AI - Gemini Omni 1.2 Flash makes video longer and cheaper, and OpenAI releases GPT-6 Cyber
Two very different releases landed within hours of each other yesterday 🤣 Google made AI video longer and cheaper with Gemini Omni 1.2 Flash, while OpenAI released GPT-6 Cyber, a model that can build working exploits for real-world software.
They have almost nothing in common, but both show models taking over jobs that used to need a lot of human time, so I have put them in one post. I would happily try the first one. I am rather glad the second one sits behind an application process 😊
Gemini Omni 1.2 Flash
Gemini Omni 1.2 Flash is the updated video generation and editing model from Google DeepMind. The headline changes are longer clips of up to 15 seconds, a 30% price cut and the ability to keep editing the same clip in conversation instead of regenerating it from scratch.
Official announcement
Gemini Omni 1.2 Flash is generally available on Google Cloud
Meet Gemini Omni 1.2 Flash, our updated video generation model from @GoogleDeepMind 🎬
Get stronger consistency, more creative control, and faster, more cost-effective video creation at scale:
The rest of the post lists five headline changes, which are covered below.
What is new?
According to Google, Gemini Omni 1.2 Flash brings:
- A lower price: $0.07 per second of 720p video, which is 30% less than Gemini Omni 1.1 Flash.
- Longer clips: native generation and editing of up to 15 seconds, up from 10 seconds, with any length from 3 to 15 seconds in one-second increments.
- Draft then upscale: draft at 360p or 720p, then upscale the best results to 1080p or 4K.
- Editing in conversation: swap a product or background, change the lighting or weather, or switch the aspect ratio, turn by turn, across full 15-second clips.
- Better reference following: character, product or style images, short motion clips of up to three seconds and audio cues can guide the result. Google’s developer documentation says up to six reference images can be used.
- More natural speech: more fluid dialogue, clearer lip-sync and distinct multi-speaker conditioning.
- More controllable shots: continuous camera moves with significantly fewer unwanted jump cuts.
Google also says that outputs carry an imperceptible SynthID watermark, and that Provisioned Throughput is available on Google Cloud.
The price in the announcement is for 720p, so here is how the effective price per second looks across resolutions on Google’s pricing page:
| Output | Effective price per second | Output tokens per second |
|---|---|---|
| 360p draft | About $0.023 | 1,931 |
| 720p | About $0.07 (about $0.10 on Omni 1.1 Flash) | 5,792 |
| 1080p with upsampling | About $0.104 | 8,688 |
| 4K with upsampling | About $0.209 | 17,376 |
These are Gemini API Standard paid-tier prices as listed on 6 November. Google’s pricing page converts the 720p rate to roughly $0.70 for a 10-second clip and $1.04 for a 15-second clip.
Availability depends on how you want to use it:
- Gemini API: generally available on the paid tier only. There is no free tier for Omni 1.2 Flash.
- Gemini Omni 1.1 Flash: remains available during a four-week transition. From 3 December 2026, requests to
gemini-omni-1.1-flashare routed automatically togemini-omni-1.2-flashand billed at the lower 1.2 rates. - Audio references: Google says these are gated behind allowlisting on Google Cloud.
- Google apps: 15-second videos are rolling out to Google AI Ultra subscribers in the Gemini app, and Google says Google Vids can generate 15-second clips at no additional cost with a Google account or Google Workspace subscription.
To try it, use Google AI Studio or the Google Cloud console.
My take
Video generation is priced per second, and every second you throw away still costs money, so the workflow matters as much as the headline price. Drafting cheaply, fixing a detail with a follow-up edit and only upscaling the keepers is a sensible way to keep the bill down. It is also exactly the pitch Google is making to businesses producing product videos and ads at volume.
A lower price per second does not automatically mean a lower total bill, though. A 15-second clip at 720p is roughly $1.04, and a handful of rejected attempts soon add up. Conversational editing should help, as long as it works reliably across the full 15 seconds. Google is also upfront that development is ongoing on fast sports action, rapid script changes and complex physics, so it is not claiming that everything is solved.
I have not tried Omni 1.2 Flash myself, so I cannot say how well the consistency and lip-sync claims hold up in practice. A leaderboard position is a useful sign, but it is a general guide and not a promise about your own prompts 😉
Google also mentioned Gemini Omni in its Genie 4 announcement, as something that helps anchor simulations in real-world appearance, which I covered in my Genie 4 post.
Relevant X posts
Before the release
TestingCatalog spotted references to the model on Google Cloud ahead of the announcement and expected a minor upgrade over Omni 1.1 Flash. A 50% increase in clip length and a 30% price cut feels like more than minor to me 😉
Release details
TestingCatalog summarises the availability, clip length, resolution and price change:
Google AI Studio shared the developer announcement as an X article. Its opening paragraph covers the rollout, the 15-second clips and the price:
Today, we’re rolling out Gemini Omni 1.2 Flash, and making it generally available across the Gemini API and Gemini for business. Gemini Omni 1.2 Flash introduces native 15-second video creation, multi-turn conversational editing capabilities, and enhanced reference following. We have dropped the price 30% from Gemini Omni 1.1 making it available at $0.07 per second for 720p video.
