ERNIE (Wenxin) is Baidu's foundation-model family, currently led by ERNIE 5.1, which Baidu says was trained for about 6% of the pre-training cost of comparable models and ranked first among Chinese models on LMArena's Search Arena in May 2026. Developers reach it through Baidu's Qianfan platform, which serves ERNIE alongside DeepSeek, Kimi, GLM and Qwen behind one OpenAI-compatible API.
Baidu Qianfan pay-as-you-go, list updated 27 August 2026: ERNIE 5.1 CN¥0.004 per 1k input tokens and CN¥0.018 per 1k output for prompts up to 32k (CN¥4/M and CN¥18/M, about $0.56 and $2.54), rising to CN¥6/M and CN¥22/M between 32k and 128k; ERNIE 5.0 CN¥6/M input and CN¥24/M output; ERNIE 4.5 Turbo (32K and 128K) CN¥0.8/M input, CN¥3.2/M output, CN¥0.2/M on cache hits and CN¥4 per triggered search-augmentation call, with batch inference at CN¥0.32/M input and CN¥1.28/M output; ERNIE 4.5 Turbo VL CN¥3/M input and CN¥9/M output. New mainland accounts get a CN¥20 voucher after real-name verification; no USD rate card is published on the international site.
Best for teams that need a top-ranked Chinese-language model with a genuine self-host escape hatch in the open-weight ERNIE 4.5 releases, and who can tolerate reading a Chinese rate card — pick it over Doubao when Chinese-language quality and open weights matter more than multimodal breadth.
Visit ERNIE