by Baidu
ERNIE-5.5-Flash-Preview pricing, plans and limits
A hosted preview in Baidu's ERNIE 5.5 Flash line, sized at 280B total and 14B active parameters.
- Best for
- Long-context agent calls on a hosted API
- Not for
- Work that needs open weights
Should you use ERNIE-5.5-Flash-Preview?
Baidu put ERNIE-5.5-Flash-Preview on OpenRouter on 10 November 2026, where it measured 203 tokens per second at a 1.44-second median latency. It suits developers who want a fast long-context API model with tool calling, and it competes with MiMo-V2.6-Flash and GLM-5.3-Flash, though Baidu has published no benchmark scores for it.
Best for
- Long-context agent calls on a hosted API OpenRouter lists a 1,024,000-token window, tool calling and JSON-schema output for the model.
- Latency-sensitive text generation OpenRouter's median of 203 tokens per second and 1.44 seconds latency, read 11 November 2026, support Baidu's speed claim.
Not for
- Work that needs open weights The weights link on OpenRouter returned a 404, and Baidu's Hugging Face lists only Base repositories.
- Decisions that need published benchmarks Baidu has published no scores for the Preview on any page read.
- Image, audio or video input OpenRouter lists text as the only input.
Can you use ERNIE-5.5-Flash-Preview?
- OpenRouter (Baidu Qianfan host): preview (Pay per token; one host, Baidu Qianfan). Listed as ERNIE-5.5-Flash-Preview.
- Free plan: not published.
Checked 11 Nov against Baidu's own pages.
ERNIE-5.5-Flash-Preview pricing
- Input, per 1M tokens
- $0.28/1M
- Output, per 1M tokens
- $1.13/1M
- Cached input, per 1M tokens
- $0.03/1M
- Pricing model
- per-token
- Free tier
- No
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0084 | $0.0011 | $0.0095 |
| Support reply | $0.0006 | $0.0003 | $0.0009 |
| One coding agent run | $0.056 | $0.023 | $0.079 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Qianfan charges 0.28 dollars in and 1.13 out per million tokens on OpenRouter, the only host, and 0.028 for cached reads. Hokai found no separate Baidu rate card.
What ERNIE-5.5-Flash-Preview can do
ERNIE-5.5-Flash-Preview is a text-in, text-out mixture-of-experts language model from Baidu with 280 billion total parameters and 14 billion active per token, listed in preview on 10 November 2026. It offers a 1,024,000-token context window through Baidu Qianfan on OpenRouter, with function calling and structured output.
What goes in and what comes out
Reads
- text
- tool-calls
Writes
- text
- tool-calls
Capabilities
- Function callingSupported
- Structured outputSupported
- Tool useSupported
- Audio inputNot supported
- Audio outputNot supported
- Video inputNot supported
- VisionNot supported
- Reasoning modes
- configurable reasoning effort
Key features
Reasoning traces on request
OpenRouter's parameter list includes include_reasoning, so reasoning output can be returned alongside the answer.
Uptime on the only host
OpenRouter showed 99.99% reachability and 99.95% availability over the three days before this page was checked.
Short list of sampling controls
temperature, top_p and max_tokens are accepted; OpenRouter's list shows no seed, stop or penalty parameters for this model.
Base architecture as a reference
Baidu's Base page describes the ERNIE 5.5 Flash design as 43 layers with hybrid sparse attention and 512 routed experts, 10 active; Baidu does not say the Preview uses it unchanged.
Who it is for
- Developers who want a fast hosted model with a very long context window
- Teams already routing through OpenRouter that want a Baidu model in the mix
- Evaluators comparing Chinese-lab mixture-of-experts models
ERNIE-5.5-Flash-Preview benchmarks
Standout result
Total and active parameters, per Baidu's listing text
Source: OpenRouter model page and API, read 11 Nov 2026Where it sits
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.492/M | #29 / 104 | peer median $1.70/M |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Strengths and limits
Strengths
- Speed backs Baidu's claim: OpenRouter's median output speed sits above the 200 tokens per second Baidu states.
- OpenRouter's API lets one completion use 90% of the context window, so very long generated documents or code fit in one call.
- Tool calling and JSON-schema output are both accepted, per OpenRouter's parameter list, which suits agent loops.
- OpenRouter lists a 1,024,000-token context window and function calling, read 11 November 2026.
- OpenRouter measured a median 203 tokens per second on the only host, against Baidu's claim of more than 200.
Limits
- No benchmark scores, system card or data policy were found on any Baidu page, so quality is unproven.
- OpenRouter shows a single host, Baidu Qianfan, so there is no provider failover on that route.
- Text only, so teams that need image, audio or video input need a second model.
ERNIE-5.5-Flash-Preview alternatives
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| ERNIE-5.5-Flash-Preview This page | Baidu | 1M | $0.28 / $1.13 | text → text |
 MiMo-V2.6-FlashPick ERNIE-5.5-Flash-Preview for a Baidu-hosted API with 14B active parameters; pick MiMo-V2.6-Flash if you need open weights, listed on Hokai at 309B total and 15B active. | Xiaomi Corporation | 1M | $0.14 / $0.28 | text, image, video, audio → text |
 GLM-5.3-FlashPick ERNIE-5.5-Flash-Preview to try Baidu's newest Flash line on a hosted route; pick GLM-5.3-Flash for an open-source model listed on Hokai at 320B total and 18B active. | Z.ai | 1M | $0.15 / $0.50 | text, image, video, pdf → text |
 Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
ERNIE-5.5-Flash-Preview specs
- Context window
- 1.0M tokens
- Max output
- 922K tokens
- Latency (P50)
- 1440 ms
- Tokens per second
- 203 tok/s
- Architecture
- Sparse Mixture-of-Experts (MoE)
- Mixture of experts
- Yes
- Parameters
- 280B total, 14B active per token
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| OpenRouter | https://openrouter.ai/baidu/ernie-5.5-flash-preview | API key |
| Baidu Qianfan | https://cloud.baidu.com/product-s/qianfan_home | API key |
Known quirks
Show 3 known quirks
Output limit differs by page
OpenRouter's API lists a maximum completion of 921,600 tokens while its FAQ text says up to 1,024,000.
