Baidu’s ERNIE 5.5 Flash costs RMB 1 per million input tokens, and weights are coming

Baidu said at its Baidu World 2026 event that ERNIE 5.5 Flash Preview is live on its API. The model has 280 billion total parameters, charges RMB 1 per million input tokens, and will have open weights by November 25.

The Beijing company also launched a marketplace for model services called Token Exchange and upgraded four agents. Everything below comes from Baidu’s own announcement and has not been independently verified.

A 280B model that activates only 14B

According to Baidu, ERNIE 5.5 Flash Preview uses 14 billion of its 280 billion parameters for each token. It supports a context window of 1 million tokens, and the company says it performs strongly in coding, 3D modeling, knowledge work and content creation. Baidu cites competitive scores on Terminal-Bench and AutomationBench, but the announcement gives no figures and does not name the models it was compared with.

The API can be used now on Baidu AI Cloud’s Qianfan platform. Baidu bills RMB 1 for every million tokens sent to the model and RMB 4 for every million it generates, and cache hits on input are billed at only RMB 0.04 per million. The company also says the model is free on OpenRouter, though it does not say for how long. The weights are to be open-sourced by November 25, and the licence terms have not been published.

Baidu says the model was tuned with ERNIE Evolve, its own recursive self-improvement technique, to raise inference efficiency and cut cost. It gives no detail on how the method works.

Cheap tokens meet a market for tokens

The price is the main hook here. Third-party listing sites, which are not primary sources, show the earlier ERNIE 5.0 at $1.40 input and $5.60 output per million tokens, citing Baidu’s international Qianfan page. Those are dollar figures and Baidu’s new prices are in yuan, so a direct comparison needs a current exchange rate. Even so, a preview model at a flat RMB 1 for input is pitched squarely at high-volume agent and coding workloads.

The same type of listing puts DeepSeek’s V4-Flash at $0.14 input and $0.28 output per million tokens. That means ERNIE 5.5 Flash Preview is entering a market where Chinese open-weight rivals already compete on very low prices. These listings are secondary sources, so the figures should be checked against each vendor’s pricing page. Baidu is betting that open weights, a long context window and a free OpenRouter listing will bring developers in before the preview label comes off.

Token Exchange is the other half of that bet. Baidu describes it as a neutral marketplace where customers can compare model services on price and performance in one place. It says it verifies each resource by comparing model fingerprints and testing capability, performance and availability, and rates providers through regular spot checks and real-time call monitoring.

After one month, Baidu says the exchange has more than 100 model resources and handles over 50 million calls and more than 700 billion tokens a day. It also says the exchange could put underused compute to work. The announcement does not say how Baidu earns money from the exchange, who the providers are, or how it keeps its own models neutral against third-party ones. Because Baidu both sells models and runs the marketplace, that neutrality claim is one for readers to watch.

Video models and agents get cheaper to run

Baidu also released two application models on Qianfan. ERNIE Wizshot Micro-Expression makes expressive video from text, images and speech, accepts up to 4,000 Chinese characters, and is said to lead the industry for podcasts and short animated dramas at half the usual cost. ERNIE Light Deep Thinking began in multimedia search. It pairs a fast model that answers users with a slow model that keeps refining it, and it uses Baidu Sparse Attention to compress memory during decoding.

Among the agents, Baidu Yijing now covers short animated dramas from concept to final cut at RMB 0.2 per second of video. DuMate becomes a single entry point to Baidu’s other agents, and Baidu has opened its Harness engine and Agent API. KOOKO, upgraded from GenFlow 3.0, draws on 1.8 billion professional resources and 750 million academic papers. Famou 3.0 pairs an agent platform with forward-deployed engineers, and Baidu says its customers span 25 of China’s 41 industrial categories.

CEO Robin Li framed the launches around what he called the intelligent organization. “If AI is the new electricity, then the intelligent organization is the new factory of the AI era, where productivity truly happens,” he said. Baidu says its own daily active agents rose from 5,000 to 37,000 in a year and now finish more than 1 million tasks a week.

Hardware and robotaxi updates

On infrastructure, Baidu AI Cloud launched the Tianchi 1024-card Supernode, an all-optical interconnect design. Its Kunlunxin chip unit said the M100 inference chip has shipped and that the flagship M300 is expected in 2027. Baidu also said Apollo Go robotaxis now operate in 28 cities and have covered more than 350 million autonomous kilometers, including driverless service in Abu Dhabi and Dubai.

What is still missing is independent testing. Benchmark numbers, licence terms for the open weights and the sources of the usage statistics remain unpublished, so the real test comes when the weights land and developers can run the model themselves.