Field SOP
Field SOP

A 360p Draft, Then 4K: Why One Word Still Gets Rebilled?

Take gemini-omni-1.2-flash from an API key to a finished clip in five documented steps. Confirm the channel and billing first: the Gemini API paid tier has no free tier, input is $1.50, text output $9.00 and video output $12.00 per 1M tokens versus $17.50 for 1.1, and the Gemini API docs say provisioned throughput is not supported. Start with a cheap 360p draft through the Interactions API response_format, with duration from 3s to 15s, a 10s default and aspect_ratio 9:16 or 16:9. Add up to six reference images while remembering that first and last frame interpolation only supports 3 to 10 seconds. Edit conversationally with previous_interaction_id instead of regenerating, and note that store=false removes later editability. Fetch clips over 4MB with delivery set to uri and poll the Files API until ACTIVE; 1080p and 4K are upscales. Cost is derived from official per-second prices, with the pricing page quoting about $1.04 for a 15-second 720p clip. Three traps get their own sections: the free tier and the December 3 rerouting, the list of unsupported features, and size and state limits.

Published November 7, 20269 min read
<!-- gemini-omni-12-flash-api-setup-sop | sop | A 360p Draft, Then 4K: Why One Word Still Gets Rebilled? -->

You have a Gemini API key and want a clip from Gemini Omni 1.2 Flash. The hard part is rarely the first request; it is the bill and the errors: why a one-line edit still costs money, why a first and last frame errors out, why a large file never arrives. The Google API changelog marked gemini-omni-1.2-flash generally available (GA) on November 5, 2026, a GA milestone, not a first release, since a 1.1 model already existed. The Google Cloud blog is dated November 6, 2026, and Chinese-language coverage dated November 6, 2026 is a reprint, not a release date. This follows the official changelog, pricing page, model page, and video and Omni docs, splitting the path from key to finished clip into five steps plus three traps.

Scope first. This piece is wiring up and settling the bill: five steps from key to finished clip, with parameters, limits, and cost estimates. The hotspot column covers the GA event and official capability and pricing line; the open-source column covers craft-anything; the comparison column covers the parameter table for five GitHub tools. It differs in angle from the November 4, 2026 llm-api-pricing-comparison-review, which unpacks pricing illusions across vendors. This computes only the official token conversion for this one Google video model, with no cross-vendor comparison and no provider recommendation.

One boundary holds throughout: only parameters, endpoints, and price conversions in the official documentation appear here, with no verified run, no configuration checklist, and no timing guarantee. Per-second multiplications are labeled estimated from official per-second rates; official page figures are quoted as written. Nothing here is legal or procurement advice.

Step 1: Fix the channel and billing assumptions before writing code

Step one is not code; it is pinning down the channel and billing premise. The official pricing page marks the Free Tier as Not available, so there is no free tier: neither input nor output carries a free allowance, and a request needs the Gemini API paid tier with the model string gemini-omni-1.2-flash.

Copy prices exactly from the official pricing page: input $1.50 per 1M tokens covering text, image, and video; text output $9.00; video output $12.00. For contrast, the 1.1 video output is $17.50, and the official changelog describes 1.2 video output as 30% lower. Those belong to two models, so the accounting uses only $12.00, never $17.50. The official pricing page also states the billing basis: total output token consumption, at 5,792 tokens per second of 720p video.

Keep the channels separate. The Google Cloud blog dated November 6, 2026 calls the model production-ready and offers Provisioned Throughput, while the Gemini API documentation states Provisioned throughput is not supported on the API side. This article follows the Gemini API paid tier, so Provisioned Throughput and audio-reference allowlisting belong to the Cloud side and do not infer what the API can do. Some official pages still carry 1.1 example model strings; copying them mixes up model and bill, so every code block here uses gemini-omni-1.2-flash.

