Tech • AI • Robotics • Game
AI Went Crazy Again: Google Omni 1.2, Claude's 2X Acceleration, GPT-6.2
9.4/10TL;DR
Google, Anthropic, OpenAI, and World Labs are accelerating competition across AI video, automated research, coding tools, and real-time 3D world generation.
KEY POINTS
Google’s video model takes the lead
Google DeepMind released Gemini Omni 1.2 Flash on November 6, pushing clip length from 10 to 15 seconds and cutting cost by 30%. At 720p, pricing is about $0.07 per second, or roughly $1.05 for a full 15-second clip, undercutting several rivals. On human-preference arena rankings, it placed first in text-to-video and image-to-video, and third in video editing.
Longer clips and iterative editing sharpen the commercial pitch
The model is built to generate and edit video from mixed inputs, including character photos, product shots, style references, short action clips, and audio cues. It is designed to preserve identity, clothing details, lighting, and framing across camera changes, making it especially relevant for ads and virtual try-on formats. Editing happens through repeated prompts on the same clip, allowing background, product, weather, or aspect-ratio changes without starting over.
Camera movement and dialogue quality improve
The latest version expands motion range with FPV-style shots, large zoom-outs, and orbiting camera paths. Multi-speaker dialogue handling and lip-sync have also improved, reducing one of the biggest weak points in AI-generated video. The broader market is increasingly shifting from pure visual quality to a contest over length, price, and editability.
Anthropic’s internal progress appears to be accelerating
External researchers from METR spent about 11 weeks reviewing internal data at Anthropic, with findings released on November 5. Using Anthropic’s own capability metric, model progress since October 2025 has risen about 1.9 times faster than in the prior period, from roughly 15 points per year to 29. That implies about two years of previous progress being compressed into one year.
AI is now doing much of the work of improving AI
METR estimated that roughly 80% of Anthropic’s acceleration now comes from AI assisting research rather than from simple growth in training compute. Researchers described a workflow where coding is less of a bottleneck, and many more experiments can be launched in parallel. The result is a faster cycle for testing ideas that previously would have been too costly or slow to pursue.
Unreleased Anthropic systems are reportedly boosting staff output
The review described three unreleased models trained between July and October, with the strongest, Model 4, scoring above 90% on many common benchmarks. In an internal October survey, staff reportedly estimated that this model made them about 12 times more productive on average. Earlier internal estimates were around 1 times at the start of 2025, 2.5 times with Opus 4.5, and 5 times with Mythos 5.
Debate remains over whether recursive self-improvement has arrived
Anthropic researchers reportedly do not consider full recursive self-improvement to be here yet because AI has not independently driven the most critical breakthroughs in model training. Still, internal work was described as showing meaningful early signs, with AI handling substantial portions of research tasks. Researchers and reviewers also disagreed over how to describe whether some internal warning thresholds had effectively been reached.
OpenAI is preparing a new model family
Reports indicate OpenAI may soon unveil a broader release centered on GPT-6.2, alongside new systems referred to as Astra, Soul, and a cybersecurity-focused model. Early claims describe high cost but major gains in reasoning and generation quality, including stronger performance even at lower compute settings. None of those details are fully confirmed, but the launch expectations have intensified rivalry at the top end of the market.
Anthropic loses key Claude Code personnel
Claude Code lost four prominent team members in under two weeks, including product lead Cat Wu, who moved to OpenAI. Other departures included Tariq Shihipar, Anthony Morris, and engineering manager Sunnil Shaw. Wu joined OpenAI’s core products group under Thibo Sottos, whose team leads Codex, a direct competitor in AI coding tools.
World Labs brings live AI 3D worlds to a handheld-style interface
World Labs, the startup co-founded by Fei-Fei Li, introduced FrameBoy, a browser-streamed research preview that presents a Game Boy-like interface for navigating AI-generated 3D scenes in real time. Built on its Atlas world model, the system can create explorable spaces from a single image, a few photos, or even text. After optimization, it now streams at up to 15 frames per second, a major jump from earlier workflows that required waits of tens of seconds or minutes for each camera move.
AMD’s acquisition raises the stakes in physical AI
AMD announced on September 28 that it would buy World Labs in an $8.2 billion all-stock deal, expected to close by year’s end. The acquisition would make Fei-Fei Li an executive vice president and chief scientist at AMD, reporting to Lisa Su. The move signals that 3D world models, simulation, and physical AI are becoming a strategic front in the hardware race against Nvidia.
CONCLUSION
The latest wave of releases and research suggests AI competition is no longer centered on a single benchmark. The new battleground is broader: cheaper video generation, faster AI-assisted R&D, stronger coding products, and real-time world models that tie software advances directly to chip strategy.
Ask a question