Gemini Omni 1.1 Flash's 40-Second Scenes Are Four Separate Bills

The number in every headline this week is forty. Forty-second scenes from Google's new video model — which sounds like you type a prompt and get forty seconds back.

You do not. A single call to gemini-omni-1.1-flash returns three to ten seconds. Forty is what you reach by extending that clip three more times — four jobs in total, each one separately billed. Nobody lied — Google's own announcement describes the continuations plainly — but the number travelled a long way from that sentence, and it arrived at a lot of people as a per-generation figure.

This is the briefing we would want before putting a card down: what forty seconds costs, what the 360p tier is for, and the constraints buried in the docs — one of which blocks European developers from the headline feature entirely.

Forty seconds is four jobs, and four invoices

Video models bill per second of output, and extension carries no discount — a continuation costs what a fresh generation costs, so forty seconds prices out as four ten-second clips. At every tier:

Four times ten seconds at each per-second rate. Ours come from the live catalog; the model pages carry the current figure if this ages.

That Google column deserves a word, because the rates are unexpectedly hard to find. Two official pages carry the full ladder and a text search catches neither. In the launch announcement the prices sit inside a picture — legible to a human, invisible to Ctrl-F. On the Cloud pricing page they appear as tokens per second against a $17.50-per-million meter, so they do not look like prices until you multiply. The page you would open first, the API pricing docs, gives one rung of four: "an effective price of approximately $0.10 per second" at 720p.

Both agree, with a rounding wrinkle — the announcement shows a tidy $0.03 / $0.10 / $0.15 / $0.30, the token maths lands a little above each. Billing follows the tokens, so the column above does too; the launch post has the tier-by-tier working.

Five dollars forty for forty seconds of 1080p, six and change from Google. That is the sticker before a single alternative take — and you will take them, because the last extension is where drift shows up. Budget the cut you keep plus the two you bin and one finished forty-second sequence is sixteen dollars, not five.

The 360p row used to be the one we lost, and we printed that here until recently. It moved: our draft tier is now $0.0297 a second against Google's $0.0338, twelve percent under. Do not read much into it. fal sits at $0.03 — a one percent gap, statistical noise dressed as a win, and the sort of thing that flips when somebody edits a multiplier. The draft workflow below holds whoever you buy it from; it has never depended on the price being ours.

The real 1.1 upgrade is memory, not length

Length was roughly reachable before. What makes it usable is what the model sees when it continues.

The June preview read the final second of the clip it was extending. One second tells you what is in frame and nearly nothing about what was happening, so extensions drifted. The tracking shot stopped tracking. A jacket shifted a shade. Light rebuilt itself at every join.

Gemini Omni 1.1 Flash reads up to ten seconds of prior footage. Ten seconds is long enough to know the dolly was still travelling, which side the key light came from, and roughly what colour the coat was — which is why a four-part sequence now holds together at all. If someone tells you 1.1's headline feature is length, they have it backwards. Length was arithmetic. Memory is the model.

1080p and 4K are upscales, not renders

One parenthesis in Google's documentation changes what you should pay for. From the model reference, verbatim: "1080p output (upscaled)" and "4K output (upscaled)". The model renders below those resolutions and scales up. The top of the ladder is not more detail — it is a larger file carrying the same information.

Which recolours the 4K row above. Twelve dollars and change, from Google, for forty seconds of upscaled output. If a client spec demands 4K dimensions, pay it. If you picked 4K assuming it looks better, you are paying exactly three times Google's 720p rate for a resize you could run in ffmpeg.

Draft at 360p, finish at 720p

This is the part that saves money, and nobody put it in a headline.

360p costs $0.0297 a second here — a four-second draft is twelve cents, against $0.36 at 720p and $0.54 at 1080p. Google also claims 360p runs up to 60% faster, a vendor number rather than one we measured, though the direction is obviously right.

Now the arithmetic. Say a shot takes ten attempts before the motion is right — realistic for anything with a specific camera move.