Step 2: Start with a cheap draft by keeping duration and resolution low

Do not ask for 1080p or 4K first. The official video documentation enters through client.interactions.create(...) from the Interactions API, with response_format marking a video and setting length and resolution. Start at the lowest tier to confirm prompt, framing, and lighting, then decide whether to raise resolution; that holds the bill down best. In the shape shown in the official documentation:

python
interaction = client.interactions.create(
    model="gemini-omni-1.2-flash",
    input="A slow push-in on a ceramic cup on a sunlit wooden table, warm side light.",
    response_format={"type": "video", "duration": "5s", "resolution": "360p"},
)

Memorize a few response_format fields. duration accepts "3s" through "15s", default "10s"; over 15 seconds returns an invalid argument error, so a long clip is not one stretched request. resolution accepts 360p, 720p, 1080p, and 4K, default 720p, with 1080p and 4K upscaled in Step 5. For vertical set aspect_ratio to "9:16"; omit it for default "16:9" landscape. 5s at 360p is the cheapest tier here.

The official per-second conversions show why. The official pricing page gives 360p at 1,931 tokens per second, about $0.023 per second, and 720p at 5,792 tokens per second, about $0.07 per second, the default tier. Estimated from official per-second rates, 5 seconds at 720p is roughly $0.35 and at 360p roughly $0.12; the official pricing page gives about $0.70 for a 10-second 720p clip directly.

Step 3: Add reference material, and respect the hard limits

Once the draft points the right way, add references to lock character and style. The official documentation allows up to 6 reference images per request, up to 3 images per character and 2 characters, so the six are grouped by character. Reference videos follow a separate rule: a 3 to 10 second output allows up to 3, an 11 to 15 second output allows 1, each at most 3 seconds, and reasoning across multiple videos is not supported.

First and last frame interpolation supports only 3 to 10 seconds, and supplying both frames while duration exceeds 10 seconds errors out. A longer clip comes from conversational extension in Step 4, not from pulling duration to 15 seconds here.

The official documentation also gives prompt guidance: write scene, camera, and light clearly, and for a single shot say "In a single unbroken scene" or "No scene cuts", or the model may cut between shots. Negative prompts are not supported, so put "No dialogue" or "No embellishments" into the normal prompt. This keeps the response_format from Step 2, with more material in input, and the model string stays gemini-omni-1.2-flash.

Step 4: Edit in conversation instead of rebuilding every time

After the first pass, changing a product, the lighting, or the camera angle does not need a fresh generation. The official Omni documentation describes stateful editing: pass previous_interaction_id from the prior interaction and the model reuses the existing video state without re-uploading material. Keep edit prompts short, ending with "Keep everything else the same" as the official guidance suggests:

python
refine = client.interactions.create(
    model="gemini-omni-1.2-flash",
    input="Swap the ceramic cup for a matte black one. Keep everything else the same",
    previous_interaction_id=interaction.id,
)

Be clear what this saves: previous_interaction_id avoids re-uploading references and rebuilding state, but every create is a new generation request billed on that turn's output tokens. That is why a one-word edit is still charged; it is not a free tweak.

Video extension uses the same mechanism: 3 to 10 seconds per turn, up to 30 seconds total, appended only at the tail. The official best practices list background=false, store=false, stream=false for faster synchronous generation, but store=false makes the clip non-editable in later turns via previous_interaction_id. If you want to keep editing, do not turn store off for speed; decide before sending, because it cannot be recovered.

Step 5: Fetch, upscale, and settle the bill, quoting $1.04 rather than recomputing it

Only after draft and edits settle do resolution and fetching come in. The official model page marks 1080p and 4K as upscaled tiers; the official pricing page gives 1080p with upsampling at 8,688 tokens per second, about $0.104 per second, and 4K at 17,376 tokens per second, about $0.209 per second, both above the roughly $0.07 per second default for 720p. The Google Cloud blog gives the same pattern: draft at 360p or 720p, then upscale chosen creations to 1080p or 4K, rather than paying upscaling prices while drafting.

