Gemini Omni 1.2 Flash
TL;DR Highlight
Gemini Omni 1.2 Flash makes 15-second videos with chat editing at 30% lower price than 1.1
Who Should Read
A backend developer weighing cost and quality before adding AI video generation to an ad or commerce service
Core Mechanics
- Clips go from 10 to 15 seconds
- Omni 1.1 topped out at 10 seconds; 1.2 Flash goes up to 15.
- You can set duration in one-second steps from 3 to 15 seconds.
- Price is 30% lower than 1.1
- It costs $0.07 per second for 720p video.
- Draft at 360p or 720p, then upscale only the best clips to 1080p or 4K.
- Edit by chatting, not regenerating
- Swap the product, background, lighting, weather, or aspect ratio turn by turn on the same clip.
- Multi-turn editing works across full 15-second clips.
- References hold from start to finish
- You can combine character, product, or style images, motion clips up to 3 seconds, and audio cues.
- Faces, garments, packaging text, and brand colors stay consistent across camera angles, per Google.
- Cut-free camera moves and better speech
- Google says continuous moves like FPV flythroughs, zoom-outs, and orbits have far fewer unwanted jump cuts.
- Lip-sync is clearer, and multiple speakers are assigned distinct voices.
- Now GA with enterprise options
- It is generally available on the Gemini API and Google Cloud, with Provisioned Throughput.
- Outputs carry an imperceptible SynthID watermark.
- Google says fast sports action, rapid script changes, and complex physics are still being improved.
Community Reactions
- Pricing confusion
- One commenter listed $1.50 input and $9.00 (text) / $12.00 (video) output per 1M tokens.
- Another said it actually cost more despite the cheaper claim.
- Quality still needs many iterations
- A user making AI videos since early 2024 said 15 seconds helps but complex prompts still fail.
- They said that without first/last frames on longer videos, unusual characters regress to bland designs.
- Early Flow use showed little beyond consistency
- A commenter said it is only worth using when character consistency is the problem and outputs didn't look much better.
- They said it's aimed at creatives rather than coders.
- Rollout is confusing
- Many complained it was unclear when Vids, the Gemini app, or paid plans get access.
- One user trying it on OpenRouter reported Gemini uptime of 68%.
How to Apply
- Calculate your 720p draft cost first
- Run the same prompt today in AI Studio at 5 and 15 seconds in 720p.
- Record cost per clip at $0.07/sec and how many retries you need.
- Build a draft-edit-upscale pipeline
- Generate drafts at 360p or 720p, then fix details with follow-up edits.
- Upscale only the final cut to 1080p or 4K to avoid regeneration costs.
- Make ad variants with multi-turn edits
- Generate one base clip, then send follow-ups that change only background, lighting, or aspect ratio.
- Include product photos and character sheets as references and check consistency.
- Build a consistency test set before adopting
- Generate several 15-second clips from three of your product photos and an unusual character design.
- Count how often faces, logos, and text hold up, then decide.
Try It
# Paste into Claude or ChatGPT: video prompt helper for Omni 1.2 Flash (the article's features turned into a prompt)
You are an AI video prompt director. From the product info below, write one 15-second video prompt and three follow-up edit prompts.
Product: {{product description}}
Audience and channel: {{audience, placement}}
Reference materials: {{product photos, character sheets, etc.}}
Requirements:
1. Length is 15 seconds, with one continuous camera move and no cuts.
2. State faces, garments, packaging text, and brand colors as elements to keep consistent.
3. Each of the three follow-ups changes only one of background, lighting, or aspect ratio and keeps everything else.
Next step: Run the result in AI Studio at 720p, tell me which elements drifted, and ask me to revise the prompt.Terminology
Related Papers
Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU
Shrinking DeepSeek-V4-Flash 284B to 1.71 bpw runs it on a 64GB CPU-only machine
Progress on provable inference for large language models
Proving 400B model outputs with zkSNARKs works, at 70-95x the GPU time of plain vLLM
Tell HN: Gemini trains on chat history, even for subscribers (personal accounts)
Gemini personal accounts may train on and human-review chats, even for paying subscribers
You don't need long context
Compacting context to 16K–48K kept accuracy while cutting estimated costs 2–8x
The AI cost of JavaScript frameworks
Building the same app with 12 JS frameworks via agents, 5 runs each, showed up to ~2x cost differences.
Claude Code regularly sends each background agent’s transcript to its model
Disabling Claude Code's 30-second agent status summaries removed 19.7% of API requests in one team's telemetry
Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)
A KV pool smaller than the context window leaked other customers' data into 155 long-prompt responses.
Microsoft Decision-1: 9B Qwen Model Launch Explained
Microsoft's 9B Qwen-based model returns only per-option probabilities for routing and classification.