blog.google
Introducing Gemini Omni — Google Blog
Gemini Omni Flash rolls out today to all Google AI Plus, Pro, and Ultra subscribers globally.
AI Product Update Daily: May 24, 2026
Google launches Gemini Omni Flash for conversational video editing. Anthropic's Mythos AI finds 10,000+ critical security bugs. DeepSeek makes a 75% price cut permanent. ChatGPT lands as a PowerPoint add-in. And OpenAI's safety head departs.
Google pushed a new video-generation model into the hands of every AI subscriber. Anthropic's security experiment crossed 10,000 critical bugs found. DeepSeek made its steep price cut permanent. And OpenAI officially landed inside PowerPoint. Here is everything that shipped or changed in the past 24 hours.
Google launches Gemini Omni Flash — video generation from any input
Google released Gemini Omni Flash, the first model in its new Omni family, on May 23-24, rolling it out globally to all Google AI Plus, Pro, and Ultra subscribers through the Gemini app, Google Flow, and YouTube Shorts. 1
Omni Flash accepts any combination of image, audio, video, and text as input and produces video output. The key capability: iterative, conversational video editing. You give it a clip, tell it to change something — the physics of an object, the environment, a character's action — and it carries forward the scene state across multiple instructions. Edits build on each other rather than restarting from scratch. Every output carries Google's SynthID watermark; users can verify AI origin via the Gemini app or Google Search.
Loading link preview…
Developer and enterprise API access is coming in the next few weeks. Omni Flash supports image and audio outputs later; the initial rollout is video-only. Benchmarks against Sora or other generators were not disclosed in the launch post.
YouTube Shorts and the YouTube Create App are also getting Omni Flash this week at no cost — the first time a full Gemini generative model has been given to consumers for free at this tier.
Anthropic's Mythos hits 10,000 critical security bugs found
Anthropic published an update on Project Glasswing confirming that partners using its Claude Mythos model have identified more than 10,000 high- or critical-severity vulnerabilities across software they describe as "the most systemically important in the world." 2
The breakdown: Mythos examined 1,000 open-source projects and found 6,202 high- or critical-severity flaws. Cloudflare, one of roughly 50 partner organizations with access, found 2,000 bugs on its own systems, 400 of which were classified high or critical — with a false-positive rate Anthropic says was lower than human testers. 3
One disclosed vulnerability: CVE-2026-5194, a flaw in wolfSSL (an SSL/TLS library common in IoT and smart home devices) that Mythos constructed an exploit for. The exploit would allow an attacker to forge certificates and host fake banking or email sites. Anthropic says a technical write-up is coming in the coming weeks.
Loading link preview…
Mozilla had earlier used Mythos to find 271 vulnerabilities in Firefox (April). Researchers using the same model earlier this month reported bypassing Apple macOS security. The tool debuted in April but has not been publicly released; Anthropic says it is "too powerful" for general availability.
Not everyone agrees with that framing. Gary McGraw, former VP at Synopsys, told The New York Times: "If you don't release a tool like this — or you hoard it — you are not solving the real problem." Google security researcher Michał Zalewski told The Wall Street Journal that some of the hype is "overblown."
DeepSeek makes 75% price cut on V4-Pro permanent
DeepSeek announced on May 23 that the 75% price reduction it applied to its V4-Pro API is now permanent, keeping costs at 0.025–6 yuan per million tokens ($0.0035–$0.83) — down from the original 0.1–24 yuan range. 4
When V4-Pro launched in April, DeepSeek attributed the higher price tier to "constraints in high-end compute capacity," and said prices would fall once Huawei Ascend 950 supernodes shipped at scale in the second half of 2026. The permanent cut suggests supply has improved ahead of schedule. DeepSeek did not confirm whether Huawei Ascend 950 availability was the driver.
