Gemini Omni API
Gemini Omni 多模态视频模型，支持文生、图生、参考生成和视频编辑。
What is Gemini Omni
Gemini Omni is Google's multimodal video model family announced at I/O 2026 (as relayed by upstream EvoLink documentation), with Omni Flash as its short-form, multi-input, conversationally editable route. Roko exposes it through the unified async video task API with four routes — text-to-video, image-to-video, reference-to-video, and video-edit — mapped to gemini-omni-flash-* submodels.
Unlike per-second models such as Veo, Gemini Omni bills in tokens: $1.43 per million input tokens, $16.50 per million video-output tokens, and $8.47 per million other-output tokens. Actual cost depends on the usage returned per task, so you cannot budget from a fixed unit price — calibrate with small batches before scaling.
Current output is 720p at 3–10 seconds or auto (model-decided; reservations estimate 10 seconds), 16:9, 9:16, or auto aspect ratio, with native audio always generated. The workflow differentiator is conversational editing: the video-edit route accepts one MP4 up to 10 seconds and rewrites the footage from a natural-language instruction.
Model pricing
Pricing follows the current production configuration and the billing dimensions supported by each route.
| Route | Price | Billing rule |
|---|---|---|
| Gemini Omni | Calculated from task parameters on the pricing page | 输入 token $1.43/M；视频输出 token $16.50/M；其他输出 token $8.47/M |
Cost estimate examples
Token billing has no fixed unit price; the examples below only demonstrate the calculation. Assumed volumes are not measured — actual charges follow each task's usage and the pricing page.
input tokens/1M × $1.43 + video-output tokens/1M × $16.50 + other-output tokens/1M × $8.47
The three terms map to input, video-frame tokens, and other output (such as text); exact reconciliation is possible once the task reaches its terminal state.
assume 2,000 input + 60,000 video-output + 500 other-output → 0.0029 + 0.99 + 0.0042 ≈ $1.00
Assumptions only illustrate the method. Usage varies with duration and scene complexity — calibrate with small real batches.
Model information
Only provider-documented routes, capabilities, and limits implemented by this gateway are listed.
gemini-omni-flash-text-to-video
- Limits
- —
- Duration
- 3–10s / auto
gemini-omni-flash-image-to-video
- Limits
- Images 1
- Duration
- 3–10s / auto
gemini-omni-flash-reference-to-video
- Limits
- Images 1–8
- Duration
- 3–10s / auto
gemini-omni-flash-video-edit
- Limits
- Videos 1
- Duration
- 3–10s / auto
Key capabilities
Four routes, one task API
text-to-video, image-to-video (1 image), reference-to-video (1–8 images), and video-edit (1 clip ≤10s) share the same async interface and task lifecycle — switching routes is just a different model parameter.
Conversational video editing
Describe the change in natural language (for example, "switch to golden-hour light") and the video-edit route rewrites the footage while preserving subject and scene. Whether identity and motion survive is something you must validate on your own assets.
Native audio always on
Audio (ambience, effects, dialogue) is generated in the same pass with no toggle parameter; its cost lands in the video-output and other-output token meters.
Callback and polling
Pass an HTTPS callback_url on submit and the upstream POSTs updates when the task reaches a terminal state; otherwise poll the task endpoint. Both match Roko's task lifecycle.
Main parameters
Parameters come from structured model configuration; unsupported or unconfigured fields are omitted.
modelstringSubmodel ID
Default:gemini-omni-flash-text-to-video / gemini-omni-flash-image-to-video / gemini-omni-flash-reference-to-video / gemini-omni-flash-video-editpromptstringGemini Omni 多模态视频模型，支持文生、图生、参考生成和视频编辑。
Default:RequireddurationintegerOutput video seconds
Default:3–10s / autoqualityenumCurrent billing
Default:720paspect_ratioenumCapability
Default:auto / 16:9 / 9:16image_urlsarrayImage to video: 1 · Reference to video: 1–8
Default:Varies by routevideo_urlsarrayVideo editing: 1
Default:Varies by routeCapabilities and usage guidance
Key features
- Text to video · gemini-omni-flash-text-to-video
- Image to video · gemini-omni-flash-image-to-video
- Reference to video · gemini-omni-flash-reference-to-video
- Video editing · gemini-omni-flash-video-edit
- 720p output
- 3–10s / auto Duration
- Generated audio output
- Audio-capable workflow
Limits
- Image to video: Images 1
- Reference to video: Images 1–8
- Video editing: Videos 1
- Capability source: https://evolink.ai/zh/gemini-omni
Limits and caveats
- 720p is the only current resolution; choose Veo 3.1 or Gemini Omni 1.1 Flash when you need 1080p/4K.