Workaround: Keep max_tokens at or below 921,600 and check the finish reason on long generations.
Weights link returns a 404
OpenRouter links baidu/ERNIE-5.5-Flash-Preview on Hugging Face, but the page returned a 404 on 11 November 2026.
Workaround: Use the hosted API, or the separate ERNIE-5.5-Flash-Base checkpoint if you need weights, knowing it is pretrain-only.
One host, no failover
OpenRouter says the model is hosted by one provider and forwards every request to it.
Workaround: Add a second model to the fallback list in your request so an outage on Qianfan does not stop your job.
ERNIE-5.5-Flash-Preview facts
- Deployment
- API · OpenRouter · Baidu Qianfan
- Licence
- Proprietary (API access through Baidu Qianfan)
- Released
- 10 Nov 2026
- Current version
- ERNIE-5.5-Flash-Preview
- Family
- ERNIE 5.5
- Variant
- Chat
Version history
ERNIE-5.5-Flash-Preview · 10 Nov 2026
Listed on OpenRouter, served by Baidu Qianfan.
What changed
- Release ERNIE-5.5-Flash-Preview listed on OpenRouter source
Questions about ERNIE-5.5-Flash-Preview
What does ERNIE-5.5-Flash-Preview cost per million tokens?
OpenRouter lists one host, Baidu Qianfan, read on 11 November 2026: $0.28 per 1M input tokens and $1.13 per 1M output tokens, with cache reads at $0.028 per 1M. Hokai found no separate Baidu rate card for the Preview, so these OpenRouter prices are the ones on this page.
Can I download the weights of ERNIE-5.5-Flash-Preview?
Not as far as Hokai could find. OpenRouter links a Hugging Face page for the model, which returned a 404 on 11 November 2026, and Baidu's Hugging Face organisation lists only the two ERNIE-5.5-Flash-Base repositories. No free tier was listed on the pages read.
Is ERNIE-5.5-Flash-Preview the same model as ERNIE-5.5-Flash-Base?
Hokai found no Baidu page that says so. Baidu released the Base on 5 November 2026 as a pretrain-only checkpoint trained on 33.5 trillion tokens, with a 128K training context and 1M-token inference support. The Preview appeared on OpenRouter five days later with the same 280B total and 14B active sizing.
How much context and output does ERNIE-5.5-Flash-Preview allow?
OpenRouter's API lists a context window of 1,024,000 tokens and a maximum completion of 921,600 tokens. OpenRouter's FAQ text says completions can reach 1,024,000 tokens, and Baidu's description says 1 million tokens of context. Plan around the API figures.
Does ERNIE-5.5-Flash-Preview support tool calling and structured output?
Yes, per OpenRouter's parameter list read on 11 November 2026: it accepts tools and tool_choice for function calling, a JSON-schema response format, and a reasoning setting with an effort level. Input is text only.
About ERNIE-5.5-Flash-Preview
ERNIE-5.5-Flash-Preview is a preview language model from Baidu, listed on OpenRouter on 10 November 2026. Baidu's listing text, as OpenRouter prints it, calls it the latest foundation model in the ERNIE family, with improvements in coding, agentic capabilities, knowledge work and deep research. Its size, per the same text, is a sparse mixture-of-experts design with 280 billion parameters in total and 14 billion activated for each token. It is separate from ERNIE-5.5-Flash-Base, the pretrain-only checkpoint Baidu published on 5 November, and Hokai found no Baidu page that says how the two relate.
Read the full background
Baidu has published no benchmark scores for the Preview on any page Hokai read, so this record carries none. What exists is two measurements from OpenRouter, read on 11 November 2026: a median output speed of 203 tokens per second and a median latency of 1.44 seconds, both from Baidu Qianfan, the only host serving the model. Baidu's own listing text says generation runs at more than 200 tokens per second.
The context window is 1,024,000 tokens according to OpenRouter's API, which also lists a maximum completion of 921,600 tokens. A separate FAQ line on the same page gives a higher ceiling for completions, so budget with the API numbers.
The model takes text in and gives text out. OpenRouter's API lists tools and tool_choice for function calling, JSON-schema structured outputs, plus a configurable reasoning effort. It accepts no image, audio or video input.
The only listed route is Baidu Qianfan, Baidu's cloud model service, reached through OpenRouter. OpenRouter's weights link leads to a missing Hugging Face page, and the only ERNIE 5.5 Flash weights Baidu has published are the Base checkpoints. Treat the Preview as API-only until Baidu says otherwise.
Teams that need open weights today should look at GLM-5.3-Flash or MiMo-V2.6-Flash, open-source models in the same size class on Hokai, or at Hy3. Teams that need published benchmark evidence before choosing should wait, because Baidu has released none for the Preview. Anything with images, audio or video as input needs another model.
Baidu publishes no training cutoff, data policy or safety card for the Preview on the pages Hokai read. For the Base checkpoint, Baidu's release page says pre-training used 33.5 trillion tokens; that figure describes the Base, not the Preview.
Sources and guides
Official sources
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Baidu's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Lists that include ERNIE-5.5-Flash-Preview
The provider
More AI Models on Hokai
- Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.
Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens.
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.
JevTypeSafe's first System One Model, launched in 2026: skips text generation for typed, calibrated decisions software can act on directly.