Context: According to Latent.Space analysis from May 22, running DeepSeek's Intelligence Index on V4-Pro costs approximately 3× less than Gemini 3.1 Pro Preview, 12× less than GPT-5.5, and 19× less than Claude Opus 4.7. [cite:5|[AINews] All Model Labs are now Agent Labs — Latent.Space|[https://www.latent.space/p/ainews-all-model-labs-are-now-agent]]
Loading link preview…
ChatGPT add-in lands in Microsoft PowerPoint — global beta
OpenAI launched a ChatGPT add-in for Microsoft PowerPoint, now in global beta. The integration puts a sidebar inside PowerPoint that lets users create presentations from scratch using text prompts, incorporating documents or images as source material, or edit existing slides through natural language instructions. 5
Access is open to all ChatGPT plan tiers — Business, Enterprise, Education, Teacher, K-12, Free, Go, Pro, and Plus. The add-in is separate from Microsoft's own Copilot for Office 365.
OpenAI planning a cybersecurity model for Japanese firms
Asahi Shimbun reported on May 22 that OpenAI plans to provide a cybersecurity-focused AI model to select Japanese companies, with performance it characterizes as comparable to Claude Mythos. 6 No launch date was given. The report follows active AI security competition between OpenAI and Anthropic on the capabilities front.
Figma launches native AI agent on the design canvas
Figma introduced its own AI agent embedded directly in its collaborative design canvas, letting users generate new designs, edit existing work, and iterate using natural language prompts — without leaving the canvas. 7 The agent rolled out initially in Figma Design and is backed by the company's Anthropic Code to Canvas integration with Claude Code. Figma's revenue grew 46% year over year as of the launch announcement.
OpenAI safety head Aleksander Madry departs
Aleksander Madry, who built and led OpenAI's Preparedness team before being reassigned to AI reasoning work last summer, announced his departure from OpenAI after nearly three years. 8 In a statement, he said he plans to focus on studying AI's economic impact. Madry is an MIT professor on leave. No replacement for the Preparedness role has been announced.
In brief
- Anthropic's Stainless acquisition: The New Stack reported that Anthropic's ~$300M acquisition of Stainless (completed May 18) shut down the shared SDK generator that OpenAI, Google, and Cloudflare had all used — the clearest example yet of the Stainless deal's competitive impact on developer tooling. 9
- Anthropic + Microsoft compute talks: The Verge reported Anthropic is in early talks with Microsoft to rent Azure servers running Maia 200 AI chips to supplement the $1.25B/month compute deal it signed with xAI. 10
- YouTube Shorts AI remix: YouTube began rolling out an AI remix feature powered by Gemini Omni that lets users create new Shorts from other creators' videos. 11
- Ferrari + IBM AI for F1 fans: Ferrari is using IBM's AI platform to build personalized engagement tools for Formula 1 fans, including real-time race commentary and predictive analytics. 12
References
- 1Introducing Gemini Omni — Google Blog
- 2Companies Using Anthropic's Mythos AI Uncover 10K+ Serious Software Bugs — PCMag
- 3Anthropic Project Glasswing initial update — Anthropic Research
- 4China's DeepSeek to make permanent 75% price cut on flagship V4‑Pro AI model — Reuters
- 5OpenAI launches ChatGPT for PowerPoint in global beta — MSN
- 6OpenAI to provide latest cybersecurity AI model to Japanese firms — The Asahi Shimbun
- 7Figma builds its own AI assistant that can design alongside you — The Next Web
- 8OpenAI's Top Safety Expert Exits Company: 'Not An Easy Decision' — Benzinga
- 9Anthropic's $300M Stainless deal lands hardest on OpenAI and Google — The New Stack
- 10Anthropic is paying $15 billion a year for access to Elon Musk's data centers — The Verge
- 11You can now remix other people's YouTube Shorts with AI — The Verge
- 12Ferrari is using IBM's AI to create F1 superfans — TechCrunch
Add more perspectives or context around this Drop.