Baidu’s ERNIE-5.5-Flash-Preview lands on OpenRouter at $0.28 per million input tokens

OpenRouter has added Baidu Qianfan: ERNIE-5.5-Flash-Preview to its catalog. The listing prices the text model at $0.28 per million input tokens and $1.13 per million output tokens. It gives no context length, release notes or benchmark results.

What the listing shows and what it leaves out

The model is registered under the identifier baidu/ernie-5.5-flash-preview and is attributed to Baidu Qianfan, the company’s model platform. The page documents a single interface, the Chat Completions endpoint, with text messages as input. It describes no image, audio or video support.

Developers can pass stream, max_tokens, temperature, top_p, tools, tool_choice and response_format. The list also includes reasoning and reasoning_effort, which suggests the model can run in a reasoning mode. The page does not explain how that mode works. Streaming responses arrive as server-sent events, and, according to the page, failed generations that return a 502 error are not billed.

One limit matters for European customers: the page says in-region routing for the EU is not currently available for this model. The listing also mentions an additional price but leaves it blank and points to the model page for the full list, so the $0.28 and $1.13 figures may not be the whole price sheet.

Several basics are missing. The page states no context length, no maximum output, no release date and no model description. It does not say whether the model is open or proprietary, or how it was trained. Baidu had not published a matching announcement that we could find, so everything above comes from the listing alone.

A name that echoes an open-weight release

The name invites a comparison with another recent release. On November 5, Baidu published ERNIE-5.5-Flash-Base, according to Let’s Data Science. It is a sparse mixture-of-experts model with 280 billion parameters, 14 billion of them active per token, trained on 33.5 trillion tokens. It is a pre-training checkpoint with no instruction tuning, and a post-trained version is expected without a stated date.

Neither the OpenRouter page nor the coverage we read says the hosted Preview is that model or a post-trained version of it. The shared name is a reason to watch, not evidence of a link. The size, active-parameter count and context figures belong to the Base model and should not be applied to the listing.

Priced against the cheapest new arrivals

The analysis here is about cost. At $0.28 for input and $1.13 for output, the output price is about four times the input price, a typical split for text models that bill generated tokens at a premium. Reasoning-capable models can also spend many tokens on intermediate thinking. If the preview does that, the output price will drive real bills more than the input price suggests. The listing gives no information on how many tokens its reasoning mode uses.

Compared with other new entries, the price is not the lowest. Our earlier report on Meta’s Muse Ember on OpenRouter put that model at $0.10 per million input tokens, so the Baidu preview’s input price is 2.8 times higher. Against that, Muse Ember’s price comes without any promise about quality, and the same is true here: the page offers no benchmark that would let a developer judge whether the extra cost buys anything.

Chinese labs have used OpenRouter for early exposure before. We reported that Tencent’s Hy4 appeared on the platform at half the preview’s price, a pattern where a preview listing is followed by a cheaper general release. Whether Baidu will do the same is not stated anywhere. The word “Preview” in the name suggests that terms, limits and prices may still change, though the listing itself makes no commitment either way.

Who can try it now

Anyone with an OpenRouter account and API key can call the model through the standard endpoint at https://openrouter.ai/api/v1/chat/completions, using the identifier above. Because OpenRouter exposes a common interface, switching an existing integration to this model means changing the model string. That makes it cheap to test against a current workhorse model on real prompts, which is the only comparison available until Baidu publishes figures.

Developers in the EU should note the missing in-region routing. Teams with data residency requirements will have to look elsewhere for now.

What the source does not establish is considerable: who runs the endpoint, how the model scores on any test, how long its context window is, whether it is multimodal, and when it will leave preview. Baidu’s own claims, once it makes them, will also need independent checking. Until then, the listing is a price and a parameter list, and the model’s real strengths remain unknown.