A 4MB threshold governs delivery. For a clip larger than 4MB, the official video documentation requires delivery: "uri" inside response_format so the result returns as a URI, then the Files API is polled until the file state reaches ACTIVE before download. Clips under 4MB return directly; a clip over 4MB without delivery: "uri" hits the payload limit.

python
final = client.interactions.create(
    model="gemini-omni-1.2-flash",
    input=prompt,
    response_format={"type": "video", "duration": "15s", "resolution": "720p", "delivery": "uri"},
)

The accounting rule: official per-second price times seconds times resolution tier, with every multiplied figure labeled estimated from official per-second rates. But one figure is quoted, not multiplied: the official pricing page gives about $1.04 for a single 15-second 720p clip directly, official token-converted wording quoted as written, not replaced by a fresh multiplication. About $0.70 for a 10-second 720p clip is likewise taken directly from the official pricing page. Keep those two classes apart and the bill stays consistent.

Every clip carries a SynthID watermark, described in the official documentation as invisible and programmatically detectable, attached on output with no extra parameter.

Trap 1: No free tier, plus a reprice routing that only takes effect on December 3, 2026

The first trap is expectation mismatch. The official pricing page writes Free Tier as Not available, so there is no free tier; free allowances from other products are not API free credits, and every Gemini API paid-tier request is billed on output tokens. The second sits in a date: the official changelog and pricing page state that effective December 3, 2026, requests to gemini-omni-1.1-flash automatically route to gemini-omni-1.2-flash and are billed at 1.2 rates. This is future-effective and not yet in force as of November 7, 2026, so do not call it already applied. Keeping the old 1.1 string will not preserve a billing line outside the $17.50 video output, because after routing the bill follows the 1.2 figure of $12.00. Change the model string to gemini-omni-1.2-flash and check the bill yourself.

Trap 2: The unsupported list is easier to trip over than the supported one

The Gemini API version documented here does not support uploading audio references, distinct from outputting video with native audio: the model page says Output: Video with native Audio, but input-side audio references are unavailable on the API, while Cloud-side audio references are gated behind allowlisting, a separate channel's line. System instructions, temperature, top_p, stop sequences, and negative prompts are not supported, so negatives go into the normal prompt. Provisioned throughput is not supported on the Gemini API side and cannot be reversed from the Cloud blog's Provisioned Throughput. Editing and extending uploaded videos are unavailable in the European Economic Area, Switzerland, and the UK, a documented regional limit. Voice editing and YouTube as a media source are not supported. English is fully supported while other languages are not evaluated, so do not assume a Chinese prompt performs like an English one.

Trap 3: Size, state, and the payload limit account for most errors

Grouping size and state problems removes most avoidable errors. A duration over 15 seconds returns an invalid argument error, so 15 seconds is the single-clip ceiling. An input video for editing or extension needs at most 10 seconds, with multi-turn extension at 3 to 10 seconds per turn and up to 30 seconds total. Reference videos are capped at 3 clips, with no reasoning across multiple videos. On state, store=false drops later editability: once store is off for synchronous speed, the clip cannot be edited with previous_interaction_id. On delivery, a clip over 4MB without delivery: "uri" and polling the Files API until ACTIVE hits the payload limit. These are request sizes or states that miss the official constraints, so check them before tuning the prompt.

FAQ

Q1: Is there a free allowance to try first? A1: No API free tier exists for this model. The official pricing page marks the Gemini Omni 1.2 Flash Free Tier as Not available, so there is no free tier, and both input and output are billed per token on the Gemini API paid tier.

Q2: Can I pass an audio clip as a reference? A2: The Gemini API version documented here does not support uploading audio references, an explicit official limitation. Keep two things apart: the model outputs video with native audio, but input-side audio references are unavailable on the API, and Cloud-side audio references need allowlisting, another channel's line that does not imply the API can do it.

Q3: What is the longest clip I can make? A3: Two answers. A single generation accepts duration from "3s" to "15s", default "10s", and over 15 seconds returns an invalid argument error, so one clip tops out at 15 seconds. For longer, use conversational extension: 3 to 10 seconds per turn, up to 30 seconds total, appended only at the tail with no start or middle insertion.

Q4: Why is a one-word edit billed again? A4: Billing is based on each request's output token consumption. previous_interaction_id reuses prior video state and avoids re-uploading, but every client.interactions.create is a new generation turn producing new output tokens, which the official pricing page converts at 5,792 tokens per second for 720p. Editing saves the upload and state rebuild, not the output charge for that turn.

