Baidu’s ERNIE Light DT 1.0 costs substantially more than Alibaba’s Qwen 3.8 Flash, while accepting text only and offering no published benchmark in the comparison. OrcaRouter’s side-by-side analysis gives developers a useful reason to test before they commit: a million-token context window is not much of a selling point if the price and performance case remain unclear.
Watch Desk analysis
What happened
The comparison from OrcaRouter says ERNIE Light DT charges $1.13 per million input tokens and $3.94 per million output tokens. Qwen 3.8 Flash is listed at $0.15 and $0.47 respectively. Both models have a roughly one-million-token context window and a maximum output of 131,072 tokens, according to the comparison.
The differences are not just on the invoice. OrcaRouter says ERNIE Light DT accepts text only and has no published benchmark, while Qwen accepts text, images and video. The article describes Qwen’s open-weights sibling as available under the Qwen Community License, but that sibling is not the same thing as the hosted Qwen endpoint.
Why it matters
For text-only tasks, the price gap makes the missing evidence especially important. OrcaRouter argues that Baidu’s sparse-attention and history-compression approach could reduce the cost of serving long requests, but says Baidu has not published the cache-footprint data or benchmark results needed to assess that claim. A million-token context limit tells you how much a model can accept, not how well it handles information buried within it.
For work involving screenshots, scanned pages, charts or video, the modality gap may settle the choice before price does. A separate OCR or vision system could fill the gap, but it also adds another step, cost and potential failure point.
Our read
Treat the quoted prices and specifications as OrcaRouter’s comparison, not a universal bill or independent benchmark verdict. Its strongest practical suggestion is also the least glamorous: test both against your own workload, particularly if long-context performance is the reason to consider paying more. The premium needs evidence; a promising-sounding architecture is not the evidence.
What to watch
- Whether Baidu publishes matched performance and throughput results.
- How both models perform on questions whose answers appear at different depths in long documents.
- Whether either provider changes its prices, modality support or published technical detail.
Discussion spark: What evidence would justify paying more for a text-only model: stronger results on your own long-context tests, or published benchmarks anyone can reproduce?
Sources and evidence
- ERNIE Light DT 1.0 vs Qwen 3.8 Flash: Dearer, Text Only – OrcaRouter (10 November 2026, 20:39 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.