ERNIE 5.5 Flash Preview Wants Your Hardest Task: What the Eleven-Hour Build Actually Proves

ERNIE 5.5 Flash Preview went live as a billed endpoint on Baidu's Qianfan platform on 10 November 2026, and the official launch post followed on 11 November with a single headline claim: give it your hardest tasks. That claim rests on one artefact more than any other — a third-person Unreal Engine 5.8 game built in what Baidu describes as a single eleven-hour coding task — and on twenty benchmark rows that Baidu ran on its own harnesses with judges it selected. Neither the eleven hours nor the twenty numbers has been reproduced outside the company. The interesting question is not whether the demo is impressive. It is whether an eleven-hour agentic run is even the right thing to measure, and what the spec sheet says about the bill you would pay for one. The catalogue holds 228 models from 17 providers and there is no ERNIE entry in a catalogue of 228 models from 17 providers so the comparison group is one key away even though the subject is not.

The claim is about duration, not difficulty

Every model launch post says the model is good at hard things. This one defines hard differently: it defines it as long. The game build, the 3D scenes, the multi-thousand-file workspaces and the 116-second animated short are all exercises in holding a plan together across hours of edits and tool calls, not in answering a question that a smaller model would get wrong.

That distinction matters because the two properties come apart. A model can be excellent at single-turn reasoning and collapse at hour four of a refactor, when the context is full of its own earlier mistakes and the plan needs revising rather than executing. A model can also be mediocre on a one-shot puzzle and still be the right choice for an overnight build. Baidu is betting you care about the second thing.

The most useful sentence in the whole post is not a benchmark. It is in the game-development section, describing how the model worked without being able to see: because it is text-only and cannot look at the screen, it converted bugs spotted in playtesting — a character hovering above the ground, an inconsistent running direction, a misplaced camera — into measurable tests in code. That is a candid description of a workaround, and it tells you something the spec sheet does not. The model's competence on visual work is delegated to tooling it writes itself, which means the quality of the result depends on whether it can invent a verifiable test for the thing it cannot see.

What the reel actually contains

The four case studies are worth listing because they are more specific than launch copy usually gets, and specificity is what makes them checkable in principle.

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

• 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.

The training figure behind the knowledge-work section is the largest number in the post: 300,000 workspaces holding 10 million reference files across more than 40 file types. That is a statement about what the model was trained to survive, and it is the part of the argument a benchmark table cannot make at all.

The twenty rows, read for what they say about long runs

The appendix is a single image, which means the numbers cannot be copied, sorted or checked against a spreadsheet. Read off that image, the ERNIE column runs as follows, all vendor-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, AA-Omniscience-Public at 25.2, AA-LCR v1.1 at 79.0.

• 3D — BlenderWorld text-only at 79.0.

The table is far more useful read as a map of what still fails than as a scoreboard. Four rows sit below 40: ProgramBench at 29.5, Agents' Last Exam on the command line at 29.0, AA-Omniscience-Public at 25.2, and Terminal-Bench-Science at 21.0. Those are not soft benchmarks. Agents' Last Exam is designed to be near the frontier of what an agent can complete autonomously; omniscience punishes confident wrong answers; Terminal-Bench-Science tests laboratory-grade tasks where a plausible-looking procedure is worthless. A model that wants your hardest task and scores 21.0 on the science variant of the hardest agentic benchmark is telling you where the ceiling currently sits. SpreadsheetBench V2 at 36.4 says something similar about spreadsheet work specifically, which sits awkwardly beside the launch post's 1,000-path Monte Carlo workbook.

The rows above 80 are equally informative for a different reason. BrowseComp at 87.4, MCP-Atlas at 84.6, CookieBench at 88.2 and DRACO at 84.9 are all tasks where the model is scored on a completed, checkable end state rather than on an answer. That pattern — strong on verifiable completion, weak on open-ended scientific judgement — is exactly what you would expect from a model trained on 300,000 workspaces full of files that either build or do not.

