ERNIE 5.5 Flash Preview's 1668 on WebDev Arena Has Zero Human Votes and a Proprietary Tag Despite the "#2 Open Model" Billing
Arena said on November 12 that Baidu’s ERNIE 5.5 Flash Preview was “trending towards becoming the #2 open model” on Code Arena WebDev at 1668. Baidu’s ERNIE account repeated it and added that Apache 2.0 weights are coming soon. The leaderboard row says something narrower.
The Row
On the November 11 snapshot of the WebDev board (158 models, about 998,000 votes), ERNIE 5.5 Flash Preview shows:
- Score: 1668, +16/-16
- Rank: blank
- Votes: N/A
- Label: AutoEval
- License tag: Proprietary
- Price: $0.28 per million input tokens, $1.13 per million output
- Context: N/A
AutoEval is Arena’s method for scoring a model before enough human votes arrive. A reward model trained on human preferences casts synthetic votes on regenerated responses, and those are blended with live votes on other models. Arena says the score is updated once real votes validate it. The 1668 is a placeholder with an error bar, not a ranking earned through battles.
Who Is Inside the Interval
A 1668 score with a 16-point interval spans roughly 1652 to 1684. Seven ranked models on the board fall inside it:
| Model | WebDev score | Licence |
|---|---|---|
| Gemini 4 Argon (High) | 1681 | Proprietary |
| GPT-6.1 Terra (Max) | 1678 | Proprietary |
| Qwen3.8-Max (0902) | 1676 | Proprietary |
| GPT-6 Sol (Max) | 1673 | Proprietary |
| Qwen3.8-Max | 1670 | Proprietary |
| ERNIE 5.5 Flash Preview | 1668 (AutoEval, no votes) | Proprietary tag |
| DeepSeek-V4.1-Flash (Max) | 1664 | Open |
| Kimi K3 (Max) | 1658 | Open |
The two open-weights models Arena says ERNIE sits “just above”, DeepSeek-V4.1-Flash at 1664 and Kimi K3 at 1658, are 4 and 10 points behind on point estimates. Both sit well inside ERNIE’s own interval. On the board’s published uncertainty the three cannot be separated.
The Open-Weights Claim
Baidu’s release page for the Preview, dated November 11, says the weights “will be open-sourced under Apache 2.0 soon.” They are not published. Until they are, the model is available only through Baidu’s Qianfan API, and Arena’s row correctly labels it Proprietary. “#2 open model” is a forecast about a licence and a rank at the same time.
Baidu also calls the Preview an intermediate post-training checkpoint and warns that some capabilities may still fall short. The model is built on ERNIE 5.5 Flash Base, the raw 280B-parameter mixture-of-experts with 14B active that Baidu released as open weights on November 4. The release page offers no benchmark scores for the Preview itself, only demos: an Unreal Engine 5.8 game written in C++ over an 11-hour task, nearly 100 mobile mini-games, and Blender and FreeCAD builds.
Read It This Way
The 1668 is a useful early signal. At $0.28 and $1.13 per million tokens, a model in the same WebDev band as GPT-6 Sol and Qwen3.8-Max would be a large cost drop. Claude Fable 5.2 leads the board at 1838, 170 points ahead, so ERNIE is not near the frontier. But it is only worth trusting once human votes replace the estimate and the weights ship. Check the row again when the vote count stops reading N/A.