ERNIE 5.5 Flash Preview: Two Prices, a Free Endpoint, and No Rank to Show For It

Baidu's pitch for ERNIE 5.5 Flash Preview, posted to X on 13 November 2026, is that the model "has range": playable worlds, 3D models, research reports, interactive workbooks. The pitch underneath it is "open-weight foundation model." Three days after the 11 November announcement, the things a buyer would check first do not line up with either word. The model has two published paid rate cards that differ by roughly half, a free hosted variant that predates the launch by a day, no independent benchmark score, and no rank on the arena roster that carries its name. And the weights that are actually downloadable belong to ERNIE 5.5 Flash Base, not to the model the demos were made with.

None of that makes it a bad model. It makes it a preview whose marketing has outrun its paperwork, which is normal for a three-day-old release and worth knowing before you point a pipeline at it.

What Baidu actually shipped on 11 November

The ERNIE team's own product-updates page is dated 11 November 2026 and describes ERNIE 5.5 Flash as a 280B-parameter mixture-of-experts model with 14B active parameters and a context window of 1,024,000 tokens. The page calls the release "an intermediate checkpoint from post-training" — a phrase that is doing a lot of work, since it means the model is explicitly not the finished article. The weights, it says, "will be open-sourced under Apache 2.0 soon."

That page is where the demo reel comes from, and the demos are the reason the model got attention. A playable Unreal Engine 5.8 build produced in a single 11-hour task. Roughly 100 mini-games across Cocos, Unity, LayaAir, PixiJS and Canvas 2D. A Blender scene of a 747 and Daxing airport driven by two cooperating agents. A FreeCAD W16 engine, Shenzhen towers and a lunar rover. A 20-page report on AI power costs, a workbook with a 1,000-path Monte Carlo, 4,169 NHTSA crash reports parsed, a 6m34s orchestral WAV, and a 116-second "Pelican Test" that emits 10,000+ lines of Python across 20 art styles. Baidu also cites 300,000 workspaces, 10 million reference files and more than 40 supported file types as the environment these ran in.

The part readers get wrong: it does not see

ERNIE 5.5 Flash Preview renders 3D models and game worlds, so the natural assumption is that it is a multimodal model that outputs images. It is not. The arena roster that carries the model's row records its image input capability as false, and Baidu's own write-ups describe every visual artefact as code: Blender scripts, FreeCAD scripts, Pycairo and Pillow drawing calls, Unity projects. The model writes the program that produces the picture. That distinction matters for how you budget it, because a text-only model doing 3D work burns tokens on source code rather than on image patches, and it fails in a different way — a broken script rather than a wrong pixel.

The 116-second Pelican Test is the cleanest example of what the model is actually good at. Ten thousand lines of Python, twenty art styles, twenty-two audio recordings, all in under two minutes; Baidu's own framing is that the model converts playtest bugs into code tests. That is a coding-agent story with a rendering flourish on top, not an image-generation story.

Three price tags, one model

Baidu Qianfan's rate card for ERNIE 5.5 Flash Preview, read on the Chinese-language product page this week, lists ¥0.001 per thousand input tokens during the 08:00–22:00 window, ¥0.0001 per thousand cached input tokens, and ¥0.004 per thousand output tokens. Per million tokens that is ¥1 in, ¥0.1 cached, ¥4 out. At the 13 November reference rate of 6.685 CNY to the dollar, that converts to roughly $0.15 input, $0.015 cached and $0.60 output per million tokens.

The USD listing that earlier coverage of this model used — the one most write-ups still repeat — is $0.28 input, $0.028 cached and $1.13 output. Same model, same week, published by the same vendor, about 47% apart on every line.

Then there is zero. A free variant of the same model, same 1,024,000-token context and same 921,600-token completion ceiling, appeared on a third-party model catalogue on 10 November — a day before Baidu's own release page, and it is still listed at $0 on both prompt and completion tokens. That is the endpoint Baidu's 13 November thread is pointing readers toward when it says to try the model. So the range in the tweet is real; the price is whatever the surface says it is.

Take a concrete agent run: 40 model calls, 30,000 input tokens each, 260,000 output tokens total.

• On the USD card — 1.2M input at $0.28 plus 0.26M output at $1.13 comes to about $0.63 per run.

• On the CNY card — the same tokens convert to about $0.34 per run.

• On the free variant — $0.00, whatever the rate limits turn out to be.

