Gemini Omni 1.2 Flash goes GA: 15-second clips, 30% cheaper
Google’s video model is live on the API with multi-turn re-edits — plus GPT-6 Cyber ships gated, OpenAI buys Pierre, and a trick to cut agent cost 73%.

Copy markdown
Gemini Omni 1.2 Flash is generally available
Google’s video model now does native 15-second clips (up from 10, selectable 3–15s in one-second steps) and multi-turn conversational re-edits of existing shots. 720p lands at $0.07/sec — a 30% cut from Omni 1.1 — and you can draft at 360/720p then upscale to 1080p or 4K. Live on the Gemini API and Vertex AI today, so longer, cheaper, editable clips are shippable now.
GPT-6 Cyber ships — approved users only
OpenAI’s new security model is rated “Critical” on its Preparedness Framework and solves 48.5% of ExploitGym tasks, up from 32.1% for GPT-5.6 Cyber. Access is gated behind “Daybreak Red” with identity checks and IP allowlisting on by default; vetted defensive researchers get “Daybreak Blue.” Most builders won’t get a key, but this is the clearest template yet for how frontier capability gets rationed.
OpenAI is buying Pierre to deepen Codex
OpenAI is acquiring dev-tools startup Pierre Computer, whose code-navigation tech already powers parts of Codex, in a move Axios frames as a shot at GitHub. If you live in Codex, expect tighter repo-aware navigation and review features to follow.
ChatGPT Images 2.6 makes animated GIFs
OpenAI’s new image model reportedly turns text prompts or photos into animated GIFs, with frame-by-frame editing and speed control, and generates stills up to 50% faster — rolling out across ChatGPT and Codex on web, desktop and mobile. A quick weekend unlock for looping product shots and reaction clips.
EnvOpt cuts agent task cost 72.7%
A new arXiv method reorganizes an agent’s environment — not the model or the harness — and cut Claude Haiku 4.5’s cost per task by 72.7% on AutomationBench while raising pass rate, across 200 tasks. If you pay per agent run, that’s free leverage: restructure the workspace, not just the prompt.
METR clocks Anthropic’s self-speedup near 2x
METR estimates Anthropic is improving model capabilities on well-specified tasks at roughly twice its Mar 2024–Oct 2025 pace (a 1.3–3x range), though Anthropic redacted most of the supporting evidence. The recursive-acceleration story finally has a third-party number on it.