Announce in Cloud and Chips?

Baidu AI Cloud launched the Tianchi 1024-card Supernode, an all-optical interconnect system built for high-speed data transmission. Kunlunxin, Baidu's AI chip business, said its M100 inference chip has shipped, while the flagship M300 is expected in 2027 with improvements in computing power, memory and interconnection.

Together, the announcements connect Baidu's silicon and cloud infrastructure to foundation models, workflow agents and a marketplace for model services, tightening the integration across the full stack.

Why Does the Full-Stack Play Matter?

Token Exchange extends that stack beyond Baidu's own products, opening the verification and credit-rating layer to outside providers. The move positions Baidu not just as a model vendor but as an infrastructure gatekeeper for how Chinese enterprises select among competing AI services.

The open question, as the source notes, is whether Token Exchange's verification and credit ratings can turn early usage into a trusted channel for enterprises choosing among model providers.