Baidu ERNIE Lightning and Wizshot: API access, pricing and limits
Baidu’s ERNIE-Light-DT-1.0 and ERNIE-Wizshot-ME are available through Qianfan for two different jobs: long-context text reasoning and short portrait videos. Their access routes, billing units and input requirements matter more than the shared ERNIE name.
Baidu’s launch page lists both alongside the separate ERNIE 5.5 Flash Preview. These are not interchangeable endpoints, and the Preview model’s free OpenRouter offer does not make Lightning or Wizshot free.
Lightning: a million-token text model with paid access
The Qianfan model list identifies ernie-light-dt-1.0 with a 1M-token context, a maximum 128K-token output and default limits of 60 requests and one million tokens per minute. Its output allowance includes reasoning as well as the final answer. Those are service limits, not proof that every long-document task will succeed.
Baidu describes sparse attention and feedback-driven history compression as ways to reduce memory and computation for longer conversations. Its performance claims remain vendor claims. We have not benchmarked the model.
Qianfan’s billing table lists CNY 8 per million input tokens, CNY 0.80 for cache-hit input and CNY 28 per million output tokens. The documentation publishes these amounts per thousand tokens; the figures here multiply them by 1,000. Baidu’s product page says Lightning has one rate across the day, rather than the peak and off-peak schedule advertised for Flash Preview.
For another access route, OpenRouter lists Baidu Qianfan as the provider for baidu/ernie-light-dt-1.0-intl, at $1.13 per million input tokens, $0.113 for cache reads and $3.94 for output. These are that service’s listed dollar prices, not a currency conversion we calculated. Its listing supports function calling and JSON-schema structured output, but excludes EU in-region routing. That does not mean all European users are barred.
Thinking controls need model-specific settings
Baidu’s reasoning guide lists Lightning among models that can enable or disable thinking. Its specific entry for reasoning_effort gives max as the default, with low, high and max as options. Do not assume the page’s general default-high description applies to this model.
For a useful trial, keep the document set and question constant while changing the reasoning setting. Compare the answer’s citations, latency and total billed output, not just its length. This is a suggested evaluation method, not a test ByteForward has run.
Wizshot: portrait video, not a text-only endpoint
Baidu markets ERNIE-Wizshot-ME for 1080p clips lasting four to 15 seconds, with an emphasis on facial expressions in close-up scenes. Qianfan lists CNY 0.70 per generated video second. At that rate, a five-second clip costs CNY 3.50 before any separate services.
The creation API documentation shows the preparation involved. Requests require character portraits, reference voice audio and structured shot and dialogue information. Reference audio must be WAV, contain speech and last 10 to 60 seconds. The service accepts up to ten characters but recommends no more than three, and up to ten shots but recommends no more than four.
The advertised duration and API validation are not identical: the product page describes four to 15 seconds, while the API checks synthesized dialogue plus silence against a four-to-16-second interval. Work within the advertised range rather than treating the extra second as a guaranteed creative option. Generation uses an authenticated create request followed by a separate task-status lookup.
Use portraits and voices you have permission to use, and review the result before publishing. A generated performance should not be presented as footage of a real conversation.
Availability dates and weights are separate questions
The official model list and Wizshot creation documentation are dated November 11, 2026. OpenRouter labels Lightning’s release November 10. Those dates do not establish one simultaneous worldwide launch, so this guide does not treat November 12 as the first day either API existed.
OpenRouter describes Lightning as drawing on open-source GLM models and Baidu Search research. That description is not a downloadable Lightning release or a license for its weights. Start with the documented hosted service, and verify a specific repository and license before planning local deployment.
Featured image: AI generated editorial illustration by ByteForward, referencing the official Baidu mark. Conceptual artwork, not a product demonstration. No endorsement implied.
