Baidu's ERNIE is one brand covering three different things: a research programme running since 2019, a family of foundation models, and a consumer chatbot that has changed both its name and its price since launch. Untangling them is the first step in deciding whether any of it is useful to you.
- Consumer chatbot has been free since Baidu dropped the paid tier in March 2025
- Qianfan API pricing of ¥4/¥18 per million tokens at the base tier is very cheap for a flagship
- ERNIE 5.1 offers a 128K context with up to 65,536 output tokens in one call
- ERNIE 4.5 weights — including the 300B-A47B and 424B-A47B VL models — are Apache 2.0 on Hugging Face
- Ranks 5th globally and 1st among Chinese models on the arena.ai search leaderboard
- ERNIE 5.0 and 5.1 weights have never been released; the licence is proprietary and self-hosting is impossible
- Baidu publishes no absolute parameter count for ERNIE 5.1, only a ratio against ERNIE 5.0
- Every headline benchmark is vendor-reported; Artificial Analysis does not track any Baidu model and OpenRouter does not serve ERNIE 5.x
- Qianfan bills in yuan only, with no USD rate card and a China-centric onboarding path
- Mainland content rules apply; refusals on political topics are well documented by the BBC, Reuters and AFP
- Mainland China users wanting a free flagship chatbot
- Teams already inside Baidu AI Cloud
- Search-grounded and agentic Chinese-language workloads
- Developers who can use Apache 2.0 ERNIE 4.5 weights
- You need to download or self-host the current flagship
- You require independent benchmark verification
- You bill in USD or operate outside China
- You need uncensored discussion of mainland-sensitive topics
Pricing
Free access to the latest ERNIE model at ernie.baidu.com and in the Wenxiaoyan app. Baidu retired the paid consumer subscription in March 2025 alongside the ERNIE 4.5 launch.
Per Baidu's Qianfan billing page as of July 2026: ¥0.004 per 1,000 input and ¥0.018 per 1,000 output tokens for prompts up to 32K, rising to ¥0.006 / ¥0.022 for prompts between 32K and 128K. Yuan only — Baidu publishes no USD rate card.
The ERNIE 4.5 family, including ERNIE-4.5-300B-A47B-PT and the 424B-A47B vision model, is downloadable under Apache 2.0. ERNIE-Image is Apache 2.0 too. No ERNIE 5.x weights exist.
Names, and what happened to the paid tier
The chatbot launched for invited testing on 16 March 2023 and opened to the public on 31 August 2023 after Chinese regulatory approval. Baidu reported 100 million users by December 2023, 200 million by April 2024 and 300 million by June 2024. In September 2024 the Chinese consumer app was renamed from 文心一言 (Wenxin Yiyan) to 文小言 (Wenxiaoyan), repositioning it as a search assistant rather than a ChatGPT clone. English coverage still calls it ERNIE Bot; Baidu's own English blog links to it as ERNIE Bot too.
The pricing history matters more. Baidu launched a paid ERNIE 4.0 professional tier in November 2023, one of the first Chinese chatbot subscriptions. It killed it in March 2025: the announcement of ERNIE 4.5 and the ERNIE X1 reasoning model carried the headline "Makes ERNIE Bot Free Ahead of Schedule". Since then the consumer product has been free, and any page still quoting a monthly ERNIE Pro fee is out of date.
The 5.x flagships
ERNIE 5.0 arrived in early 2026, pitched as a 2.4-trillion-parameter unified multimodal foundation model with an activation rate under 3%. ERNIE 5.1 followed on 9 May 2026, and it is a compression rather than a scale-up: Baidu says it inherits ERNIE 5.0's pre-training foundation while cutting total parameters to roughly one third and active parameters to roughly one half, using a "Multi-Dimensional Elastic Pre-Training" scheme with elastic depth, elastic expert capacity and variable top-k routing sparsity. The headline claim is that pre-training cost about 6% of comparable models at the same scale.
