ERNIE 5.5 Flash Preview Finally Has a Launch Post, Twenty Benchmarks and a Weights Promise

46.5 on Terminal-Bench 4.0 is the first score Baidu has ever published for ERNIE 5.5 Flash Preview, and it landed on 11 November 2026 — one day after the model itself went live on Qianfan as a billed endpoint with a price list and no documentation at all. The release post carrying that number is the first official statement about the model, and it answers the three questions the endpoint could not: what it is (280 billion total parameters with 14 billion active, a 1M-token context, post-trained on the ERNIE 5.5 Flash Base checkpoint from 5 November), how Baidu says it compares (twenty benchmark rows set beside GPT-6.1 Sol, Claude Opus 5.5, GLM-5.3, GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017 and a model listed only as Hy4), and whether you will ever be able to run it yourself (the weights "will be open-sourced under Apache 2.0 soon"). Every figure in that table is vendor-reported, produced on Baidu's own harnesses with judge models Baidu selected, and none of it has been reproduced outside the company. The benchmark group is the part of this comparison you can actually 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.

What Baidu put on the page

The post is structured as a demo reel with a benchmark appendix, which tells you what Baidu thinks the argument is. Four capability sections — game development, 3D modelling and design, knowledge work, video creation — each built from artefacts the model produced end to end inside an agent harness, with external assets and tools credited piece by piece. The specification sits in the opening paragraph and never comes back: 280B total, 14B active, 1M-token context, API on Qianfan, weights promised under Apache 2.0. There is no model card, no knowledge cutoff, no latency figure, no pricing, and no statement about how many tokens of reasoning the model emits per task.

Two sentences in the post matter more than the rest of it combined. The first is in the game-development section, describing the Unreal Engine 5.8 project: "Because it is a text-only model and can't see the screen, it turned problems spotted in playtesting, such as a character hovering above the ground, an inconsistent running direction and a misplaced camera, into measurable tests in code." That is an unusually candid description of a workaround — the model wrote tests because it could not look at the game — and it is the clearest statement yet that this is a text-in, text-out system whose visual competence is delegated to tooling.

The second is the closing paragraph: ERNIE 5.5 Flash 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." A launch post that tells you the launch is not finished is worth more than one that does not, and it reframes everything above it. You are not reading a product announcement. You are reading a progress report with a scoreboard attached, published so that users will report back before the official release.

The twenty benchmark rows, read carefully

The appendix table is a single image, which is itself a small tell: the numbers cannot be copied, sorted or checked against a spreadsheet. Read off the image, the ERNIE column runs as follows, all Baidu-reported and unreproduced:

• Agentic coding — Terminal-Bench 4.0 at 46.5, DeepSWE v1.1 at 71.2, ProgramBench (Almost Solved) at 29.5, NL2Repo-Bench at 64.8, Terminal-Bench-Science 0.1 at 21.0.

• Computer use and automation — CookieBench at 88.2, AutomationBench at 61.9, APEX-Agents 1.1 at 64.2, DRACO at 84.9.

• Knowledge work — SpreadsheetBench V2 at 36.4, OfficeQA Pro V2 at 67.9, Agents' Last Exam (CLI) at 29.0, SkillsBench 1.1 text-only at 73.5.

• General agents and tool use — Toolathlon-Verified at 79.0, MCP-Atlas at 84.6, BrowseComp at 87.4.

• Reasoning — HLE without tools at 46.4 with a footnote marker, AA-Omniscience-Public at 25.2, AA-LCR v1.1 at 79.0.

• 3D — BlenderWorld text-only at 79.0, the one row that speaks directly to the CAD and Blender claims in the body.

The footnotes are where the honesty lives, and they are worth reading before any of the cells. Each benchmark ran on a named harness — Claude Code 2.1.x, mini-swe-agent 2.4.6, Codex 0.162.1 and something abbreviated DSH all appear — and the judge models are Qwen3.8-Max, DeepSeek-V4.1-Flash-1017 and Gemini 3.1 Flash-Lite. One footnote concedes that "GLM-5.3 and our model lack native multimodal capabilities, image understanding is provided through a tool powered by Claude Opus 5.5." In other words, at least some of the vision-dependent benchmarks were completed by ERNIE 5.5 Flash Preview driving a tool that calls a competitor's model. That is a legitimate engineering choice and a genuinely strange thing to discover in a scoreboard, and it means the rows that involve looking at anything measure the whole harness, not the model alone.

The two rows carrying an AA prefix deserve their own caveat. AA-Omniscience-Public and AA-LCR v1.1 are benchmarks published by Artificial Analysis, the independent tracker — but these are Baidu's runs of them, not Artificial Analysis' runs. As of 11 November there is no ERNIE 5.5 row on the Artificial Analysis leaderboard, so there is no independent number to set against either cell.

The demo reel is the actual argument

If the benchmark table is the part that will be quoted, the four case studies are the part that is hard to fake, because they describe work products rather than scores. The details are specific enough to be checkable in principle and concrete enough to be useful when you are deciding what to point this model at.

