ERNIE 5.5 Flash vs ERNIE 5.5 Flash Preview: Which One Are You Actually Calling?

Baidu's own website lists two ERNIE 5.5 artefacts and neither of them is called ERNIE 5.5 Flash. There is ERNIE 5.5 Flash Base, the 280B raw checkpoint published on 5 November, and ERNIE 5.5 Flash Preview, the post-trained model published on 11 November. ERNIE 5.5 Flash — the name without a suffix, the one that appears in price lists and model pickers — is the model the Preview is a preview of. That is not a marketing quirk to be glossed over; it is the whole substance of the comparison, because everything that separates ERNIE 5.5 Flash Preview from ERNIE 5.5 Flash is a difference between a checkpoint and a release, not between two architectures. Same 280 billion total parameters, same 14 billion active, same 1M-token context. The differences are a licence, a date, a stability guarantee and a set of numbers that Baidu has told you will change. The comparison group is the half of this decision that is under your control. provider list price with 0% markup GPT-6.1 Sol, Claude Opus 5.5, GLM-5.3 and DeepSeek V4.1 Flash sit on one key at provider list price, with the OrcaRouter fee shown as $0.00.

The naming is the confusing part, so start there

Three names, two models, one release pipeline. ERNIE 5.5 Flash Base is the pretrained checkpoint: 280B total, 14B active, 43 layers, a sparse mixture-of-experts stack with a shared expert, trained on 33.5 trillion tokens, downloadable today on Hugging Face under a restricted community licence. ERNIE 5.5 Flash Preview is that checkpoint after post-training — the thing that answers instructions, uses tools, writes game code and produces slide decks. ERNIE 5.5 Flash, when it exists, is the same model after post-training finishes.

Baidu states the relationship plainly in the release post: the Preview "is an intermediate checkpoint from post-training. It is already competitive with models of its size, but because training is not finished, some capabilities may still fall short. Post-training is ongoing and ERNIE 5.5 Flash is improving quickly." Read that twice. The company is telling you that the model you can call today and the model you will be calling in a few months share a name stem, an architecture and a context window, and differ in how finished they are. If you are choosing between them, you are not choosing between two systems. You are choosing between buying now and waiting.

The comparison, dimension by dimension

Here is what actually differs, with both sides on every line:

• Identity — ERNIE 5.5 Flash Preview is a named, callable checkpoint published 11 November 2026; ERNIE 5.5 Flash is an unreleased successor with no date, no model card and no endpoint of its own.

• Architecture and size — identical on both sides: 280B total parameters, 14B active, 1M-token context, inherited from the ERNIE 5.5 Flash Base checkpoint of 5 November. Post-training does not resize a model.

• Weights and licence — the Preview is promised under Apache 2.0 "soon", with no repository on Hugging Face as of 11 November; the eventual ERNIE 5.5 Flash will be governed by whatever licence Baidu ships at official release, which the company has not stated and which is not guaranteed to be the same. Treat the Apache 2.0 promise as a promise about the preview, not about the family.

• Stability — the Preview is explicitly an intermediate artifact whose capabilities "may still fall short" and whose scores Baidu expects to improve; the official release is by definition the version Baidu is willing to stand behind.

• Benchmark standing — the Preview has twenty vendor-reported rows and no independent score; ERNIE 5.5 Flash will have whatever Baidu publishes at release plus, if the pattern holds, a second round of numbers that make the preview's look dated. Neither has an Artificial Analysis entry today.

• Price — the Preview is $0.28 per million input tokens and $1.13 per million output on Qianfan, with cached reads at $0.028. There is no price for ERNIE 5.5 Flash, and assuming it will match the preview's is an assumption: preview pricing is frequently a placeholder.

• What you can do with it — the Preview can be called today and cannot be downloaded; ERNIE 5.5 Flash can be neither called nor downloaded, because it does not exist yet.

The asymmetry that matters is at the bottom of that list. The preview is the only one of the two you can test, and it is the only one you cannot keep.

What the preview does that the name undersells

Preview implies thin, and on the evidence published on 11 November this one is not thin. It is a text-only model with tool use, structured output and a million-token window, priced at the cheap end of the flash tier, and Baidu's demos describe work products rather than chat exchanges: an Unreal Engine 5.8 game assembled in a single 11-hour coding run, roughly 100 mobile mini-games generated from design documents, a Boeing 747 and Daxing airport modelled through the Blender Python API by two coordinated agents, a W16 engine and a lunar rover built as editable FreeCAD assemblies, a 20-page research report on data-centre electricity costs, a 24-page autonomous-driving analysis built on 4,169 federal crash reports, and a 116-second animated short written in more than 10,000 lines of Python.

