Baidu ERNIE 5.5 Flash Preview: A 280B AI Model That Costs Almost Nothing to Run

Amir Farid
7 min read

Baidu ERNIE 5.5 Flash Preview: A 280B AI Model That Costs Almost Nothing to Run
Baidu ERNIE 5.5 Flash Preview is a 280-billion-parameter AI model with a 1-million-token context window, priced at roughly $0.14 per million input tokens and $0.56 per million output tokens. Announced 12 November 2026 at Baidu World 2026 in Beijing, it targets businesses running AI agents all day. The catch is in the name: Baidu calls it an intermediate checkpoint and warns some capabilities may be incomplete, with weights open-sourced by 25 November 2026. Cheap enough to test; too early to bet a core workflow on.
What ERNIE 5.5 Flash Preview actually is
ERNIE 5.5 Flash Preview is Baidu's latest large language model, built for agent-style work: long conversations, tool use, document handling and multi-step tasks. The "Flash" label signals speed and low cost rather than maximum capability. Baidu announced it alongside two application models and a marketplace called Token Exchange.
The headline numbers
The model has 280 billion total parameters, with 14 billion activated per token, using a sparse mixture-of-experts design that keeps costs down. It offers a 1-million-token context window, and Baidu says it was pre-trained from scratch on 33.5 trillion tokens. That window matches the headline figure of Gemini 4 Hydrogen.
What "Preview" really means
Baidu describes this release as an intermediate post-training checkpoint, not the finished ERNIE 5.5. Some capabilities may be incomplete. Treat it as a cheap way to learn what the model can and cannot do, not as a stable production dependency — keep a fallback model ready.
What it costs — and why the price matters

Price is the whole story. Most capable models charge enough per token that a business running agents around the clock must watch usage or limit which tasks get automated. A model this cheap changes that: agents can read long documents and retry failed steps cheaply. Our guide to AI calling agent costs in 2026 shows how token prices feed into the real monthly bill.
The published prices
Baidu lists RMB 1 per million input tokens, RMB 4 per million output tokens, and RMB 0.04 per million cached input tokens — roughly $0.14 and $0.56, with cached input near zero. The API is live on Baidu AI Cloud's Qianfan platform, and the model is free to try on OpenRouter during the preview. When tokens cost almost nothing, a week of testing on real customer questions might cost a few pounds rather than a few hundred.
How it compares on price with Gemini 4 Hydrogen and LongCat-2.5
Two recent releases give useful price anchors. A lower price says nothing about answer quality — that must be tested, not assumed.
A simple cost comparison
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| ERNIE 5.5 Flash Preview | About $0.14 | About $0.56 |
| Meituan LongCat-2.5 | $0.30 | $1.20 |
| Gemini 4 Hydrogen | $1.50 | $7.50 |
On these figures ERNIE is roughly half the price of LongCat-2.5 and roughly a tenth of Gemini 4 Hydrogen for input and output. That gap matters for high-volume work. It is not, on its own, a reason to switch a working system: a model that costs a tenth as much but answers half as well is a bad deal for anything a customer sees.
Price is not quality
There is no independent comparison of these three models on the same business tasks yet. Baidu says its model posts competitive scores on Terminal-Bench and AutomationBench and decodes at over 200 tokens per second — Baidu's own claims, not independently verified. Run a small fair test: fifty real examples from your business, scored blind against your current model.
The 1-million-token context window: where it helps
One million tokens is roughly hundreds of thousands of words, so an agent can hold plenty of background material at once. Retrieval still matters: pointing the model at the right documents beats dumping everything in. Our RAG chatbot guide covers combining retrieval with a large model, and our chatbot versus AI agent explainer helps decide what you actually need.
Where long context does not help
A long window does not fix a model that misunderstands the task. Short jobs — a delivery question, an appointment booking, a ticket classification — rarely need more than a few thousand tokens, and sending a million tokens with every question adds up. Our e-commerce AI agent guide shows how support, recommendations and abandoned-cart recovery can share one agent.
Open weights, Token Exchange and the Baidu stack
Baidu has promised to open-source the weights by 25 November 2026 under Apache 2.0. Base weights are already on Hugging Face under a Baidu custom licence, which is not fully open, so read the licence first. If the Apache 2.0 release lands, businesses could run the model on their own infrastructure. Our Mistral Large 4 open-weights post covers the hosted-versus-self-hosted trade-offs.
Token Exchange in plain English
Token Exchange is Baidu's marketplace for buying model services from multiple providers. Baidu says it handled 50 million-plus calls and 700 billion-plus tokens a day after a month of internal operation, with 100-plus model resources onboarded — company-reported figures, no independent audit. Judge it on whether it lowers cost or simplifies billing versus going direct.
The catches: preview limits, benchmarks and practical checks
Before building on a new model, check preview limits, benchmark claims, availability, and data handling. Baidu's ERNIE Evolve technology — described as recursive self-improvement — is claimed to boost inference efficiency, but that is also unverified. The PR Newswire release from Baidu is the primary source for the announcement; read it alongside independent testing. The same caution applied to our Qwen4 Max preview post.
Availability and compliance checks before you build
Check availability, data residency, and whether the provider's terms fit your obligations to customers. Baidu is a Chinese company, and businesses in the UK, EU, UAE or Saudi Arabia should confirm where data is processed and what the terms say about retention. Our custom software cost guide for the UK and mobile app cost guide for the UAE make the same point.
What Coheser would do with it
At Coheser we build AI agents, AI automation and chatbots, so a near-free 280B model is interesting — as a tool to test. Start with low-risk, high-volume work where a wrong answer is easy to catch: drafting replies for a human to approve, summarising support threads, triaging messages. A WhatsApp chatbot that drafts replies for a human, or a voice agent in a supervised pilot, can run on a cheap model while a human stays in the loop. Measure accuracy, not just cost: if the model needs more human correction than the one it replaces, the saving is an illusion.
What to avoid for now
Do not put a preview model in front of customers unsupervised, let it take irreversible actions such as refunds, or rebuild a working system around it until the finished version is tested. Our guide to scaling a SaaS MVP without rebuilding argues the same: keep the model behind a clean interface so you can swap it later, and spend the saving on evaluation.
Frequently asked questions
Is ERNIE 5.5 Flash Preview really that cheap?
On Baidu's published list prices, yes: about $0.14 per million input tokens and $0.56 per million output tokens, with cached input far cheaper still. Total cost matters more: a cheap model needing more retries or correction can still cost more than a dearer one that gets it right first time.
Can I use it in my business today?
You can try it today through the Qianfan API or free on OpenRouter during the preview, but Baidu itself warns that some capabilities may be incomplete. That makes it suitable for testing and supervised pilots, not unsupervised customer-facing work yet. Check availability and terms for your market.
Will the weights really be open source?
Baidu has promised an Apache 2.0 release of the weights by 25 November 2026. Base weights are already on Hugging Face under a Baidu custom licence, which is not fully open, so watch for the Apache 2.0 release. Until it lands, treat self-hosting as a plan, not a fact.
Does it beat Gemini 4 Hydrogen or LongCat-2.5?
On price, yes: roughly half the price of LongCat-2.5 and roughly a tenth of Gemini 4 Hydrogen on the figures in our earlier posts. On quality, there is no independent comparison yet. Run your own test on your own tasks and compare the results blind.