Google Cloud Tech announced general availability on Google Cloud, aimed at product videos, try-ons and ads:
Wes Roth summarises the 3 to 15 second clip lengths, the follow-up edits and the lip-sync and camera improvements:
Early reaction and the leaderboard
Vamsi Batchu works at Google DeepMind, so this is an enthusiastic insider reaction. The chart attached to his post shows Omni 1.2 Flash at #1 on Arena’s Text-to-Video leaderboard with a score of 1,541, ahead of Omni 1.1 Flash (1,514) and Omni Flash (1,509). I could not load the live leaderboard to confirm it, so I am treating the chart as his claim:
Gemini Omni 1.2 Flash to the top 🔥 can’t wait to play with this !!!
OpenAI GPT-6 Cyber
GPT-6 Cyber is OpenAI’s new cybersecurity model, designed for finding and validating vulnerabilities and for developing exploits within the boundaries of authorised testing. It is a step up from GPT-5.6 Cyber and the model behind the Linux kernel research OpenAI shared the day before.
Official announcement
Introducing GPT-6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work.
GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
GPT-6 Cyber is not generally available. It is initially available through the API and Codex to a select group of approved Daybreak Red customers, with access expanding progressively to all Daybreak Red customers. OpenAI is also making GPT-5.6 Cyber available to all Daybreak Red customers and expanding Daybreak Blue alpha access for GPT-6 Astra. I could not find pricing in the announcement.
If you are a security team or researcher, new users and organisations can apply for Daybreak.
What is new?
OpenAI reports the following results for GPT-6 Cyber:
| Evaluation | GPT-6 Cyber | Comparison |
|---|---|---|
| ExploitGym: exploiting a specified vulnerability in a remote target | 48.5% | GPT-5.6 Cyber: 32.1% |
| ExploitGym Honeypot: avoiding out-of-scope services, without the deployment prompt | 99.93% | GPT-5.6 Cyber: 94.49% |
| Offensive web penetration testing: vulnerability recall | 40.3% | GPT-6 Astra: 32.0% |
| RemoteExploitBench: tasks completed | 95% | GPT-6 Astra with Daybreak Blue: refused 100% of tasks |
OpenAI says the model also stayed within scope in 100% of the Honeypot tests when its deployment prompt was used. The other headline claims are:
- 81 new high-severity Firefox vulnerabilities, plus 30 known issues that were more severe than previously reported. OpenAI says it disclosed these to Mozilla with proposed patches, and that Mozilla began landing fixes shortly after.
- 948 working Linux kernel privilege escalation exploits. OpenAI’s research pipeline reproduced 1,724 vulnerabilities, and the 948 figure is more than five times the 166 exploited by GPT-5.6 Sol using the same research harness.
- A Critical rating in cybersecurity under OpenAI’s Preparedness Framework, up from High for GPT-5.6 Cyber.
- Added safeguards, including cyber-specific misalignment monitoring, Private Safety Processing and IP allowlisting for Daybreak Red API access by default.
- Critical infrastructure access: a new Daybreak Red for Critical Infrastructure group, alongside OpenAI’s $1 billion Daybreak for Frontline Defenders grant programme.
My take
I find it hard to decide whether to be impressed or slightly nervous 😊 Finding and fixing vulnerabilities in widely used open-source software such as Linux and Firefox is exactly what I want defenders to have. But OpenAI rates the model Critical under its own framework, and in one of its own comparisons the mainline GPT-6 Astra refused every RemoteExploitBench task while GPT-6 Cyber completed 95% of them. That is the dual-use problem in a single line.
That is why the gated access makes sense to me, and it is becoming a pattern. Google started Gemini 4 Carbon with a subset of cyber defenders, as I covered in my Carbon post, and Anthropic keeps Mythos 5.1 to its trusted access programmes, as I covered in my Fable 5.1 and Mythos 5.1 post. Anthropic is also scanning open-source projects for maintainers, which I wrote about in my OSS Scanner post.
A few caveats are worth keeping in mind:
- Every number is vendor-reported. Google reported 54.8% on ExploitGym for Carbon, against OpenAI’s 48.5% for GPT-6 Cyber, but each company used its own setup, so I would not treat that as a head-to-head result.
- The Firefox figure is OpenAI’s. Mozilla’s advisory for Firefox 159 lists many high-impact fixes, but I could not tie the figure of 81 to individual entries from the advisory alone.
- Finding bugs is not the whole job. OpenAI’s own Linux write-up says the challenge was not discovery but working out which findings could be reproduced and which deserved maintainers’ attention first. Maintainers still need good triage and patches, not just a larger pile of reports.
Relevant X posts
Chubby gives a quick summary of the release. The comment about GPT-6.2 is his own and is not part of OpenAI’s announcement:
Alex summarises the main figures, including the Linux exploits, the Firefox findings and the Critical rating, which match OpenAI’s announcement:
Codex Updates says more announcements are on the way. I could not find anything in OpenAI’s announcement to support that, so I would treat it as a teaser and not as a confirmed plan:
Useful links
#Blog #AI #Google #Google Deepmind #Gemini Omni #Gemini Omni 1.2 Flash #Video Generation #OpenAI #GPT-6 Cyber #Daybreak #Cybersecurity #Open-Source