The footnotes carry the honesty. Each benchmark ran on a named harness, with Claude Code 2.1.x, mini-swe-agent 2.4.6, Codex 0.162.1 and something abbreviated DSH all appearing, 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 this model lack native multimodal capability and that image understanding was provided through a tool powered by Claude Opus 5.5. At least some of the vision-dependent rows therefore measure a harness in which ERNIE 5.5 Flash Preview calls a competitor's model to look at things. That is a legitimate engineering choice and a strange thing to find in a scoreboard.

The two rows prefixed AA need 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. 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.

What an eleven-hour run would cost, and what Baidu did not say

The endpoint documentation is the part of this release that turns a demo into an engineering decision. Qianfan lists ERNIE 5.5 Flash Preview at $0.28 per million input tokens and $1.13 per million output, with cached reads at $0.028, a context window of 1,024,000 tokens, an output ceiling of 921,600 tokens, and reasoning that is always on at a single high effort level with no way to switch it off.

That output ceiling is the mechanical reason a long single run is even conceivable: the model can emit nearly as many tokens as its whole context window holds. It is also the reason the bill is not obvious. Prompting is priced at a quarter of the output rate, but everything the model thinks is billed at the output rate, and a model that cannot stop thinking is a model whose cost per task is dominated by tokens nobody reads.

Run the arithmetic on a maximal single call, which is the most favourable case available. Filling the 1,024,000-token window costs about $0.29 at the input rate. Emitting the full 921,600-token output ceiling costs about $1.04. That is roughly $1.33 for one call that exhausts both limits. The eleven-hour Unreal build very plausibly did not do that, and the launch post does not say how many tokens it consumed, how many separate calls it took, or what it cost. Until someone measures output tokens per completed task on their own workload, any claim about this model being cheap or expensive on hard work is a comparison of rate cards rather than a comparison of results.

Two things the release still owes you

The post closes by calling ERNIE 5.5 Flash Preview "an intermediate checkpoint from post-training," conceding that training is not finished and that some capabilities may yet fall short. It also promises the weights will be open-sourced under Apache 2.0 soon, with no date, and as of 11 November there is no preview repository on Hugging Face — the only ERNIE 5.5 artefacts published there are the base checkpoint and its BF16 variant. There is no model card, no knowledge cutoff, and no latency figure anywhere in the release.

So the honest summary of "give it your hardest task" is narrower than the slogan. ERNIE 5.5 Flash Preview is a text-only model with a very large context, an unusually large output allowance, a strong record on tasks that end in a build artefact, a weak record on tasks that end in a scientific judgement, and zero independent verification. That is a real and specific profile. It is also one that a second checkpoint will probably overwrite.

How to test that profile without betting a path on it

Baidu serves this model through its own Qianfan API and through third-party platforms that resell it. OrcaRouter does not route ERNIE 5.5 Flash Preview — there is no ERNIE entry in a catalogue of 228 models from 17 providers and no Baidu endpoint behind it — so nothing here should be read as an offer to serve it.

What the catalogue is good for in this situation is the other half of the experiment. The models this one is measured against are reachable through one key at provider list price with no markup, which means you can run the same hard tasks through them and compare completion rates rather than published scores. 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 at hour three.

The evaluation worth running this week is small. Take ten of your own long-horizon agent tasks — the ones you would not normally trust to a single unattended run — and put each one through ERNIE 5.5 Flash Preview and through two models you already trust. Record whether it finished, how many output tokens it burned, and what it cost. That produces the two numbers Baidu left out: tokens per task and cost per completed task. Everything else in the launch post, including the 46.5 and including the eleven hours, is somebody else's harness on somebody else's machine.

What would settle the question is not another demo reel. It is a date on the Apache 2.0 release, a model card with a knowledge cutoff, an Artificial Analysis row produced by Artificial Analysis, and a second checkpoint whose numbers can be set against this one. Baidu has published the story it should have published alongside the endpoint. The next thing it owes the people it just invited to hand over their hardest tasks is the artefact that lets them stop taking its word for how the hardest ones turned out.