Baidu ERNIE 5.5 Flash Preview Explained: $0.28 Price, 1M Context and No Model Card Yet

Illustrative image: Baidu's headquarters at Shangdi, Beijing. No official visual of the model exists — Baidu published no launch imagery. Image: Wikimedia Commons.

TL;DR: Baidu ERNIE 5.5 Flash Preview explained in one line: Baidu quietly put the finished, chat-ready version of its 280-billion-parameter ERNIE 5.5 Flash model on sale on 10 November 2026 — $0.28 per million input tokens, $1.13 per million output, a 1-million-token context window — with no announcement, no model card and no benchmark scores at all.

📌 Key Highlights

• What: ERNIE 5.5 Flash Preview — the post-trained, chat-ready version of Baidu's 280B ERNIE 5.5 Flash — went live on Baidu Qianfan on 10 November 2026.
• Price: $0.28 per million input tokens, $1.13 per million output, $0.028 per million cached input — roughly ₹25 and ₹100 per million tokens.
• Specs: 280B total / 14B active per token, 1,024,000-token context, 921,600-token maximum output, text only, function calling and structured outputs.
• The catch: no launch post, no model card, no benchmark table, no published knowledge cutoff — and no independent score anywhere yet.
• Context: the base weights shipped on 5 November with 30 vendor benchmark rows — but those describe the unfinished checkpoint, not this model.

What happened?

On 10 November 2026, ERNIE 5.5 Flash Preview started answering prompts on Baidu Qianfan — and almost nobody noticed, because Baidu said nothing about it. No blog post, no press release, no model card; the listing simply appeared on OpenRouter the same day, while Baidu's own ERNIE site on 11 November still led with the base checkpoint and a "coming soon" promise. It is the quietest launch of a major model this month — the opposite of Claude Fable 5.2, which arrived with a full announcement and a benchmark table.

It is the finished version of the system Baidu opened up on 5 November. That release, ERNIE 5.5 Flash Base, was a raw pre-training checkpoint — 280 billion parameters, 14 billion active per token, downloadable for further training, but unable to chat or call tools. Post-training has turned it into a product: it answers prompts, calls functions and reasons before it speaks.

Baidu ERNIE 5.5 Flash Preview explained: the price and the specs

Everything knowable about the model fits in one table, because Baidu has published little else:

Spec	ERNIE 5.5 Flash Preview
Input price	$0.28 per million tokens
Output price	$1.13 per million tokens
Cached input price	$0.028 per million tokens
Total parameters	280 billion
Active parameters per token	14 billion
Context window	1,024,000 tokens
Maximum output	921,600 tokens
Input type	Text only — no image, audio or PDF parsing
Reasoning	Mandatory, high effort, always on
Tool use	Function calling and structured outputs
Access	Baidu Qianfan API, or resold through OpenRouter
Early speed (OpenRouter)	102 tokens/second, 2.03 s latency

The output ceiling is the interesting number: at 921,600 tokens it is roughly ten times what most chat models allow, because a model that must think before it answers needs room to think. The catch is the bill — reasoning tokens are charged at the output rate, so a request that thinks for 8,000 tokens and answers in 500 costs about a cent but puts 8,500 output tokens on the invoice. Across a forty-step agent loop, price per completed task matters more than price per million tokens.

Illustrative image: the entrance of Baidu's headquarters at Shangdi, Beijing. Image: Wikimedia Commons.

What is missing — and why it matters

What Baidu has not published matters as much as what it has: no model card, no evaluation table, no knowledge cutoff, no context-length caveats — just a three-sentence description of architecture, not quality. Nor can the one set of numbers that exists be reused: the thirty benchmark rows Baidu published for the base checkpoint on 5 November come from a different pipeline, so they are a ceiling on what post-training had to work with, not a forecast of what it produced.

Nobody else has measured it yet — Artificial Analysis carries no ERNIE 5.5 row — and there is no published measure of how verbose it is, the largest swing factor between two models at the same per-token price. Until somebody runs a fixed set of their own prompts and counts output tokens, any cost comparison is a comparison of rate cards, not bills.