Q5: When does the old 1.1 model string change price? A5: Per the official changelog and pricing page, effective December 3, 2026, requests to gemini-omni-1.1-flash automatically route to gemini-omni-1.2-flash and are billed at 1.2 rates; future-effective and not yet in force as of November 7, 2026. 1.1 video output is $17.50 and 1.2 is $12.00. Do not expect the 1.1 string to preserve an old bill; switch to gemini-omni-1.2-flash and check the account yourself.

Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-07

FAQ

Is there a free allowance to try first?
No API free tier exists for this model. The official pricing page marks the Gemini Omni 1.2 Flash Free Tier as Not available, so there is no free tier, and both input and output are billed per token on the Gemini API paid tier.
Can I pass an audio clip as a reference?
The Gemini API version documented here does not support uploading audio references, an explicit official limitation. Keep two things apart: the model outputs video with native audio, but input-side audio references are unavailable on the API, and Cloud-side audio references need allowlisting, another channel's line that does not imply the API can do it.
What is the longest clip I can make?
Two answers. A single generation accepts duration from "3s" to "15s", default "10s", and over 15 seconds returns an invalid argument error, so one clip tops out at 15 seconds. For longer, use conversational extension: 3 to 10 seconds per turn, up to 30 seconds total, appended only at the tail with no start or middle insertion.
Why is a one-word edit billed again?
Billing is based on each request's output token consumption. previous_interaction_id reuses prior video state and avoids re-uploading, but every client.interactions.create is a new generation turn producing new output tokens, which the official pricing page converts at 5,792 tokens per second for 720p. Editing saves the upload and state rebuild, not the output charge for that turn.
When does the old 1.1 model string change price?
Per the official changelog and pricing page, effective December 3, 2026, requests to gemini-omni-1.1-flash automatically route to gemini-omni-1.2-flash and are billed at 1.2 rates; future-effective and not yet in force as of November 7, 2026. 1.1 video output is $17.50 and 1.2 is $12.00. Do not expect the 1.1 string to preserve an old bill; switch to gemini-omni-1.2-flash and check the account yourself.

Related

Field SOP

ComfyUI Video Generation SOP: Run LTX-2.5 on 12GB VRAM

A ComfyUI local video generation SOP, using LTX-2.5 to get a 12GB card through the full pipeline. Model files and directories (int8_convrot builds, not BF16): the transformer goes to models/diffusion_models/, gemma4-12b-with-proj to models/text_encoders/, the video and audio VAEs (both bf16) to models/vae/, and the 2x latent upscaler to models/latent_upscale_models/. Pitfalls first: the audio VAE is a separate file - skip it and you render a silent video; the prompt enhancer is an extra ~5GB model adding 1-2 minutes per generation and is off by default; HF gated=auto means you must log in and accept terms to download (even to read the model card). Hardware floor: 12GB VRAM with distilled/int8, CUDA 12.7+, PyTorch ~2.7, Python 3.12+ (innfactory's compilation); on 8GB cards, switch to Wan 2.2 TI2V-5B with offload instead. Three workflow types: two-stage T2V/I2V (generate, then 2x latent upsample refinement), single-stage (fast, for previews), and A2V (audio-driven video). VRAM strategy: quantized/distilled first, ComfyUI native offload, short low-res generation then upscale. Repos: github.com/Lightricks/ComfyUI-LTXVideo, huggingface.co/Lightricks/LTX-2.5.

Oct 10, 20269 min read
Field SOP

GitHub Copilot: Six Models Deprecated December 14, What to Fix

Six GitHub Copilot models are deprecated on December 14, 2026, and the official November 13, 2026 changelog gives six migration sentences but no configuration file name, model ID string or migration script. This SOP turns those sentences into a checkable list: replace hardcoded model names in workflows and integrations, enable model policies in Copilot settings, and understand what global default, the model selector, Copilot Chat, inline edits and code completions each do. All six model names and suggested alternatives are quoted as published, Kimi K3 stays To be announced, and the Claude Haiku 4.5 annual-plan exception for individual subscribers is spelled out.

Nov 16, 20269 min read