A boutique social agency showed us their OpenAI invoice last week after an account coordinator ran twelve 45-second founder clips through a coding agent pinned to GPT-6.2 Astra.
One afternoon of cutting talking-head videos cost $164.80 in API usage.
Nobody on their team realized that every time Astra inspected extracted video frames and rewrote a local React composition to fix subtitle wrapping, it was billing $8 per million input tokens and $40 per million output tokens across 300,000-token trajectories.
OpenAI’s November 2026 lineup already gives you two much cheaper workhorses: GPT-6 Luna at $0.10 input and $0.50 output per million tokens, and GPT-6.2 Sol at $2.00 input and $10.00 output per million tokens. Both fall apart if you force them to calculate 29.97 fps frame offsets and multi-track audio ducking in raw code, yet both shine when paired with a stage-based video MCP.
Point Codex or your OpenAI MCP client at https://mixi.mixbash.com/api/agent/mcp over Streamable HTTP, authenticate via OAuth 2.1 PKCE or bearer headers, and split your pipeline where it actually makes economic sense.
For straight execution runs, GPT-6 Luna is unbeatable. Give it a folder of phone clips encoded with the mixbash-intake-v1 preset (1080p H.264 CRF 23 AAC) plus local Whisper word timestamps, and Luna cleanly fires get_styles, create_project, prepare_upload, complete_upload, register_media, plan_visuals, and generate_preview for roughly two-tenths of a cent.
Where GPT-6.2 Sol earns its keep is messy human communication. When a founder replies in Slack with three contradictory voice notes about changing the callout text at second twelve while keeping the original pacing, Sol translates that feedback into crisp, structured directives for revise_script or revise_preview at roughly two to three cents total.
Let’s be candid about Luna’s limits so you do not get burned. Luna is fast and literal-minded; if you dump an unedited 1,200-word brain dump of brand guidelines into its prompt and expect subtle editorial synthesis on the client side, its phrasing can feel stiff. Either route creative synthesis through GPT-6.2 Sol or let Mixbash’s server-side draft_script engine author the hook (social, vsl, or explainer) directly.
Regardless of whether Luna or Sol drives the session, Mixbash’s worker harness handles brand typography, word-accurate jump cuts, and automated visual QA on the server, then pauses at get_project_page until a human explicitly unlocks approve_review and render_final.