Release guide · August 31, 2026
Gemini Omni 1.1 Flash is a video generator, not a long-video analyst
Google's Gemini Omni 1.1 Flash generates and edits short video. Here are the exact API ID, prices, limits, and why a long-video tool benchmark does not apply.

Gemini Omni 1.1 Flash is a paid, generally available video-generation and editing model with useful creative controls and token-based billing. It is not the 1M-context long-video analysis and tool-chaining model implied by early launch summaries: Google limits source video to 10 seconds, documents video-only output on the Gemini API, and does not support function calling on the Enterprise route.
Google released Gemini Omni 1.1 Flash on August 27, 2026 as a model for conversational video generation and editing. The stable Gemini API identifier is exactly gemini-omni-1.1-flash. Google lists it on the paid tier and exposes it through the Interactions API in Google AI Studio.
That sounds straightforward. The confusing part is the name. “Gemini,” “Omni,” and a 1,048,576-token context window can make the product look like a general multimodal reasoning model for ingesting long recordings, writing subtitles, and calling tools. Google’s own capability tables describe a much narrower product.
Sources checked October 28, 2026. This refresh adds the retired preview identifier, current workflow restrictions and a proposed short-shot evaluation. No Gemini video-generation call was made for this update; the earlier long-video benchmark remains not run.
The verified route and price
The Gemini API model page lists the stable model code as gemini-omni-1.1-flash. It accepts text, images, and up to ten seconds of video for editing or extension, and returns video. Output can run from three to ten seconds at 24 FPS in 360p, 720p, 1080p, or 4K.
The official Gemini API pricing page, checked October 28, 2026, lists no free tier. Standard paid pricing is:
| Meter | Price |
|---|---|
| Text, image, video, or audio input | $1.50 per 1M tokens |
| Text output, including thinking | $9.00 per 1M tokens |
| Video output | $17.50 per 1M tokens |
| 720p video, approximate effective rate | $0.10 per generated second |
Google says 720p video consumes 5,792 output tokens per second, making the token rate approximately $0.10 per generated second. A ten-second accepted 720p shot therefore has a list-price video-output component of about $1 before input tokens and failed attempts. The calculation is an illustration, not an observed bill. This refresh removes the earlier 360p, 1080p and 4K dollar figures because we could not re-verify them in the current pricing text. Check the selected resolution’s current billing details before approving a run; do not extrapolate its rate from 720p.
Two official routes use different identifiers
Google’s developer surfaces do not use one universal string. The stable Gemini API route uses gemini-omni-1.1-flash. The Gemini Enterprise Agent Platform model page lists gemini-omni-1.1-flash-preview and labels that route Preview.
Preserve those spellings literally. Do not add -preview to a Gemini API request because an Enterprise page uses it, and do not remove -preview from an Enterprise deployment. The Enterprise page lists the global region, fixed quota, a 131,072-token input limit, and a 57,920-token output limit. Those route-specific figures should not overwrite the Gemini API’s documented 1,048,576-token context window.
Check the old preview name before retrying
The current model page, updated October 22, marks gemini-omni-flash-preview as shut down. That old identifier is different from both the stable Gemini API ID and the Enterprise preview ID above.
If an existing integration still contains the retired name, inspect its provider, endpoint and saved configuration before changing anything. On the Gemini API route, use the documented stable ID and validate one request against the current guide. Keep the old configuration in version control and record the replacement. Do not treat repeated requests to a retired model as a quality or latency test.
What Omni 1.1 actually adds
Google’s launch post presents five practical additions:
- extend scenes in ten-second increments to a cumulative 40 seconds;
- use up to ten seconds of prior context during extension;
- interpolate between specified first and last frames;
- generate 360p drafts before rendering higher-resolution output;
- upscale output to 1080p or 4K and use short reference videos.
Google claims 360p drafts can be up to 60% faster than 720p and cost one third as much. That is a vendor throughput claim, not an independent latency result. The same distinction applies to Google’s descriptions of improved consistency, control, and production readiness.
Why the proposed long-video benchmark was stopped
We designed a fixed test around three roughly 30-minute English, Spanish, and Hindi videos. Each run would ask the model to extract subtitles and make a function call that writes a structured chapter map. The gate was intentionally checked before sourcing or uploading media.
The route failed that preflight on three independent constraints:
- Input duration: Google documents at most ten seconds of source video for editing and extension, not a 30-minute recording.
- Output type: the Gemini API model page documents video output rather than subtitle text output.
- Tools: the Enterprise capability table explicitly marks function calling as unsupported.
The 1,048,576-token context figure does not override those modality-specific limits. Uploading public-domain videos anyway would create spend without testing the stated hypothesis. Replacing them with ten-second clips would silently change the task. Using a different Gemini model would test another route. No benchmark call was made, and there is no measured WER, latency improvement, or cost-per-accepted-result to report.
The complete preflight, proposed prompt, tool schema, and acceptance gate are preserved under benchmarks/gemini-omni-1-1-flash/. The benchmark remains explicitly not run. It can only be activated if Google documents a route that accepts the full fixtures, returns subtitle text, and supports the required function call.
Data use and deployment caveats
The model is paid-only in the Gemini API pricing table. Google’s Gemini API terms say paid-service prompts and responses are not used to improve its products, though safety and operational logging still applies. Region eligibility remains governed by Google’s available-regions list. The Enterprise page is internally inconsistent: its consumption table marks Pay-as-you-go unsupported, while its regional section lists Standard PayGo in global. Fixed quota is listed as supported. Confirm the purchasing route in your project before budgeting; this page does not establish that PayGo is available to your account.