- Token billing means cost is not precisely predictable before submit; auto durations reserve as 10 seconds and settle on the terminal state.
- Audio is always generated with no off switch; pick another model when silence is required.
- No dedicated negative-prompt parameter — write exclusions into the prompt body.
- Upstream output links expire after 24 hours; bake rehosting into your pipeline, especially across time zones.
- Claims about Google announcing this family come from upstream EvoLink documentation and are not quoted from a Google page.
How to call it
- POST /v1/videos/tasks with model set to the route's submodel (gemini-omni-flash-text-to-video / image-to-video / reference-to-video / video-edit), plus prompt, duration (3–10 or auto), and aspect_ratio.
- Attach media per route: one image for image-to-video, 1–8 images for reference-to-video, or one MP4 up to 10 seconds via video_urls for video-edit.
- Poll GET /v1/tasks/{task_id} or receive the callback_url POST until the task reaches succeeded / failed.
- On success, read the video URL and rehost it promptly; upstream output links are valid for 24 hours.
- On failure, read the reason. This model has no separate negative-prompt parameter — write exclusions into the prompt and resubmit. A successful submit does not guarantee a successful generation.
Typical use cases
Short-form batch production
Input: topic prompt + 9:16 vertical; action: batch-submit 3–10s or auto durations; expectation: 720p clips with audio; limit: 720p only, and costs must be reconciled per task from usage.
Conversational re-editing of existing footage
Input: an MP4 up to 10 seconds plus an edit instruction; action: video-edit route; expectation: lighting, mood, or local elements rewritten while the subject holds; limit: fidelity varies by source — verify cheaply first.
Multi-asset character/scene compositing
Input: 1–8 reference images of characters, props, and scenes; action: reference-to-video; expectation: assets fused into one coherent shot; limit: 8 images is the ceiling, and consistency depends on inputs.
Choosing against other video models on Roko
A coarse split by billing model and workflow; final decisions should rely on each catalog and your own tests.
Gemini Omni vs Veo 3.1
Pick Gemini Omni for editing existing footage, multi-asset fusion, and conversational rewrites; pick Veo 3.1 for precomputable per-second pricing, 1080p/4K, and batch n.
Gemini Omni vs Gemini Omni 1.1 Flash
The 1.1 adds a video-extend route, 360p–4K resolution tiers, and a 5-second default, and accepts videos among reference assets; the first Omni is 720p-only with no extension route.
vs per-request channels (Grok)
The Grok channel's fixed per-request pricing is fully budgetable but spec-bound; Gemini Omni trades cost variance for editing and multi-asset workflows.
Frequently asked questions
How do I call Gemini Omni on Roko?
Submit via POST /v1/videos/tasks with model set to the route's submodel ID (for example gemini-omni-flash-text-to-video), then poll the task to a terminal state.
How much does one Gemini Omni video cost?
Token billing: $1.43/M input, $16.50/M video output, $8.47/M other output. Cost depends on task usage and cannot be quoted per clip; calibrate with small batches.
Which durations, resolutions, and ratios are supported?
3–10 seconds or auto (reserved as 10s); 720p; 16:9, 9:16, or auto. Subject to the current Roko catalog.
How is it different from Veo 3.1?
Gemini Omni focuses on conversational editing and multi-input workflows, billed in tokens, currently 720p-only; Veo 3.1 bills per second with predictable pricing, 4/6/8-second clips, and up to 4K.
What footage can the video-edit route take?
One video (per upstream docs, an MP4 up to 10 seconds) plus a natural-language edit instruction. Validate edit fidelity on your own assets.
How long do output links stay valid?
Upstream links are valid for 24 hours; rehost files to your own storage as soon as the task completes.
Can I disable audio or batch results?
No. Audio is always generated with no toggle, and the current catalog has no batch n parameter — one result per task.
Am I charged for failed tasks?
The terminal state governs: failed tasks settle per platform rules and generally do not bill successful-result fees; reservation differences are refunded or charged automatically. See Roko's billing docs.
Sources and verification scope
Family and workflow claims come from upstream EvoLink public docs; callable routes, parameters, and prices come from Roko's current production catalog. Reviewed: 2026-10-08.
- EvoLink: Gemini Omni Flash API model page Upstream provider page: modes, token billing, FAQ
- EvoLink docs: gemini-omni-flash-text-to-video Upstream API manual: duration, aspect ratio, 24h link validity
- Roko API video task docs Async submit, task query, and billing lifecycle on Roko
- Roko API model catalog Current routes, prices, and generation config
Models in this family
Explore other available variants from the same model family.
Other video models
Explore more models available through the unified asynchronous video task API.