Baidu launched ERNIE 4.5 Turbo and ERNIE X1 Turbo at Create 2025, its annual developer conference in Wuhan, on April 25, 2025. The announcement is primarily a pricing story: two models positioned to undercut the existing competitive field while claiming benchmark parity with GPT-4.1 and outperformance of DeepSeek-R1.
The framing at Create 2025 — themed "Models Lead, APPs Rule" — reflects Baidu CEO Robin Li's stated view that AI models and chips are worthless without good applications. The launch includes two models plus a multi-agent platform (Xinxiang) and a digital avatar tool (Huiboxing), signaling that Baidu is pushing the full stack, not just the model API.
What's New: Two Turbo Models at the Price Floor
ERNIE 4.5 Turbo is the general-purpose multimodal model. Key claims:
- Multimodal capabilities (text + image) on par with GPT-4.1, superior to GPT-4o, across multiple benchmarks
- Faster response, reduced hallucinations, improved logical reasoning and coding versus ERNIE 4.5
- 80% cheaper than ERNIE 4.5: input at RMB 0.8/M tokens (~$0.11/M), output at RMB 3.2/M tokens (~$0.44/M)
- Free on ERNIE Bot
ERNIE X1 Turbo is the reasoning model. Key claims:
- Outperforms DeepSeek-R1 and DeepSeek-V3 on reasoning benchmarks (per Baidu)
- Improved multimodal functions alongside reasoning
- 25% of DeepSeek R1's price: RMB 1/M input (~$0.14/M), RMB 4/M output (~$0.55/M)
- Half the price of the original ERNIE X1
ERNIE X1 (the base reasoning model) was first released in mid-March 2025, where Baidu claimed it matched DeepSeek-R1 performance at half the price. X1 Turbo improves performance further while cutting the price again to one-quarter of DeepSeek R1.
Architecture: 424B MoE with Separate Text and Vision Experts
The ERNIE 4.5 technical report (released June 2025) provides the architecture details for the model family backing the Turbo variants:
ERNIE 4.5 model family specs:
| Variant | Total Params | Active Params | Layers | Text Experts | Vision Experts |
|---|---|---|---|---|---|
| ERNIE-4.5-A47B | 424B | 47B | 54 | 64 total / 8 active | 64 total / 8 active |
| ERNIE-4.5-A3B | 28B | ~3B | 28 | 64 total / 6 active | 64 total / 6 active |
| ERNIE-4.5-0.3B | ~0.36B | — | 18 | — | — |
Context length: 131,072 tokens (128K) across all variants.
The defining architectural choice is separate expert pools for text and image/video tasks. Unlike standard MoE models that use the same expert pool for all modalities, ERNIE 4.5 maintains distinct expert groups — text experts activate for language tasks, vision experts activate for image/video inputs. Baidu's claim is that this eliminates cross-modal interference that degrades specialist performance.
Inference throughput (per ERNIE technical report): 56K input TPS (tokens per second) and 18K output TPS per H800 node — unusually high for a frontier-class model.
Note: The Turbo variants are production API products; the June 2025 open-source release of the ERNIE 4.5 base models added Apache 2.0 weights to HuggingFace. The Turbo models are API-only products.
Benchmarks: GPT-4.1 Parity Claims
Baidu's benchmark claims follow the standard Chinese lab playbook: compare against the Western frontier and show competitive or superior numbers on standard benchmarks, using Baidu's internal evaluation conditions.
ERNIE 4.5 Turbo (Baidu claims, multimodal):
- On par with GPT-4.1 across text and multimodal benchmarks
- Superior to GPT-4o on multiple benchmarks
- DocVQA: competitive with GPT-4.1 (specific numbers in official materials)
- MMLU: competitive with GPT-4.1
ERNIE X1 Turbo (Baidu claims, reasoning):
- Outperforms DeepSeek-R1 on reasoning benchmarks
- Outperforms DeepSeek-V3 on reasoning benchmarks
- These benchmark results are Baidu's own evaluations, not independent audits
The caveats matter. Baidu's benchmark methodology uses internal test conditions; third-party evaluations on the same benchmarks sometimes show different ordering. The more defensible claim is the pricing comparison — the RMB numbers are verifiable, and the pricing positions are clear.
