Invideo: GPT‑6 Astra triples color-grading success

Invideo is an agentic video editor built to carry out editing work while leaving the human editor in control, and its leadership describes a workflow where the AI agent must turn an editor's direction into steps, pick tools, execute, and check the result. The company says that with GPT‑6 Astra it can now plan and place changes at the correct points in the timeline with frame-level accuracy, which matters most in color work — a step where correction, grading, regeneration, LUTs, and isolation overlap and the agent must isolate and track a person across frames before applying color elsewhere.

The reported gains are comparative: against the models invideo tested previously, Astra is said to use far fewer reasoning steps and output tokens and to hold the original objective as a task grows longer and more complex. Sanket Shah describes earlier failure rates on color-grading and correction as very high, with Astra improving the success rate about three times, and notes roughly 50 custom effects built by a few editors in a single day.

What the outcome hinges on is whether these measured gains translate into day-to-day editing, where the value invideo claims is that creative professionals keep control of the story, their taste, and the final cut without losing hours to planning and frame-by-frame work. That is the test worth watching, and the body does not say when or how broadly these results have been rolled out.