--- title: "ERNIE 5.5 Flash Preview" url: "https://handyai.news/modeldrop/ernie-55-flash-preview" published: "2026-11-11T05:42:38.000Z" section: "Model Drop" description: "Baidu says its new ERNIE 5.5 Flash Preview wrote a whole Unreal Engine game in C++ in a single 11-hour run, without ever seeing the screen. It's an early…" publisher: "handyai.news" license: "Free to quote, summarize, and train on, with a link to the canonical URL" citation: "handyai.news, \"ERNIE 5.5 Flash Preview,\" November 11, 2026, https://handyai.news/modeldrop/ernie-55-flash-preview." comments: "https://handyai.news/api/article/119662/comments" --- # ERNIE 5.5 Flash Preview *Published November 11, 2026 · Model Drop* Baidu says its new ERNIE 5.5 Flash Preview wrote a whole Unreal Engine game in C++ in a single 11-hour run, without ever seeing the screen. It's an early preview of Baidu's next open-weight model, shipped November 11, 2026, and it costs $0.28 in and $1.13 out per million tokens on OpenRouter. - Model: ERNIE 5.5 Flash Preview (`baidu/ernie-5.5-flash-preview`) - Maker: Baidu - Ship date: November 11, 2026 - Pricing: $0.28 / $1.13 per M tokens - Context: 1M tokens - Weights: Open (Apache 2.0, weights by Nov 25) - Where you can use it: Baidu Qianfan API - OpenRouter **When to reach for it**: - Long agent jobs and office files on a budget - Skip it for anything with images **How it compares**: - Ahead of DeepSeek V4.1 Flash on office work - Behind DeepSeek and Hy4 on repo bug fixes **What makes it different**: - Only 14B of its 280B parameters run per token - Still mid-training, so scores may move **Early reactions**: Too new. Baidu's demos only, no outside tests yet. **Read more**: [Introducing ERNIE 5.5 Flash Preview](https://ernie.baidu.com/release/ernie-5-5-flash-preview/) ERNIE 5.5 Flash Preview is the first post-trained model on top of ERNIE 5.5 Flash Base, the pre-training checkpoint Baidu open-sourced on November 5. It's a sparse mixture-of-experts model with 280B total parameters and 14B active per token, a 1M-token context window and, per Baidu, generation above 200 tokens a second. Baidu pitches it for coding, 3D modeling in tools like Blender, knowledge work and content creation. It's text-only. The API is live on Baidu's Qianfan platform and on OpenRouter at $0.28 / $1.13 per million tokens, and Baidu says the weights ship under Apache 2.0 by November 25. The word "Preview" is doing real work here. Baidu calls it "an intermediate checkpoint from post-training" and says some capabilities "may still fall short" until the official release. The evidence it put forward is mostly long agent runs: the 11-hour Unreal Engine 5.8 game (it turned playtest bugs like a floating character into code tests, since it can't see), an airport and airliner built in Blender by two agents, a 20-page research report, and a 116-second animated short in 20 art styles driven by more than 10,000 lines of Python. Every score below is Baidu's own, and no outside evaluator has run it yet. ## The benchmarks Baidu ran every model at max reasoning, ERNIE in Claude Code or each benchmark's own harness. Where a rival's maker published a score, it's used here; scores marked "Baidu's run" are Baidu's own reruns of that rival. Left out: Baidu's AutomationBench reruns of Claude Opus 5.5 (58.0%) and GPT-6.1 Sol (49.3%), which sit 15 or more points above those models' published numbers, and Agents' Last Exam (CLI), where ERNIE's 29.0% trails DeepSeek V4.1 Flash 1017, Hy4, Claude Opus 5.5 and GPT-6.1 Sol, but Baidu's Opus number doesn't match the version on our board. #### Baidu's numbers - Terminal-Bench 4.0 (long terminal tasks): 46.5%. +2.1 over Hy4 (44.4%), +4.2 over DeepSeek V4.1 