Baidu releases ERNIE 5.5 Flash Preview: 280B parameters, 14B active, 1M context, $0.28/1M input

Baidu has released ERNIE 5.5 Flash Preview, a model with 280 billion total parameters that activates 14 billion parameters per token and supports a 1 million token context window. It is available through Baidu Qianfan and listed on OpenRouter as baidu/ernie-5.5-flash-preview, with a release date of Nov 10, 2026.

Specifications and pricing

Spec	Value
Total parameters	280B
Active parameters per token	14B
Context window	1.0M tokens
Input price	$0.28 per 1M tokens
Output price	$1.13 per 1M tokens
Cache read price	$0.028 per 1M tokens
Release date	Nov 10, 2026

Baidu Qianfan is the only provider listed for the model on OpenRouter. The listing notes that the model is not currently available through EU in-region routing.

What Baidu claims

The model description calls ERNIE 5.5 Flash Preview the latest foundation model in the ERNIE family, with "comprehensive improvements in coding, agentic capabilities, knowledge work, and deep research." It says a new architecture combines strong model capabilities with efficient inference, delivering generation speeds of more than 200 tokens per second. No benchmark scores are published on the listing, so these capability claims are not backed by figures there.

Observed performance on OpenRouter

As read on Nov 10, 2026, the OpenRouter listing shows the following for Baidu Qianfan:

• Throughput: 99 tokens per second (P50, best across providers)
• Latency: 1.98 seconds (P50, best provider)
• OpenRouter availability over the last 24 hours: 100.00%
• Per-provider uptime: not shown

These are live statistics and will change as usage grows. The 99 tokens per second figure is the OpenRouter-measured number at the time of reading. The 200 tokens per second figure is Baidu's own description of generation speed.

On the listing's usage panel, Open WebUI is the top app by token volume, at 46K tokens.

Baidu's other models on OpenRouter

The listing also shows Baidu's other available models:

• ERNIE-Light-DT-1.0: based on open-source GLM models, with a 1.0M context window, priced at $1.13 input / $3.94 output per 1M tokens. It uses Baidu Sparse Attention (BSA), which Baidu says reduces KV cache storage and computational costs.
• CoBuddy: a code generation model with a 131K context window and up to 65K output tokens, running on fp8 quantization.
• ERNIE 4.5 300B A47B: a 300B-parameter MoE model that activates 47B parameters per token, with a 131K context window.
• ERNIE 4.5 VL 424B A47B: a multimodal MoE model with 424B total and 47B active parameters.
• Qianfan-OCR-Fast: an OCR-focused multimodal model with a 66K context window.

Training cutoff date and benchmark results for ERNIE 5.5 Flash Preview are not listed.

What this means

ERNIE 5.5 Flash Preview pairs a sparse design (14B active of 280B total parameters) with a 1M-token context window at $0.28 input and $1.13 output per 1M tokens. That makes it a long-context option for the coding, agentic and deep research workloads Baidu names. The cache read price of $0.028 per 1M tokens is relevant for workloads that resend large shared prompts.

Two caveats apply. The "Preview" label is part of the model name, and Baidu has published no benchmark scores on the listing, so quality claims cannot yet be checked against figures. Measured throughput on OpenRouter was 99 tokens per second on Nov 10, 2026, while Baidu's description cites more than 200 tokens per second. Teams evaluating the model should test latency and throughput under their own workloads, and EU-based teams should note the absence of in-region routing.