Baidu releases ERNIE-Light-DT-1.0 with 1.0M-token context at $1.13 input, $3.94 output per 1M tokens

Baidu has released ERNIE-Light-DT-1.0, a model with a 1.0M-token context window priced at $1.13 per 1M input tokens and $3.94 per 1M output tokens. It is listed on OpenRouter as Baidu Qianfan: ERNIE-Light-DT-1.0 (model ID baidu/ernie-light-dt-1.0-intl), with a release date of Nov 10, 2026.

Key specifications

Spec	Detail
Model	ERNIE-Light-DT-1.0
Provider	Baidu Qianfan
Context window	1.0M tokens
Input price	$1.13 per 1M tokens
Output price	$3.94 per 1M tokens
Cache read price	$0.113 per 1M tokens
Release date	Nov 10, 2026
Parameter count	Not stated
Training cutoff	Not stated
Benchmark scores	None published on the listing

Architecture

According to the listing, ERNIE-Light-DT-1.0 is based on open-source GLM models and includes research findings from Baidu Search's Lightning large model.

The model uses Baidu Sparse Attention (BSA), a sparse attention mechanism developed by Baidu ERNIE. BSA applies feedback-driven history compression, which the listing says improves the storage and indexing of historical information. Baidu says this significantly reduces KV cache storage and computational costs while maintaining model performance, and that the result is longer context windows at a lower cost. These are Baidu's claims, and no benchmark data accompanies them on the listing.

Availability and live performance

The model is served by a single provider, Baidu Qianfan, on OpenRouter. The listing notes it is not currently available through EU in-region routing.

Live statistics read from the OpenRouter page on Nov 10, 2026 show a P50 latency of 2.30s, throughput of 35 tok/s and uptime of 100.00% for Baidu Qianfan. The page reports not enough usage data yet for apps using this model. These figures reflect the release day only and will change as traffic grows.

Other Baidu models on the platform

The same listing shows ERNIE 5.5 Flash Preview, which Baidu describes as having 280 billion total parameters, 14 billion activated per token, and a 1 million token context window, priced at $0.28 input and $1.13 output per 1M tokens. The ERNIE 4.5 series entries on the page list context lengths of 131K tokens.

What this means

ERNIE-Light-DT-1.0 is a long-context offering built around cost. Baidu's stated aim with BSA is to reduce KV cache storage and compute so that a 1.0M-token window is cheaper to serve. Whether that holds up depends on independent testing, since no benchmark scores or parameter counts are given.

Developers considering the model should note the single-provider setup, the lack of EU in-region routing, and the early throughput of 35 tok/s. The $0.113 cache read price is listed separately from standard input pricing.

The GLM base and the Lightning research lineage tell builders where the model comes from. Baidu has not published figures in the listing that show how it performs against its own ERNIE models.