ERNIE 5.5 Flash Preview Is Callable: $0.28 In, $1.13 Out, No Model Card

Six days after Baidu published the raw weights of ERNIE 5.5 Flash Base, the model those weights were supposed to become started answering prompts. ERNIE 5.5 Flash — served as ERNIE 5.5 Flash Preview — went live as a billed endpoint on Baidu Qianfan on 10 November 2026 at $0.28 per million input tokens and $1.13 per million output, with cached input reads at $0.028. There was no launch post, no model card, no benchmark table and nothing on Baidu's own ERNIE site, which on 11 November still leads with the base checkpoint and the promise that the post-trained model is "coming soon." What exists is a working endpoint with a price, a 1,024,000-token context window, an output ceiling of 921,600 tokens, and a specification sheet that stops short of telling anyone how good the thing is. The control group is billed at provider list prices with the OrcaRouter fee shown as $0.00 which is the only reason a preview with no model card can be tested against something.

What actually went live

The endpoint is text in, text out. No image, no audio, no PDF parsing — the same modality limit as the base checkpoint, and the first thing to check before planning anything around it. Everything else on the sheet is more generous than the price suggests:

• Context — 1,024,000 input tokens, with a maximum output of 921,600. That output ceiling is unusual: most models cap generation in the low hundreds of thousands at most, and this one allows nearly a million tokens of response.

• Price — $0.28 per million input, $1.13 per million output, $0.028 per million cached input reads. Cache reads are billed at a tenth of the input rate, which is the one line on this sheet that rewards a stable system prompt.

• Reasoning — mandatory, enabled by default, and offered at a single effort level: high. There is no cheap, non-thinking mode to fall back on. Every request pays for reasoning tokens, and reasoning tokens are output tokens.

• Tooling — function calling, structured outputs and response-format constraints are all exposed, which is what separates a preview you can build against from a preview you can only chat with.

• Availability — a single provider region, with no EU in-region routing. For a European team the latency is one problem; the base model's licence, discussed below, is a different and larger one.

• Documentation — no knowledge cutoff is published, no model card, no evaluation table, no context-length caveats. The description that accompanies the endpoint is three sentences long and describes architecture, not quality.

The specification that is missing matters as much as the ones that are present. Baidu published thirty benchmark rows for the base checkpoint on 5 November. Not one of them describes this model. A base checkpoint's scores come from a different pipeline, a different evaluation setup and a different task framing than an instruction-tuned model's, and the base numbers cannot be inherited by the chat model — they are a ceiling on what post-training had to work with, not a forecast of what it produced.

The output ceiling is the interesting number

921,600 output tokens is not a typo, and it is not a marketing flourish either — it is a direct consequence of mandatory reasoning. When a model has to think before it answers and cannot be told not to, the only way to keep long agentic runs from being cut off mid-plan is to raise the output ceiling by an order of magnitude and let the model use what it needs. Baidu has done that here, and the same design decision sets the bill.

Reasoning tokens are billed at the output rate. At $1.13 per million, a request that thinks for 8,000 tokens and answers in 500 costs about a cent — trivial per call, and 8,500 tokens of output on the invoice. Multiply by an agent loop that runs forty steps and the arithmetic that matters is no longer price per million tokens but price per completed task, which is the only cost figure that survives contact with production. This is the same trap that made Gemini 3.8 Flash look cheaper than it prices out: a flat per-token rate applied to a model that emits 30 to 34 percent more output tokens per task is a rate cut that never arrives.

There is no published measurement of how verbose ERNIE 5.5 Flash Preview is. That is not a small gap. Verbosity is the single largest swing factor between two models with identical per-token prices, and it is the first thing worth measuring yourself rather than reading about — a fixed set of twenty of your own prompts, run twice, counting output tokens. Until somebody does that, any cost comparison involving this model is a comparison of rate cards, not of bills.

Priced against what, exactly

The nearest price anchor is the model Baidu shipped the base of: MiniMax M3.1 Flash, at $0.15 in and $0.60 out per million tokens, scored 42 on the Artificial Analysis Intelligence Index and cost $0.22 per completed index task. ERNIE 5.5 Flash Preview is roughly twice that price per token with no independent score at all. That is not evidence it is worse — the index does not measure Chinese-language work, and Baidu's base checkpoint was strong on repository-scale code — but it is evidence that the preview is priced as a capable model rather than as a cheap one, and that the burden of proof is on Baidu to publish something.

Set against the base checkpoint the comparison gets starker in the other direction. The weights are free; the API is $1.13 per million output tokens. Anyone whose workload is a fixed, high-volume task should be doing the arithmetic on a fine-tune of the 280B checkpoint rather than paying per token forever — except that the licence, a separate document from the API terms, excludes the European Union, the United Kingdom and South Korea entirely, caps free commercial use at $50 million in related annual revenue, and requires a "Powered by ERNIE" mark on every commercial product. Those terms apply to the weights. They do not apply to calls made through Baidu's API, which is one of the few genuine arguments for the hosted route.

How to get at it, and what to do while you wait for numbers

Access is through Baidu's own Qianfan API, or through third-party platforms that resell it. OrcaRouter does not route ERNIE 5.5 Flash Preview — there is no ERNIE entry in our catalogue of more than 200 models and no Baidu endpoint for us to route to, so anything suggesting otherwise would be wrong. What the catalogue is good for in this situation is the control group: the models this one will be judged against, MiniMax M3.1 Flash, GLM-5.3-Flash, DeepSeek V4.1 Flash and the Gemini Flash line among them, are all reachable through one OpenAI-compatible key at provider list price with no markup, and automatic failover means a preview with an unproven uptime record never has to be the only thing holding a production path up.

That is the practical shape of the next few weeks. Use the preview where its two real advantages apply — a million-token input window at $0.28, and a text-only workload where the modality limit costs nothing — and keep a routed fallback behind it. Do not migrate an existing pipeline onto it on the strength of a spec sheet, because the spec sheet is the only thing Baidu has published.

What would change the picture is unglamorous and specific: a model card with an evaluation table, an independent score from a tracker that runs its own harness, and a measurement of output tokens per task. Baidu has published the first of those for the base checkpoint and none of them for the chat model, which is an odd order to do things in — the raw material documented in detail, the finished product shipped silently. Until that changes, the honest reading of ERNIE 5.5 Flash Preview is that it exists, it is cheap enough to test, and nobody outside Baidu knows yet whether it is good.