OpenAI Sora2, Sonnet 4.5, Meta Ads

AI’s footprint keeps expanding, this week, it shows up in powerful creative tools, smarter assistants, and deeper personalisation across platforms. OpenAI’s Sora 2 transforms video generation, Anthropic pushes coding AI further with Claude Sonnet 4.5, and Meta moves to tailor content and ads through Meta AI. Together, these updates reveal how AI is becoming more usable, human-like, and woven into daily life and work.

OpenAI’s Leap Into Realistic, Controllable Video

On September 30, 2025, OpenAI released Sora 2, its flagship video-and-audio generation model, which the company says is more physically accurate, realistic, and controllable than prior versions. OpenAI sees it as a step toward making AI-generated video more usable for creators.

The core of Sora 2’s advancement lies in its improved grasp of the physical world. Unlike earlier systems, Sora 2 can better respect cause and effect, for example, instead of a missed basketball shot unrealistically teleporting into the hoop, Sora 2 now shows it bouncing naturally off the rim or floor. OpenAI notes that Sora 2’s remaining errors “frequently appear to be mistakes of the internal agent that Sora 2 is implicitly modelling” - evidence, the company says, that it’s learning to model failures as well as successes. OpenAI frames this as an important step toward building a useful world simulator, noting that Sora 2 is better at following physical laws and capturing cause-and-effect than previous versions.

Beyond visuals, Sora 2 produces synchronised dialogue, realistic ambient soundscapes, and sound effects that are aligned with the video. This brings audio and video closer together than in earlier models, though results may vary depending on complexity.

A headline-grabbing new feature, Cameos, introduces a consent-based framework for digital likeness. Users can record a short video to create a personal avatar that can be inserted into any Sora-generated scene. OpenAI says Cameos are consent-based:

“Only you decide who can use your cameo, and you can revoke access or remove any video that includes it at any time.”

To house this ecosystem, OpenAI has launched a dedicated Sora app, described as a platform for ‘co-creation, not consumption.’ OpenAI says it is not optimising the Sora app for time spent in feed, instead focusing on creative collaboration, remixing, and sharing, supported by safety measures such as teen protections and parental controls. The Sora iOS app is rolling out on an invite-only basis in the U.S. and Canada, with OpenAI intending to expand access over time.

Claude Sonnet 4.5: A Leading AI Model for Coding, Complex Agents, and Long-Duration Tasks

On September 29, 2025, Anthropic released Claude Sonnet 4.5, positioning it as one of the leading AI models for coding, building complex agents, and advanced computer interactions. This model builds on previous versions by supporting longer context windows and greater task complexity, with notable improvements in reasoning, math, and domain expertise, particularly in finance, law, medicine, and STEM.

Claude Sonnet 4.5 brings practical upgrades to Anthropic’s product lineup. Anthropic says Claude Code now includes checkpoints (save/rewind), a refreshed terminal interface, and a native Visual Studio Code extension. The Claude API adds new memory management and context editing capabilities, enabling agents to operate effectively for longer sessions. Users can also run code and create spreadsheets, slides, and documents directly in conversation. Anthropic made the Claude for Chrome extension available to Max subscribers who joined the waitlist, as part of a limited preview.

In official benchmarks, Anthropic says Claude Sonnet 4.5 scores 77.2% on SWE-bench Verified and 61.4% on OSWorld, and that it can maintain focus for more than 30 hours on complex, multi-step tasks. Feedback from early customers highlights meaningful improvements in handling long-duration workflows and agent autonomy.

Significant attention has been given to safety and alignment. Anthropic describes Claude Sonnet 4.5 as its most aligned frontier model yet, showing measurable reductions in problematic behaviours like deception, sycophancy, and power-seeking. The model incorporates advanced safeguards, including classifiers to detect potentially dangerous inputs, such as those related to chemical, biological, radiological, and nuclear concerns, with efforts to reduce false positives. Robust defences against prompt injection attacks enhance security in agentic applications.

In addition, Anthropic introduced the Claude Agent SDK. Derived from the infrastructure behind Claude Code, this toolkit empowers developers to build custom AI agents capable of complex, autonomous tasks beyond coding. Alongside, the temporary research preview “Imagine with Claude” offers Max subscribers a chance to see the model dynamically generate software in real time.

Claude Sonnet 4.5 is now broadly available via Claude.ai, the Claude API, and integration with platforms like Amazon Bedrock and Google Cloud Vertex AI. Pricing remains the same as Sonnet 4: input $3 per million tokens and output $15 per million tokens, per Anthropic’s official pricing docs.

For developers and enterprises looking for AI that can navigate extended, complex tasks with stronger safety and reliability, Claude Sonnet 4.5 offers a significant step forward in capabilities and trusted deployment.

Meta to Personalise Content and Ads Based on Interactions with Meta AI

On October 1, 2025, Meta announced it would start using people’s interactions with Meta AI to personalise content and ads across its apps beginning December 16, 2025. Meta says it will start notifying people on October 7 via in-product notices and email.

Meta says Meta AI is used by over 1 billion people monthly and plans to use interactions as new signals to improve the relevance of posts, Reels, and ads in users’ feeds. For example, if a user talks to Meta AI about hiking, the system may suggest hiking-related groups, posts from friends about trails, or ads for hiking gear. This personalisation builds on existing mechanisms that tailor experiences based on user activity and interests.

Meta says users can control recommendations via Ads Preferences and feed controls; it also says the mic is only used with explicit permission, and a microphone indicator appears during voice interactions. Voice or text interactions with Meta AI will contribute to recommendations only if users actively engage with those AI features.

When people have conversations with Meta AI about topics such as their religious views, sexual orientation, political views, health, racial or ethnic origin, philosophical beliefs, or trade union membership, Meta says it will not use those topics to show them ads. Meta

also clarified that AI interactions will personalise content and ads only across accounts linked in the same Accounts Centre. For instance, without linking WhatsApp in an Accounts Centre, Meta AI chats on WhatsApp won’t inform personalisation on Facebook or Instagram.

Final Thoughts

These announcements highlight a distinct trend: AI is not only becoming smarter but also safer, more integrated, and designed to enhance genuine human creativity and productivity.

At QuensultingAI, we help companies navigate and harness these innovations, turning AI from buzzword to business advantage. Future-proof your operations with AI-powered voice agents, chatbots, and automation. Reach out to explore how QuensultingAI can support your AI journey.