How it compares

The sharpest comparison is with the models Baidu shipped alongside it — and with Meta's Muse Ember, launched the same week, which targets the same cheap-and-long-context corner:

Option	Input / output per 1M tokens	Context	Independent score
ERNIE 5.5 Flash Preview	$0.28 / $1.13	1,024,000	None published
MiniMax M3.1 Flash	$0.15 / $0.60	1,000,000	42 on the Artificial Analysis Index; $0.22 per completed task
ERNIE-Light-DT-1.0	$1.13 / $3.94	1,024,000	None published
ERNIE 5.5 Flash Base (weights)	Free download	128,000 (pre-training)	30 vendor rows only

Read plainly, the Preview costs roughly twice MiniMax M3.1 Flash per token with no independent score — priced as a capable model, with the burden of proof on Baidu. ERNIE-Light-DT-1.0, listed the same day, is the odd one out: a smaller model priced above its much larger sibling. And for a fixed, high-volume workload, the arithmetic worth doing is against the free base weights — a fine-tune of the checkpoint — rather than paying per token forever.

Why it matters for India readers

For Indian developers, the headline is a genuinely cheap million-token window: $0.28 per million input tokens — roughly ₹25 — for retrieval-heavy work such as searching long documents or feeding whole codebases to an agent. It is reachable from India through Qianfan or OpenRouter, with no India restriction published. But three limits bite first: text-only input, mandatory reasoning billed as output, and no benchmarks to show it beats the rivals Indian teams already use. Test it on your own workload before migrating anything — the wider pattern, from DeepSeek to MiniMax and now Baidu, is Chinese labs shipping capable models at low prices despite chip export restrictions, faster than they can document them.

Should you care?

✅ Yes, if you build AI features. A million-token window at $0.28 per million input tokens is genuinely cheap for retrieval-heavy and agentic work, with the tooling production use needs. Try it on a small workload and measure output tokens per task before judging the price. ⏳ Wait, if you need proven quality. With no benchmarks, no model card and no independent score, nobody outside Baidu knows yet whether it is good — keep your current model behind it as a fallback.

FAQs

What is ERNIE 5.5 Flash Preview?
The finished, chat-ready version of Baidu's ERNIE 5.5 Flash — a 280-billion-parameter mixture-of-experts model with 14 billion parameters active per token and a 1-million-token context window. It went live on Baidu Qianfan on 10 November 2026, five days after the raw base weights were published.

How much does ERNIE 5.5 Flash Preview cost?
$0.28 per million input tokens and $1.13 per million output tokens, with cached input reads at $0.028. Reasoning tokens are billed at the output rate, so the real cost depends on how much the model thinks, not just how much it answers.

What is the context length of ERNIE 5.5 Flash Preview?
The context window is 1,024,000 input tokens, with a maximum output of 921,600 tokens — an unusually large output ceiling that exists because reasoning is mandatory and always on.

Can developers in India use ERNIE 5.5 Flash Preview?
Yes — through Baidu's Qianfan API and OpenRouter, with no India restriction published. Prices are in US dollars, roughly ₹25 per million input tokens and about ₹100 per million output. The EU, UK and South Korea exclusion applies to the downloadable base weights' licence, not to API calls.

Takeaway: ERNIE 5.5 Flash Preview exists, it is cheap enough to test, and nobody outside Baidu knows yet whether it is good. The number to watch is not the $0.28 rate card but Baidu's model card — and your own output-token count. Run the same twenty prompts through it and your current model and count the output tokens; that comparison is worth more than any rate card.

Sources

• OpenRouter — ERNIE-5.5-Flash-Preview model page (10 November 2026)
• OrcaRouter — ERNIE 5.5 Flash Preview: $0.28 In, $1.13 Out, No Model Card (10 November 2026)
• OrcaRouter — ERNIE 5.5 Flash vs ERNIE 5.5 Flash Base: Download or Rent (10 November 2026)
• OrcaRouter — ERNIE Light DT 1.0 Priced Above Baidu's Own 280B Flash (10 November 2026)
• Baidu ERNIE team — ERNIE 5.5 Flash Base release page (5 November 2026)
• Avenvote — New AI Models, last 30 days (updated 11 November 2026)