• Ten passes straight at 1080p → $5.40
• Ten passes at 360p → $1.19, then one winner re-rendered at 1080p → $1.73

Sixty-eight percent off, for one extra generation at the end. Be honest about what a 360p pass can tell you: composition, camera movement, whether the model spliced in cuts you never asked for. Fine texture it cannot judge — not enough resolution to carry it.

Here is the same prompt at both tiers, same seed, four seconds each. Draft first:

Then the 720p render of the identical prompt and seed:

Both show a brass stopwatch on dark slate, camera dollying in from the left, warm key light raking the metal. The prompt for both, verbatim:

Editorial slow-motion macro shot of a brass mechanical stopwatch resting on a dark slate table, its second hand sweeping forward in a single continuous shot with no scene cuts. Camera performs a slow dolly in from the left, shallow depth of field, one warm amber key light raking across the metal, cool blue rim light behind. Fine dust drifts through the beam. Clean dark background, no text, no logos, no people. Sound design: a soft mechanical tick and low room tone, nothing else.

Twelve cents against thirty-six. The draft told us the dolly reads and the model held one shot — everything we needed before committing. Run your loop down there.

Extension only moves forward

Scene extension appends, and the shape of that limit decides how you plan a sequence. You cannot insert into the middle of a clip. You cannot extend backwards to build a lead-in. There is no "keep the ending, rebuild the first four seconds" — the capability runs one direction only. If shot three is wrong, everything downstream goes in the bin with it.

So front-load the effort. The first generation sets camera language, palette and lighting for everything downstream, and it is the only segment you can revise cheaply. By the third extension you are four jobs deep with no way to reach back.

Extension will not do one more thing. Hand it a video of somebody talking, ask for new dialogue, and it declines — Google restricts speech modification. Launch reporting reads that as a deliberate responsible-release hold rather than a missing feature; that framing is the reporter's, but the restriction is in the docs. Do not build a dubbing product on this model.

European developers, read this line twice

Buried in the docs as a single sentence: in the EEA, Switzerland and the United Kingdom, editing and extending an uploaded video is not available.

That is not a soft restriction. The headline feature — take footage, continue it — does not work across most of Europe. Text-to-video and image-to-video still run. But if the plan was a tool that extends clips your users upload, and those users sit in Berlin or Manchester, the plan is dead and the announcement never mentions it.

Check it against your deployment region before you write a line of integration code. It is the most expensive item here to discover late.

The knobs that do not exist

Coming from a text model, you will reach for parameters that are not here. Google's docs mark the whole sampling surface unsupported: no temperature, no top_p, no system instructions, no stop sequences, no negative prompt. None of it.

Your only control surface is the prompt itself, and third-party documentation puts the ceiling at 40,000 characters — a lot of rope, though that figure is not one we found in Google's own reference. Everything you would normally do with a sampling parameter has to be written into English instead. If you want to know how to write that English well, we covered the prompt syntax in its own post, including the tag structure and the scene-cut problem, and there is no point repeating it here.

One caution. Third-party schema documentation for this model lists temperature and top_p as accepted fields, while Google's reference says they are unsupported. We will not tell you which is right — we are telling you the two disagree. Do not design around those parameters on the strength of a schema listing; test whether they change anything first.

Sound you cannot switch off, English you cannot leave

Audio is generated natively with every clip and there is no off switch. No parameter, no flag, no silent mode. You can steer it — a Sound design: sentence does real work — but you cannot decline it. For a blog embed that means muting in the player, which is what both clips above do.

You also cannot supply audio as a reference, edit generated audio afterwards, or preserve the audio of a video you upload for extension. All three are reasonable expectations. None is supported.

Language support is narrower than the marketing suggests: English is the only fully supported language. Japanese and others are documented as not evaluated — a careful way of saying results are unpredictable. If your prompts or dialogue are not in English, budget for surprises.

Know the reference caps before designing an input pipeline. Video references: up to three clips, three seconds each, and Google warns that supplying several at once can degrade results — treat three as a ceiling, not a target. Aspect ratios are 16:9 or 9:16, also Google's. Two numbers quoted alongside those are not: the 24 fps frame rate and the one-to-seven range for reference-to-video images both come from third-party schema docs. Every frame carries a SynthID watermark, and we have not found a provider that exposes a switch for it.