Google’s rate-limit documentation makes Gemini API limits project- and tier-specific: the active RPM, TPM, and RPD values must be read in AI Studio for the billing project that will run the evaluation. No account-specific quota was inspected for this refresh, so no active rate-limit value is claimed.
For production evaluation, keep private footage out of exploratory accounts, record which route and billing project handled each request, and treat generated clips as unaccepted until a human checks continuity, prompt adherence, audio, safety, and unwanted artifacts.
Check editing restrictions before uploading
The generation and editing guide, checked October 28, distinguishes uploads from model-generated clips:
- Editing or extending uploaded video is unavailable in the EEA, Switzerland and UK; editing or extending model-generated video is supported there.
- Uploaded editing/extension input is limited to ten seconds. Multi-turn extension of generated video is a separate workflow.
- Reference videos have a different limit: at most three clips, each up to three seconds. Their audio is ignored.
- Audio-reference uploads and voice editing are unsupported. A generated video can still include audio.
- Setting
store=falseprevents later editing throughprevious_interaction_id. Decide whether you need a follow-up edit before the first request.
This matters for an agent wrapper: a successful initial generation does not prove that the requested follow-up operation is supported. Save the intended operation alongside the model ID, and stop when the documented restriction excludes it.
Try one bounded short-shot evaluation
This is an authored exercise, not a tested integration or a benchmark result. Use the official generation guide for the current request syntax. An agent can prepare the prompt and record results; a person should approve the spend and inspect the clip.
- Define the shot. Use a synthetic scene with an easy visual check: “A single blue wooden cube slides from left to right across a cream tabletop. Locked camera, one continuous shot, no text.” Before generation, require one cube, stable color, the requested direction and no scene cut.
- Freeze the setup. Record the provider, literal model ID, SDK version, prompt and selected duration/resolution. Start with one 360p draft. Choose a total spend ceiling and allow at most one deliberate retry; do not enable automatic retry loops.
- Review the output. Download the result, confirm it plays, inspect its duration and resolution, and watch the whole clip. Check the four written criteria and listen for unwanted audio. Record pass/fail with a reason even if the result looks attractive.
- Make one controlled edit. Only if the first shot passes and the budget permits, ask to change the cube to ochre while preserving the camera and motion. Use the guide’s conversational-edit workflow. Compare both clips and reject changes to features meant to remain fixed.
- Stop and account for attempts. Save request identifiers, elapsed time, usage and actual billing when available. If billing is pending, label the cost unknown. Upscaling is a separate paid decision after acceptance.
Use a small run record:
shot_id, provider, model_id, sdk_version, operation,
requested_seconds, actual_seconds, resolution, prompt,
attempts, elapsed_seconds, accepted, rejection_reason,
usage_tokens, billed_cost_usd, billing_status
Calculate cost per accepted shot as total billed run cost, including rejected attempts, divided by accepted shots. If none passes, report zero accepted shots and leave that ratio undefined. Keep review time separate from API latency. These records support a workflow decision; one attractive clip does not establish general model reliability.
Recover without hiding failures
| Symptom | Next check |
|---|---|
| Old configuration cannot reach the model | Compare the exact ID and provider with the two route pages; check for the retired preview name. |
| Follow-up edit cannot use the earlier interaction | Check whether the original request disabled storage and whether you retained its interaction ID. |
| Uploaded-video operation is rejected | Check region, operation and clip duration against the guide before uploading again. |
429 RESOURCE_EXHAUSTED | Inspect active project quotas and spend limits in AI Studio; pause instead of retrying continuously. |
| Clip renders but fails the shot criteria | Record a quality failure, then change one prompt requirement within the retry budget. |
Continue with choosing where your AI model should run to document data access, operating effort and the reason to keep or change a route.
The buying decision
Gemini Omni 1.1 Flash is worth testing for short-form video generation, editing, interpolation, and iterative scene extension. The 360p draft tier gives teams a sensible way to reject weak concepts before paying for larger output.
It should not be selected for multilingual long-video transcription or tool-chained chapter extraction. Those are reasonable Gemini-family workloads, but they are not supported by this model’s documented interface. The honest verdict is an interface mismatch, not a failed quality score.
Sources and checked dates
Checked October 28, 2026. The launch post is dated August 27; the model page was updated October 22, the Enterprise page October 28 and the rate-limit page October 26. Documentation checks are not runtime verification.
- Google AI for Developers: Gemini Omni Flash model page
- Google AI for Developers: Gemini Omni generation and editing guide
- Google AI for Developers: Gemini API pricing
- Google: Build with Gemini Omni 1.1 Flash
- Google Cloud: Gemini Omni 1.1 Flash Preview
- Gemini API Additional Terms of Service
- Gemini API available regions
- Gemini API rate limits
Put this to work
Separate a model's context-window number from its modality-specific input and output limits.
Try
Prototype one 360p shot before paying for 720p, 1080p, or 4K output.
Prove it worked
Record the exact route, resolution, generated seconds, retries, accepted output, and billed tokens for every test.
Where it can pay
For production video workflows, compare cost per accepted shot after review rather than cost per raw generation.
Keep in view
- Use the literal Gemini API model ID `gemini-omni-1.1-flash`; the Enterprise Agent Platform documents a different preview ID.
- Budget by generated video duration and resolution, not by assuming ordinary text-model economics.
- Do not benchmark subtitle extraction or tool chaining against this route: the documented interfaces do not support that workload.