The single most useful detail in that list is not a capability claim, it is a limitation the post volunteers. Describing the Unreal project, Baidu writes that "because it is a text-only model and can't see the screen," the model converted playtest problems — a character hovering above the ground, a misplaced camera — into measurable tests in code and iterated until the suite passed. That is the honest shape of this model's competence: it cannot look at anything, so it writes assertions instead. Where a task can be expressed as a test, it holds up well across long runs. Where the only feedback is visual judgement, the model is working through tools and, in at least some benchmark harnesses, through another company's vision model.

That last point comes from the benchmark appendix rather than the prose. One footnote in the table states that "GLM-5.3 and our model lack native multimodal capabilities, image understanding is provided through a tool powered by Claude Opus 5.5." Whatever you conclude about the scores, the architecture is unambiguous: no native vision, and vision-dependent evaluation that measures a pipeline rather than a model.

If you are waiting for ERNIE 5.5 Flash, what are you waiting for?

Four things, and only one of them is a capability improvement.

The first is a stability guarantee. An intermediate checkpoint carries no promise that behaviour will not shift between now and release, which is fine for exploration and awkward for anything with a regression suite pinned to model output.

The second is a benchmark set you can compare against. The preview's twenty rows — Terminal-Bench 4.0 at 46.5, DeepSWE v1.1 at 71.2, MCP-Atlas at 84.6, BrowseComp at 87.4, HLE without tools at 46.4, AA-LCR v1.1 at 79.0 among them — are vendor-reported and unreproduced, and Baidu has said the model is improving quickly. Numbers from a finished model are more useful even when they are higher, because they are the numbers other people will also be citing.

The third is the licence. The base checkpoint shipped under a restricted community licence that excludes the EU, the UK and South Korea, caps free commercial use by related annual revenue and requires a "Powered by ERNIE" attribution on commercial products. The preview's weights are promised under Apache 2.0. If that promise is kept, the post-trained model will be far more permissive than the checkpoint it was trained from — a genuine reversal, and one worth confirming on the actual repository rather than on the announcement.

The fourth is price. $0.28 and $1.13 are preview numbers. Final pricing for a GA model in this tier could move in either direction, and any cost model you build today should be parameterised on the rate rather than hard-coded to it.

None of those four argues for waiting if you have a task you can evaluate this week. They argue for not signing a multi-quarter commitment to a checkpoint Baidu has described, in writing, as unfinished.

Running the comparison yourself

You cannot get ERNIE 5.5 Flash Preview from us. It is served through Baidu's own Qianfan API and through third-party platforms that resell it; OrcaRouter routes more than 200 models and not one of them is an ERNIE, so any suggestion that we can hand you this endpoint would be false.

What we can do is supply the comparison group, which is the part of this decision that is actually under your control: the models Baidu benchmarked against — GPT-6.1 Sol, Claude Opus 5.5, GLM-5.3, DeepSeek V4.1 Flash — are all on one OpenAI-compatible key at provider list price with 0% markup, so a price cut from any of those vendors reaches your bill the same day rather than at the next contract renewal.

The evaluation to run is deliberately boring. Pick ten long-horizon tasks from your own work. Run each through the preview and through the two models you would otherwise use. Record whether it finished, how many output tokens went out the door, and what the call cost. Output tokens are the number that decides this comparison, because the preview bills reasoning as output at $1.13 per million and exposes no way to turn reasoning off — a verbose model at a low rate card is frequently more expensive per task than a quiet model at a high one. Once you have per-task cost, the preview-versus-GA question stops being about benchmark rows and becomes about whether the current checkpoint is good enough to build on before it is replaced.

For the workloads that pass that test, automatic failover answers the stability objection in the only way it can be answered today: a model that may change under you should never be the only model in a request path. Route the preview as the primary where it wins and keep a fallback behind it, and a mid-stream capability shift becomes a routing change rather than an outage.

Five days separate the base checkpoint from the preview, and Baidu has said the finished model is coming. The useful reading of ERNIE 5.5 Flash versus ERNIE 5.5 Flash Preview is not a head-to-head at all — it is a schedule. One of them you can call now and will not be able to keep; the other you will be able to keep and cannot call yet. Anyone building on either should be measuring, not committing, until the second column has numbers in it.