Before you spend anything, do this

Four checks, in order. The first kills more integrations than the rest combined.

1. Confirm your region. If you are in the EEA, Switzerland or the UK and your product extends uploaded video, stop here. This model cannot do that job for you.
2. Price the sequence, not the clip. Multiply by the extensions, then by the takes you will actually run. Forty seconds of 1080p is $5.40 here at best, realistically triple that once you count re-rolls.
3. Ask whether you need Omni at all. A second of Veo 3.1 Fast costs one cent here, against nine on Omni at 720p. One shot, no continuation, no frame interpolation, nobody who has to look the same two clips later — then that multiple is money set on fire.
4. Build your loop on the draft tier. Iterate at 360p, promote the winner. Sixty-eight percent is not a rounding error.

Omni earns the multiple exactly where the 1.1 features point: continuing a shot, interpolating between a first and last frame, animating a still, or carrying a subject across cuts. Outside that, it is an expensive way to buy one clip.

The other angles live next door: what actually shipped on launch day, what the leaderboards do and do not say — the arena rankings circulating this week belong to the previous model — and a platform-by-platform price breakdown. Schemas and live prices sit in the catalog.

Prices and product terms change. Check each provider's current pricing before making a purchasing decision.

Frequently asked questions

Can Gemini Omni 1.1 Flash really make 40-second videos?

Yes, but not in one call. A single generation returns 3 to 10 seconds. Forty seconds is the cumulative total you reach by extending an existing clip three more times — four jobs — and each extension is billed separately at the same per-second rate as a fresh generation. On E2X a 40-second sequence at 1080p costs $5.40, against $6.08 calling Google directly.

How much does a 40-second Gemini Omni sequence cost?

Multiply 40 seconds by the per-second rate for your resolution; extensions carry no discount. On E2X that is $1.19 at 360p, $3.60 at 720p, $5.40 at 1080p and $7.20 at 4K. Direct from Google, working from its published token-per-second rates, the same lengths cost $1.35, $4.06, $6.08 and $12.16. Add discarded takes and expect two to three times the headline.

Is Gemini Omni 1.1 Flash 4K output real 4K?

No. Google's documentation labels both the 1080p and 4K tiers "upscaled" — the model renders lower and scales the result up. The file has 4K dimensions; the picture does not carry 4K detail. Unless a delivery spec requires those dimensions, the top-tier rate buys a resize rather than more information.

Can I use Gemini Omni 1.1 Flash in the EU?

Partially. Text-to-video and image-to-video work, but editing or extending an uploaded video is unavailable in the EEA, Switzerland and the United Kingdom. Since extension is the headline capability, this blocks any European product built around continuing user-supplied footage. Verify it against your deployment region before integrating.

Does Gemini Omni 1.1 Flash support temperature or negative prompts?

Google's reference marks the sampling surface unsupported — no temperature, top_p, system instructions, stop sequences or negative prompt. Prompt text is your only control, with third-party schema docs putting the ceiling at 40,000 characters. Those same third-party docs list temperature and top_p as accepted fields, contradicting Google, so test them before depending on them.

Can I turn off the audio in Gemini Omni output?

No. The model builds a soundtrack into every clip by default, and no parameter switches it off. A sound-design instruction in the prompt steers what you get, but you cannot decline it, supply an audio reference, edit the result, or preserve the audio of a video you upload for extension. For silent playback, mute it in the player.

Can Gemini Omni extend a video backwards or edit the middle of a scene?

No. Extension appends to the end of a clip only — no backward extension to build a lead-in, no insertion into the middle of a sequence. Plan accordingly: the first generation sets the camera and lighting language for everything after it, and a mistake early invalidates every extension built on top of it.

Is Gemini Omni 1.1 Flash worth it over Veo 3.1 Fast?

Only if you need what Omni adds. Veo 3.1 Fast runs $0.01 per second on E2X against Omni's $0.09 at 720p. For a single clip with no extension, no first-and-last-frame control and no cross-shot continuity, Veo wins by a wide margin. Omni earns its premium on multi-shot sequences and reference-driven continuity work.