Baidu announced ERNIE 5.5 Flash Preview on November 11, 2026, opening hosted access for software teams, game developers and technical creators. Its Qianfan platform offers the model now, while downloadable Preview weights—the files needed to run it on your own infrastructure—remain forthcoming.
Baidu describes Preview as the first version of Flash Base with additional training for assistant-style tasks. It also says that training is unfinished, so this is not a finished release.
Access now, downloads later
The Qianfan catalogue lists a one-million-token context window, meaning the amount of material the model can work with in a request. Tokens are small units of text, not whole documents. The setup guide requires signing in and creating an appropriately permissioned API key.
An officially linked free OpenRouter variant currently lists zero input and output token prices. OpenRouter’s general free-model limits normally allow 20 requests per minute and 50 per day, with a higher daily allowance after qualifying credit purchases. Check your account’s allowance: these are requests, not completed projects, and a tool-driven task may need many calls.
For self-hosting, Baidu’s Chinese announcement promises Preview weights before November 25; its English announcement promises Apache 2.0 licensing. Both are future commitments. The already downloadable Flash Base checkpoint predates assistant-style training and is not the Preview described here.
Promotional and regular prices
Qianfan’s price table quotes yuan per thousand tokens. Converted to Chinese yuan (CNY) per million tokens, the November 11 rates are:
Rate | Uncached input | Output |
|---|---|---|
Promotional, peak | CNY 1 | CNY 4 |
Promotional, off-peak | CNY 0.50 | CNY 2 |
Regular, peak | CNY 2 | CNY 8 |
Regular, off-peak | CNY 1 | CNY 4 |
Cache-hit input has separate, lower rates. The Preview price rows specify neither a promotion end date nor a timezone for the peak/off-peak boundaries; confirm those before scheduling work around a discount.
Illustrative calculation: across one or more requests, one million uncached input tokens plus 100,000 billed output tokens would cost CNY 1.40 at promotional peak rates: 1 × CNY 1 + 0.1 × CNY 4. At regular peak rates, the same usage would cost CNY 2.80. This is token arithmetic, not a project quote; it excludes retries, tools, rendering, assets, taxes and human review.
Why the demos need more than the model
ERNIE 5.5 Flash Preview is a text-only model. In Baidu’s airport-and-airliner demonstration, it worked through agent software and Blender’s Python interface, with external assets, GPU rendering and FFmpeg video assembly. Those are company-reported results, not evidence that a standalone chat can reproduce the finished scene.
For a small studio, the practical distinction is between obtaining the model and assembling the production workflow. The creative software, assets and integration work remain separate requirements.
There is also a service-specific constraint: Qianfan’s Preview settings support tool calls but keep reasoning switched on at a fixed high setting. You cannot select a non-reasoning mode there. Do not assume another provider’s controls behave identically.
A sensible first step is a bounded hosted trial with a defined output and a budget for the whole workflow. RohitAI’s guide to coding-agent teams explains how to divide work and review the combined result. If local execution is essential, wait for the actual Preview files, license and deployment requirements before committing infrastructure.