2026-09-23

Keep the performance, replace the world

A hybrid filmmaking tutorial keeps live performance, blocking, composition, and camera movement in the practical shoot. It uses an AI-generated first frame to establish the new setting, then feeds either the original footage or an optional depth map, which represents how near or far parts of the scene are, into the AI video-generation tool Seedance 2.5 to guide the motion.

In one cliffside test, the original footage preserved the performance and camera movement, while the depth map handled the overall motion a little better. The tutorial also found that gestures and props could drift between generations, so this is controlled transformation rather than exact motion transfer.

New models, lower prices

Developer Simon Willison's first look at this week's model releases covers Claude Opus 5.5 and OpenAI's GPT-6 Sol and Luna. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Opus 5.5 cuts input and output prices by 20 percent from earlier Opus versions, while cache-read pricing falls 60 percent.

There is an important edge case. In Willison's test asking the model to draw a pelican riding a bicycle, Opus 5.5 at its max thinking level twice reached its 128,000-token output limit without returning the drawing. Each failed run cost him $2.56 and took nearly 20 minutes.

Faster disaster maps

Microsoft says its open-source HASTE system helps responders analyze before-and-after satellite imagery. An analyst labels examples of damaged and undamaged buildings, then the model applies those patterns across a larger area.

After Venezuela's June earthquake, Catholic Relief Services received a damage map by the next morning. Its digital-programming lead said a more manual assessment would have taken the organization two weeks. Microsoft reports that HASTE has supported responses to 23 major disasters in 16 countries since 2023.

Also

• A clearer trail through Search ads Google Ads plans a unified report showing which search term triggered an ad, which creative appeared, and where the visitor landed; availability details are due later this year.
• Muse follow-up: Meta patches a local exploit Meta patched a Muse Mac app flaw that required malicious code already running on the device but could redirect transcription processing and give an attacker access to the Muse account.
• Jev follow-up: a command-line plugin Simon Willison's llm-typesafe plugin adds Jev to LLM, his command-line tool for language models, for yes-or-no checks, message routing, and scoring; it requires a TypeSafe API key, and Willison says the waitlist moves quickly.