Why this is in the vault
A hands-on "vibe check" from a working creative studio (Afterimage) testing whether GPT-6's Astra can replace a hired VFX specialist on real footage — evidence for how far agentic multimodal models have gotten at compositing, masking, and tracking tasks that used to require specialist software skill.
The core argument
Afterimage (Cyrus Duff and Josh Lee, the studio behind Every's own video work) ran Astra against Sol on two foundational VFX primitives — masking (isolating a subject from a background) and tracking (following an object/camera across frames) — using real, previously-shot 16mm and iPhone footage. Astra beat Sol on both: cleaner mask edges on a boxer against a grainy 16mm background (using Robust Video Matting, scripted and run by Astra itself) and tighter tracking of a moving pan of food across hand and sleeve occlusions. Neither model nailed either task on the first pass, but Astra's outputs were usable after follow-up prompts where Sol's tracking output was not.
Beyond the primitives, the piece is really about Astra as an agentic orchestrator: it wrote Python to drive third-party tools (Robust Video Matting, Nuke's scripting API), combined generated stills (OpenAI's image model) with procedural animation (swaying rope, punching bag) it coded from scratch, and iterated from the authors' annotated-screenshot feedback rather than needing them to touch the compositing software directly. The authors frame the significance as: earlier models (Opus 4.6) could write deterministic scripts for simple fixes but failed at compositing; generative video models (Gemini Omni) produced convincing pixels but no granular, revisable control. Astra is presented as the first model that gives both — controllable, tool-using VFX work a small studio could plausibly ship, not just a demo reel.
Mapping against Ray Data Co
This is a live data point for RDCO's "agent capability is the actual bet, brands are downstream" thesis (see the L5 north star framing): the interesting claim here isn't "AI made a cool video," it's that a frontier model, given tool access and iterative human feedback, can drive professional software (Nuke, a scripting API) it was never fine-tuned on — the same agentic-orchestration pattern RDCO already leans on for Scribble Works and internal tooling, just pointed at VFX pipelines instead of code. It's a useful comparison point against 2026-02-05-every-codex-vs-opus on how fast "agent drives professional tool via its own scripting" is moving across domains, not just coding.
⚠️ Sponsorship
This issue carries two paid sponsor blocks — OpenAI and Anthropic — both generic "Builder Pack" ad placements unrelated to the article's editorial content (the Astra/Sol comparison itself is Google-model-adjacent branding, not tied to either sponsor). The authors, Afterimage's cofounders, are a for-profit creative studio with an existing paid-client relationship to Every (disclosed in the piece: "Every was our first client"), which is a closer conflict than the ad blocks — their incentive is to showcase Astra's capabilities compellingly since demonstrating agentic-tool competence is their own studio's pitch to future clients.
Related
[[2026-02-05-every-codex-vs-opus]] [[2026-03-27-stratechery-so-long-sora]] [[project_l5_north_star_strategic_direction]] [[project_scribble_works_ops_rules]]