I am not going to call the CNY figure a price cut, because I cannot show it is new rather than the regional card that always existed. What is checkable is that the number a reader lands on depends entirely on which page they open, and that no write-up published so far reconciles them. For a chat session the gap is pennies. For an agent making hundreds of thousands of calls it is the budget line, and it is worth knowing which of the three you are actually signing up for before the invoice tells you.

Unranked is not the same as bad, but it is not a result either

The arena roster that lists ERNIE 5.5 Flash Preview has a row for it and no position for it. In the web-dev category the entry carries the roster's maximum-integer placeholder rather than a rank — the value a leaderboard uses for a model that has collected votes but has not been placed. Two other entries in the same dump carry the same placeholder, so this is a normal pre-rank state, not a penalty. It does mean that when someone tells you the model is "ranked" at anything, they are reading a number that has not been computed.

Artificial Analysis has no page for the model at all — its leaderboard lists ERNIE 4.5 300B A47B and ERNIE 5.0 Thinking Preview, and nothing from the 5.5 Flash line. The benchmark table Baidu published with the release runs twenty rows, including Terminal-Bench 2.0 at 46.5, and every one of those figures is vendor-run and unreproduced by anyone outside Baidu. The same caution applies to the 1,668 Code Arena score that has been circulating since launch, which comes with an AutoEval basis and no vote count.

So the honest scoreboard for this model, as of 14 November, has exactly one empty column: independent evidence.

The weights that exist are Base's, not Preview's

Baidu's 13 November post calls ERNIE 5.5 Flash Preview an "open-weight foundation model." What is on Hugging Face is ERNIE 5.5 Flash Base — a repository created 21 October 2026 and last modified 5 November, carrying the ERNIE 5.5 Flash Base Model License v1.0 rather than Apache 2.0. That license excludes the EU, the UK and South Korea, sets a $50M revenue threshold and requires a "Powered by ERNIE" attribution. There is no Preview weights repository in the organisation's listing, only Base and Base-BF16.

So the sentence in the tweet is true of the family and not yet true of the model in it. The Apache 2.0 promise is dated to "soon" and has not landed. Anyone planning to self-host ERNIE 5.5 Flash Preview today has nothing to download; anyone planning to self-host ERNIE 5.5 Flash Base has a license to read carefully first, because the restrictions are geographic and commercial, not just attributional.

How a text-only preview with no rank gets adopted

The sensible way to use a model in this state is the way you would use any unproven endpoint: give it the workload its one clear advantage fits, and never let it be the only thing holding a production path up. That advantage is concrete — 1,024,000 tokens of input and a 921,600-token output ceiling, which is far beyond what most models will generate in a single response, plus a price that is low even on the more expensive of its two cards.

This is also where a routing layer earns its keep rather than just being convenient. OrcaRouter's catalogue does not currently include ERNIE 5.5 Flash Preview — the models list returns 215 entries and no Baidu model at all — so I am not going to write you an availability line that pretends otherwise. What the platform does give you for a preview like this one is the shape of a safe experiment: automatic failover across providers behind a single endpoint, so a preview without an uptime record never has to be a single point of failure, and 0% markup with provider list price passed through, so whichever of those two rate cards you are billed on is the one you see, the day it changes. If you build on the model through Baidu's own API and several third-party platforms and later want everything else on the same key, the routing DSL composes several models into one call without a second contract or a code change.

What would change this write-up

Three events, in the order they are likely to arrive. The Apache 2.0 weights for ERNIE 5.5 Flash landing on the Hub would settle the open-weight question and, if the license really is Apache 2.0, would make the model genuinely different from the Base checkpoint sitting there now. A first independent evaluation — an Artificial Analysis page, a ranked arena position, a reproducible Terminal-Bench run from someone who is not Baidu — would replace the vendor table with evidence. And someone reconciling the CNY and USD rate cards into a single published number would tell you what the model actually costs.

Until then, the fair summary is the one the demo reel argues against: a strong text-and-code model with unusually long output, priced cheaply on at least one card and free on another, wrapped in marketing that implies capabilities and openness it does not yet have. Use it for what it demonstrably does — write the script, generate the test, draft the report — and treat the 3D worlds, the "open-weight" label and the missing rank as claims still in transit rather than specifications you can build against. The free endpoint is the least risky way to find out whether the model suits you, precisely because nothing about it is load-bearing yet.