• Game development — a third-person Unreal Engine 5.8 game built in a single 11-hour coding task, merging assets from several engine generations into one project and writing the game in C++; separately, nearly 100 mobile mini-games generated from design documents across Cocos, Unity, LayaAir, PixiJS and Canvas 2D, most of them structuring rules as a deterministic simulation so that bot playthroughs and state-space search can verify each level is clearable.

• 3D modelling — a Boeing 747 and Beijing Daxing airport built by two agents through the Blender Python API, one on the airport and one on the aircraft, then matched for scale, lit, rendered frame by frame with Cycles and OptiX and cut with FFmpeg; the Hōryū-ji temple complex researched and built step by step; a previs sequence of roughly 2,100 objects and 1.82 million triangles across four Blender scenes; a W16 engine, two Shenzhen towers and a lunar rover modelled in FreeCAD through Python scripts.

• Knowledge work — a 20-page research report on data-centre electricity costs, a 24-page autonomous-driving analysis built on 4,169 federal crash reports, an interactive wealth-planning workbook with a 1,000-path Monte Carlo, and a 6 minute 34 second orchestral piece delivered as WAV. The training story behind this is the largest number in the post: 300,000 workspaces holding 10 million reference files across more than 40 file types.

• Video — more than 150,000 creation environments and about 5,300 minutes of finished video, headlined by a 116-second 2D animated short written in more than 10,000 lines of Python that moves a single parameterised pelican through 20 art styles.

What these have in common is that none of them is a chat interaction. Every one is a long agentic run with tool calls, and the claim being made is about endurance — whether the model can hold a plan together across hours of edits. That is precisely the property a twenty-row benchmark table measures worst, and the reason the post leads with artefacts.

Where the launch post and the endpoint disagree

The model described in the post and the model described by Qianfan's specification sheet are recognisably the same system seen from two departments that did not compare notes. The endpoint went live on 10 November at $0.28 per million input tokens and $1.13 per million output, with cached reads at $0.028, a 1,024,000-token context, an output ceiling of 921,600 tokens, mandatory reasoning at a single high effort level, and no model card. The release post, published a day later, restates the context window, ignores the price entirely, says nothing about mandatory reasoning, and supplies the model card's worth of positioning without supplying the model card.

The practical consequence is that the two halves of the story have to be read together or not at all. The post tells you the model can hold an eleven-hour coding task; the endpoint tells you that everything it thinks is billed at the output rate and that you cannot turn thinking off. A run that spends hours iterating is a run that emits an enormous number of output tokens, and at $1.13 per million the cost of the demo reel is a question the post never raises. Until somebody measures output tokens per completed task on their own workload, any comparison between this model and a cheaper one is a comparison of rate cards.

What is still missing, and it is not small

Three absences define the current state of ERNIE 5.5 Flash Preview, and each resolves on a different timescale.

• The weights. The post promises Apache 2.0 "soon" but names no date, and as of 11 November there is no preview repository on Hugging Face — the only ERNIE 5.5 artefacts published there are ERNIE-5.5-Flash-Base and its BF16 variant, both created on 21 October. The base checkpoint shipped under the ERNIE 5.5 Flash Base licence, a restricted community licence that excludes the European Union, the United Kingdom and South Korea, caps free commercial use by related annual revenue and requires a "Powered by ERNIE" attribution. If the preview's weights do arrive under Apache 2.0, that is a change of licence terms for the family and a much bigger deal than any benchmark row.

• The independent score. Nothing in this release has been run by anyone other than Baidu. The benchmarks with AA in the name are the ones to watch, because Artificial Analysis runs its own harnesses and will publish a comparable row if and when it adds the model; until then the twenty cells are a claim, not a measurement.

• The finished model. Baidu says post-training is ongoing and that the official release will improve on the preview. That is an explicit statement that today's numbers have a shelf life, and it makes any decision to build on this checkpoint a decision about a moving target.

How to evaluate it without betting a production path on it

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 behind it, so nothing here should be read as an offer to serve it.

What the catalogue is useful for in this situation is the other side of the table: the models this one is being benchmarked against are all reachable through one OpenAI-compatible key at provider list price with 0% markup, which makes it possible to run the same twenty tasks through GPT-6.1 Sol, Claude Opus 5.5, GLM-5.3 and DeepSeek V4.1 Flash yourself and see whether Baidu's ordering survives contact with your prompts. Automatic failover does the second job — a preview with no uptime history never has to be the only thing holding up a request path, because a routed fallback can take the call when it stalls.

The evaluation worth running this week is narrow and cheap. Take ten of your own long-horizon agent tasks, run each one through the preview and through two models you already trust, and record three things per task: whether it finished, how many output tokens it burned, and what it cost. That gives you the two numbers Baidu did not publish — tokens per task and cost per completed task — against a baseline you chose yourself. Everything else in the release post, including the 46.5, is someone else's harness.

What would settle the question is not another demo. It is a date on the Apache 2.0 release, a model card with a knowledge cutoff and an evaluation table, an Artificial Analysis row produced by Artificial Analysis, and a second checkpoint whose numbers can be compared with this one. Baidu has now published the story it should have published on 10 November. The next thing it owes its users is the artefact that lets them stop taking its word for the scoreboard.