Pricing comparison (per million tokens):
| Model | Input | Output | vs. DeepSeek R1 |
|---|---|---|---|
| ERNIE X1 Turbo | ~$0.14/M | ~$0.55/M | ~25% cost |
| DeepSeek-R1 | ~$0.55/M | ~$2.19/M | baseline |
| ERNIE 4.5 Turbo | ~$0.11/M | ~$0.44/M | — |
| GPT-4.1 | ~$2.00/M | ~$8.00/M | ~7× more expensive |
| Claude Sonnet 4.5 | ~$3.00/M | ~$15.00/M | ~11× more expensive |
At these prices, X1 Turbo is the cheapest reasoning API claiming DeepSeek-R1-class performance at April 2025 release. ERNIE 4.5 Turbo is cheaper than virtually every multimodal API with comparable benchmark claims.
Additional Launches: Xinxiang and Huiboxing
Beyond the models, Baidu announced two products at Create 2025:
Xinxiang — a multi-agent application for complex task resolution. Features 200 task types at launch (knowledge analysis, travel planning, office work), with plans to expand to 100,000+ task types and open developer access. Uses an enhanced version of Anthropic's Model Context Protocol (MCP) for integration. Available on Android; iOS in testing.
Huiboxing — a digital avatar platform that generates photorealistic avatars with natural voice from short video clips. Positioned for content creator and enterprise use cases.
These product launches signal the "APPs Rule" half of Robin Li's conference theme: Baidu is not just releasing model APIs but building consumer and enterprise products on top.
Access
Both models are available through:
- ERNIE Bot: free access for both models
- Baidu AI Cloud API: developer API at the listed pricing
ERNIE 4.5 base model weights (June 2025 release): Apache 2.0 on HuggingFace at ernie.baidu.com. Note that this is the base model, not the Turbo variant — Turbo remains API-only.
What It Means for Developers
The primary signal from the Turbo launch is pricing pressure on the Chinese model ecosystem. When Baidu can claim DeepSeek-R1 level reasoning at 25% of the price, the competitive floor drops further. This benefits developers building cost-sensitive applications — the cheapest viable option for production reasoning tasks just got cheaper.
ERNIE X1 Turbo vs alternatives for reasoning tasks:
- If cost is primary and you need chain-of-thought reasoning: X1 Turbo at ~$0.14/M input is the cheapest API in the category at launch
- If you need open-weight self-hosting: DeepSeek-R1 (open-weight) remains the benchmark; X1 Turbo is API-only
- If you need independent benchmark validation: X1 Turbo's benchmark numbers are Baidu self-reported; third-party validation was limited at April 2025 release
ERNIE 4.5 Turbo vs alternatives for multimodal tasks:
- At ~$0.11/M input claiming GPT-4.1 parity on multimodal: cheaper than every comparable API at release
- Trade-off: Chinese model ecosystem (Chinese language optimization, Baidu infrastructure), Baidu self-reported benchmarks
- The separate text/vision MoE expert architecture is a real differentiator for mixed-modality workflows
China-specific use cases: ERNIE models have deep Chinese language and cultural optimization that Western models lack. For applications targeting Chinese users or Chinese-language content, ERNIE 4.5 Turbo's combination of pricing and Chinese-language performance is a strong default choice.
Bottom Line
ERNIE 4.5 Turbo and X1 Turbo are Baidu's most aggressive pricing moves in the frontier model market. The 80% price reduction for ERNIE 4.5 Turbo and the 25%-of-DeepSeek-R1 pricing for X1 Turbo change the cost floor for production AI applications, assuming the benchmark claims hold under independent evaluation. The separate text/vision expert MoE architecture and 128K context window are genuine technical differentiators. Free access via ERNIE Bot means consumer-facing Chinese AI users have frontier-class reasoning at no cost.
Resources
- Baidu Create 2025 Announcement (PR Newswire) — official press release
- ERNIE 4.5 Open-Source — Apache 2.0 base model weights
- ERNIE Bot — free consumer access
- Baidu AI Cloud — developer API
ERNIE 4.5 Turbo and X1 Turbo are available via API at cloud.baidu.com and free on ERNIE Bot at yiyan.baidu.com. Pricing: ERNIE 4.5 Turbo at RMB 0.8/M in (~$0.11), ERNIE X1 Turbo at RMB 1/M in (~$0.14).
