How invideo improves color grading 3x with GPT-6 Astra
OpenAI published a customer story about invideo, an agentic video editor built to handle editing work while keeping the human editor in control. According to the account, invideo integrated GPT-6 Astra and saw it plan and execute complex edits with greater precision, placing changes at the correct points on the timeline. One line in the piece calls Astra's frame-level planning accuracy "quite stunning." invideo's Sanket says the model needs far fewer reasoning steps and output tokens than the models the company had previously tested to complete complex editing work. invideo also observed that Astra can follow several instructions in sequence without losing track of the editor's original goal as a task grows longer and more complex. The story singles out color work as an area where Astra is especially effective, since the agent has to choose among overlapping approaches such as correction, grading, regeneration, LUTs and isolation, for example isolating and tracking a person across frames to change a background while preserving their skin tone. Sanket says that before Astra, invideo saw very high failure rates on color-grading and color-correction tasks, and that with Astra the success rate improved about three times. GPT-6 Astra also helps invideo turn text descriptions and visual references into custom effects, coding an effect for the footage, placing it on the timeline and adding controls the editor can adjust; in one day, a few invideo editors used the model to create about 50 such effects.
Key facts
- invideo, an agentic video editor that keeps the human editor in control, integrated OpenAI's GPT-6 Astra
- invideo's Sanket says the success rate on color-grading and color-correction tasks improved about three times (3x) with Astra, after very high failure rates previously
- A few invideo editors created about 50 custom effects in one day using Astra to turn descriptions and references into coded, timeline-ready effects
- Sanket says Astra completes complex work using far fewer reasoning steps and output tokens than the models invideo tested before
- invideo observed that Astra can carry out multiple instructions without losing the editor's original objective as tasks grow longer and more complex
Why it matters
This is OpenAI positioning GPT-6 Astra's agentic reasoning as capable of measurable gains on a narrow, technically demanding professional task: color grading and correction in video editing, where an agent has to pick among overlapping methods like correction, grading, regeneration, LUTs and isolation and keep track of a longer sequence of steps without drifting from the original brief.
Who it affects
Video editors and post-production teams using agentic editing tools, invideo as a company building on GPT-6 Astra, and other software makers weighing whether to build editing or creative-effects agents on top of OpenAI's models.
How to use it
invideo is an agentic video editor that runs on GPT-6 Astra to plan and execute edits while the human editor stays in control; editors can also describe or reference an effect and have Astra code it, place it on the timeline and expose controls to refine it. The source gives no pricing, release date or technical detail about GPT-6 Astra itself.
How solid is it
This is a single first-party customer story published by OpenAI, with the specific figures, the 3x success-rate improvement and the roughly 50 effects, attributed to invideo and to one named speaker, Sanket, whose title and role are not given. No date is stated for when the results were measured, and the numbers are not independently verified.
Risks and caveats
The account is self-reported by invideo and framed as promotional material for GPT-6 Astra; it does not break down the 3x figure by individual color-work task, does not give a baseline sample size, and one quote about Astra's frame-level planning is not attributed to a named person.
“Earlier, we saw very high failure rates for color-grading and color-correction tasks. With Astra, the success rate improved about three times.”
— Sanket, invideo