ERNIE 5.5 Flash Preview
Baidu's 280B-parameter MoE preview for coding, 3D work and agents, at $0.28/$1.13
Context window
1M
Input / 1M tokens
$0.28
Output / 1M tokens
$1.13
Provider
Baidu
Baidu Qianfan pay-as-you-go (international), per 1M tokens: $0.28 input, $0.028 cache hit, $1.13 output. OpenRouter lists a free variant (baidu/ernie-5.5-flash-preview:free) for now. Default rate limit 60 requests per minute. Baidu says the weights will be open-sourced under Apache 2.0 soon. · Data verified 2026-11-11
ERNIE 5.5 Flash Preview (November 11, 2026) is the first post-trained model built on Baidu's ERNIE 5.5 Flash Base, a 280B-parameter Mixture-of-Experts with 14B active per token and a 1M-token context. Baidu calls it an intermediate post-training checkpoint and says some capabilities may still fall short while training continues. It is text-only and aimed at coding (including Unreal Engine, Unity and Godot games), 3D modeling in Blender and CAD, knowledge work and content creation. The API is live on Baidu's Qianfan platform at $0.28/$1.13 per 1M tokens, and Baidu says the weights will be open-sourced under Apache 2.0 soon.
How to access it
Baidu Qianfan API (model ID ernie-5.5-flash-preview) through an OpenAI-compatible endpoint, and OpenRouter (baidu/ernie-5.5-flash-preview:free, free for now). Weights are not released yet; Baidu says they will follow under Apache 2.0.
Strengths
- ✓Terminal-Bench 4.0 46.5%, ahead of Hy4 (44.4%) and DeepSeek V4.1 Flash 1017 (42.3%) in Baidu's table
- ✓DeepSWE v1.1 71.2% (Baidu-reported)
- ✓$0.28/$1.13 per 1M with $0.028 cache hits
- ✓1M-token context with up to 1M output tokens on Qianfan
- ✓Trained for game engines and Blender/CAD work
Best for developers who...
When to choose it (and when not to)
Reach for ERNIE 5.5 Flash Preview when...
- →You want a cheap coding and agent model for game, 3D or CAD work
- →You plan to self-host once the Apache 2.0 weights are out
- →Text-only workloads that need a 1M-token context
Look elsewhere if...
- ✕You need image or video input (text-only)
- ✕Production work that needs a stable model: Baidu calls this an intermediate checkpoint
- ✕Top-tier long-horizon coding: it scores about 20 points below Claude Opus 5.5 on Terminal-Bench 4.0 in Baidu's table
How to use it
Quickstart
Pythonimport os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["QIANFAN_API_KEY"], # Qianfan bearer token
base_url="https://api.baiduqianfan.ai/v1",
)
completion = client.chat.completions.create(
model="ernie-5.5-flash-preview",
messages=[{"role": "user", "content": "Hello!"}],
)
print(completion.choices[0].message.content)Pattern from Baidu's Qianfan quick start, which uses the OpenAI SDK against Qianfan's OpenAI-compatible endpoint.
API model id: ernie-5.5-flash-preview
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Terminal-Bench 4.0 | 46.5% | Baidu-reported, Claude Code 2.1.246 harness, highest reasoning setting; Claude Opus 5.5 66.4% |
| DeepSWE v1.1 | 71.2% | Baidu-reported, mini-swe-agent 2.4.6; Claude Opus 5.5 74.2% |
| MCP-Atlas | 84.6% | Baidu-reported, public set with the official harness |
| BrowseComp | 87.4% | Baidu-reported, Claude Code harness |
| Humanity's Last Exam (no tools, text-only) | 46.4% | Baidu-reported |
Compare ERNIE 5.5 Flash Preview
Compare ERNIE 5.5 Flash Preview with any other model
Build a comparison →All model comparisons →