Baidu has never published an absolute parameter count for ERNIE 5.1 — only that ratio. Third-party sites quoting a specific figure are doing arithmetic on ERNIE 5.0's number, not citing Baidu. What is documented, from the Qianfan model list, is a 128K context window and a maximum output range of 1 to 65,536 tokens, and a classification under text generation rather than multimodal. Image generation ships separately as ERNIE-Image. For the record, there is no model called ERNIE X2; the reasoning line runs X1, X1 Turbo, X1.1.
The open-source question
This is where ERNIE diverges sharply from the rest of the Chinese frontier. DeepSeek, Qwen, GLM, Kimi, MiniMax and Tencent all ship downloadable checkpoints for their current flagships. Baidu open-sources the previous generation only.
What you can download today is the ERNIE 4.5 family, published on Hugging Face under Apache 2.0 in June 2025 — ten models including ERNIE-4.5-300B-A47B-PT and the ERNIE-4.5-VL-424B-A47B-PT vision model, with an ERNIE-4.5-21B-A3B-Thinking variant added in September 2025. Outside the language line, Baidu does ship current-generation weights: ERNIE-Image, an 8B text-to-image model, went up under Apache 2.0 in April 2026, and the company's most-downloaded Hub asset by a wide margin is its OCR work rather than any ERNIE flagship.
ERNIE 5.0 and 5.1 weights have never been released. If your requirement is inspecting, adapting or hosting the model you use, that requirement rules out Baidu's current flagship entirely.
Benchmarks worth reading carefully
Baidu's release post is unusually specific in places and unusually vague in others. Specific: ERNIE 5.1 scores 99.6 on AIME26 with tool use, "second only to Gemini 3.1 Pro", and surpasses DeepSeek-V4-Pro on τ³-bench and SpreadsheetBench-Verified. Vague: on GPQA and MMLU-Pro it "approaches the performance of leading closed-source models", with no number given, and the comparison chart is published as an image rather than a table.
The independent picture comes from arena.ai — the former LMArena, rebranded in January 2026 — which is the only third-party leaderboard carrying the model. In a snapshot dated 1 July 2026, ernie-5.1 sat at rank 27 on the text board with 1468±5 from 32,742 votes, behind three other Chinese entries, and at rank 5 on the search board with 1227±10. Baidu's launch-day claim of first place among Chinese models still holds on search; it does not hold on text. Artificial Analysis does not track any Baidu model, so there is no neutral intelligence-index, latency or throughput measurement to check the efficiency story against.
The search result is the one that should carry weight. It matches Baidu's core business, it matches the way the product is actually used, and it is the metric on which ERNIE genuinely competes with anyone.
Access, cost and constraints
Three routes exist. The free consumer chat at ernie.baidu.com. The Baidu AI Studio playground. And the Qianfan platform, Baidu's equivalent of Bedrock or Vertex AI, where inference is metered per 1,000 tokens with prompt-length tiers — ¥0.004 in and ¥0.018 out below 32K, ¥0.006 and ¥0.022 between 32K and 128K, as of the July 2026 rate card. Everything is denominated in yuan. Baidu says ERNIE 5.1 is also rolling out across more than ten creative production agent platforms, including ISEKAI ZERO and Mulan AI.
Two constraints frame all of it. First, geography: onboarding assumes a mainland account and mainland billing, and the documentation is largely Chinese-only. Second, content regulation. ERNIE operates under mainland Chinese rules, and journalists at the BBC, Reuters, The Economist and AFP have repeatedly documented refusals and deflections on Xi Jinping, the 1989 Tiananmen Square protests, Xinjiang and the Hong Kong protests. That is a structural feature of a licensed Chinese consumer AI service.
If you are building for mainland Chinese users, ERNIE is one of the two or three products you have to evaluate, and the free consumer tier plus cheap Qianfan tokens make that evaluation nearly costless. If you are anywhere else, or if you need weights you can hold, the case largely evaporates.
