"proactiveVideo" in public spec, not announced by Google
An unannounced setting, "proactiveVideo," has been found in the public specification of Google's "Gemini API" for developers. @aitrackerbot, a bot that detects API changes, reported it on X on Nov. 6 (5:26 a.m. Nov. 7 Japan time). It is a setting of the "Live API," which exchanges audio and video in real time, and could be a feature that lets the model react to images without waiting for instructions.
Google is adding "proactive video" to the Gemini Live API 👀 A new option in today's API schema lets the model react to incoming camera frames on its own, without a prompt. It requires proactive audio and warns of "a significant increase in cost". Not in the docs yet.
ORIENS confirmed on Nov. 7 that the field exists in the v1beta version of the "Discovery document," in which Google publishes the structure of the API for machines. Google has not announced it, and the feature has not been confirmed as available. ORIENS has not run tests with an API key, and it is not known whether specifying the field actually works.
Description says generation "may" occur for each image
According to the Discovery document, proactiveVideo is an optional field that can be enabled or disabled, described as allowing the system, when enabled, to start generation for each image received. It also states that enabling it requires enabling proactive_audio, and that the frequency of generation depends on the frequency of incoming images, so costs can increase significantly.
The wording is "may," and it does not say generation will always occur for every image. The bot's post mentions "camera frames" and "without a prompt," which are not in the description; the description's term is "image received." When the field was added could not be confirmed from the Discovery document either.
Not in official documentation or SDKs
proactiveVideo does not appear in the official developer documentation, changelog or official SDKs (development toolkits). In every document, the only field listed in the same setting is the audio-focused proactive_audio. The changelog runs through Nov. 6 and contains no Live API entries since October.
| Public document | Mentions proactiveVideo |
|---|---|
| Gemini API Discovery document (v1beta) | Yes |
| Live API reference | No |
| Live API capabilities guide | No |
| Gemini 3.8 Live model page | No |
| Gemini API changelog | No |
| Python SDK (main branch) | No |
| JavaScript SDK (main branch) | No |
| Google Cloud Live API reference | No |
Current documentation says it does not respond to video alone
Google's documentation for its robotics models says of the Live API that video frames alone do not trigger new inference. A text or audio instruction is needed to get a response, and to have the model make decisions proactively, it recommends implementing a "heartbeat" that periodically sends the latest frame and a short instruction.
If proactiveVideo is offered as described, receiving images would trigger generation. Whether it can replace this periodic sending that developers build themselves is unknown, as Google has not explained.
The existing proactive audio is a feature that ignores irrelevant speech or stays silent until asked, and Google describes it mainly in terms of deciding not to respond. The documentation does not say why it is required.
How much costs rise is unknown; no entry in price list
Paid-tier pricing for the current "Gemini 3.8 Live" is $0.75 for text input and $4.50 for output per 1 million tokens. Image and video input is listed at $1 or $0.002 per minute, and audio output at $12 or $0.018 per minute. The price list has no entry for proactiveVideo, and the cost per generation cannot be estimated.
Video input to the Live API is limited to one frame per second, so up to 60 images can arrive per minute. On its model page, Google notes that video frames are sent to the model by default and urges developers to send them only when needed to manage the amount of context and costs. The fact that the description mentions increased costs is a factor that will affect the budgets of businesses considering adoption.
Supported models and release timing unconfirmed
No document states which models are supported. According to Gemini API documentation, the prerequisite proactive audio is always enabled in Gemini 3.8 Live, configurable in "Gemini 2.5 Flash Live" and not supported in "Gemini 3.1 Flash Live Preview."
The situation for Gemini 3.8 Live is particularly unclear. Google Cloud's developer guide asks developers not to pass the proactivity field for that model, and proactiveVideo sits within this field. Until Google indicates the timing and scope of availability, whether it can be used cannot be confirmed.
OpenAI's GPT-Live 1 does not support video
The voice model "GPT-Live 1," released in September by rival OpenAI, is listed in official documentation as not supporting images or video.
For the company's earlier "Realtime API," a November 2025 post on its developer forum said the model would not start speaking when shown video without audio input. A reply under the name "juberti" explained that responses are started only by audio or manually. This is from a year ago, and the current specification could not be confirmed.