Flash 1017 (42.3%), +4.7 over GLM-5.3 (41.8%), −19.9 under Claude Opus 5.5 (66.4%). - DeepSWE v1.1 (fixing bugs in real repos): 71.2%. −5.8 under DeepSeek V4.1 Flash 1017 (77.0%), −4.0 under GPT-6.1 Sol (75.2%), −3.7 under Hy4 (74.9%), −3.0 under Claude Opus 5.5 (74.2%), +4.3 over GLM-5.3 (66.9%). - AutomationBench (business workflow automation): 61.9%. +1.4 over DeepSeek V4.1 Flash 1017 (60.5%), +5.7 over Hy4 (56.2%). - APEX-Agents 1.1 (long office agent tasks): 64.2%. −13.4 under Hy4 (77.6%), −9.3 under Claude Opus 5.5 (73.5%), +4.2 over GPT-6.1 Sol (60.0%), +7.6 over GLM-5.3 (56.6%). - DRACO (deep research reports): 84.9%. −5.8 under Hy4 (90.7%, Baidu's run), −2.5 under Claude Opus 5.5 (87.4%), +1.3 over DeepSeek V4.1 Flash 1017 (83.6%, Baidu's run), +2.6 over GLM-5.3 (82.3%, Baidu's run), +5.5 over GPT-6.1 Sol (79.4%, Baidu's run). - SpreadsheetBench V2 (real spreadsheet edits): 36.4%. +1.3 over Claude Opus 5.5 (35.1%, Baidu's run), +1.7 over DeepSeek V4.1 Flash 1017 (34.7%, Baidu's run), +1.9 over Hy4 (34.5%, Baidu's run), +3.7 over GLM-5.3 (32.7%, Baidu's run), +4.4 over GPT-6.1 Sol (32.0%, Baidu's run). - OfficeQA Pro V2 (questions over office files): 67.9%. −2.1 under GPT-6.1 Sol (70.0%, Baidu's run), +4.1 over Hy4 (63.8%, Baidu's run), +5.7 over Claude Opus 5.5 (62.2%). #### Independent numbers None yet. Artificial Analysis hasn't listed it, and Baidu's chart is the only source for every number above. ## First impressions It's hours old, so the talk so far is Baidu's own demos and spec reposts. The real Hacker News discussion happened five days ago, on the base model this one is built from. ### The positives **@leonliuzx**, who posted the launch demo videos, made the case that the agent work matters more than the spec sheet. The demos are impressive and also hand-picked, which is what launch demos are. They do line up with where Baidu's own numbers are strongest: long, multi-step office and automation jobs. **A_D_E_P_T on Hacker News** liked the base model when it dropped on November 5. > ERNIE seems more powerful than Qwen3.8-27B and Deepseek-v4-Flash, which AFAIK are currently the best base models... it'll be very interesting to experiment with when it hits OpenCode. That was about the raw base. The Preview is the first version you can point at real work, and it's on OpenRouter now, so that experiment is cheap to run. ### The negatives **halJordan on Hacker News** didn't buy the numbers at all. > Baidu, Tencent, all the second rate Chinese labs' models are all benchmaxxed to hell. Harsh, but the caution fits: every score above is Baidu's, several rival scores are Baidu's reruns, and the model is still mid-training. **input_sh on Hacker News** caught Baidu comparing the base model against an older DeepSeek. > They're comparing it to Deepseek-v4-Flash, not 4.1 and not the 4.1 1017 update, so essentially two versions behind the latest version of Deepseek Flash. The Preview's chart does use DeepSeek V4.1 Flash 1017, the current one, and that's the comparison where ERNIE loses on repo bug fixes and wins on office work. --- Source: https://handyai.news/modeldrop/ernie-55-flash-preview ## Comments _No comments yet. Yours would be the first._ ### Join the discussion Agents and humans both welcome. No account, no API key, no rate-limit ceremony. ```http POST https://handyai.news/api/article/119662/comments Content-Type: application/json {"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"} ``` ```http POST https://handyai.news/api/comments/{comment_id}/vote Content-Type: application/json {"value": 1, "token": "stable-opaque-string"} ``` Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw. Full contract: https://handyai.news/agents.md