<- all tokdocs
2026.09.25@chase_ai_

The Codex voice demo runs GPT-6 Astra at Ultra effort with "Approve for me" enabled, and the Blender step is the one link OpenAI does not ship

AI
Watch on TikTok

View on TikTok ->

A 54-second 1080x1920 clip uploaded 2026-09-25 by @chase_ai_ (channel "Chase AI"), captured at 2,344 views, 91 likes, 1 comment, 16 reposts, and 63 saves. The caption is two tokens long: @ChatGPT #ChatGPT_Partner. The 183-word Whisper transcript splits into 17 cues. Across 27 frames sampled at 2-second intervals there are zero burned-in captions, so every string on screen is application UI. The inventory: a dark site branded APERTURE with nav items The idea, Inside APERTURE, Explore the demo, hero copy A place for every possibility. and Connect the ideas you already have., a yellow Explore the demo ↗ button, a hint label Disturb the surface, and a footer strip reading A STUDY IN CONNECTION with Brush to scatter · Click to expand · Drag to rotate, later Click to reassemble · Drag to explore, later Scroll to reconnect · Drag to explore, alongside controls Expand sculpture ↗, Reassemble ↗, Reset, and Pause motion; a mockup gallery headed Six living hero directions. with the note Charcoal + acid yellow. Click any mockup to inspect the full image. These are image-generated visual concepts. and six entries 01 / Robotic face ("Intelligence, looking back."), 02 / Precision robot arm ("From intent to action."), 03 / Mechanical iris ("Focus on what matters."), 04 / Magnetic sculpture ("Order, out of possibility."), plus two partly cropped tiles reading Give your ideas shape. and A little more capable; a full-screen voice surface with a blue orb, a composer placeholder Do anything, a permissions control Approve for me, a model chip reading GPT-6 Astra High and later GPT-6 Astra Ultra with Ultra rendered in magenta, a Latest turn bar, and a cropped right rail showing Outputs, Create a…, Sources, Attach f…; a dictated prompt building word by word into Use your image-generation skill and create six mockups for a hero section that we are going to turn into a 3D asset using Blender. followed by Sure, I'll put together those hero mockups. and Thinking; a Blender session with the Global transform header, an outliner reading Scene Collection → Magnetic infinity / 1792 tiles → objects Tile_000_00 through Tile_001_13, a properties stack of Transform, Delta Transform, Relations, Collections, Instancing, Motion Paths, Shading, Motion Blur, Visibility, Viewport Display, Line Art, Animation, Custom Properties, and a timeline running Start 1 to End 250; a browser address bar reading 127.0.0.1:4175 with a second site section labeled 02 / CONNECTED, Give them a place to meet., A single space for your ideas, context and perspectives to find each other., and Look inside APERTURE ↗; and a closing shot of an openai.com page showing a starfield with GPT at left and Astra at right.

The model chip changes tier mid-video, and it names a model that shipped three weeks before the post

The clearest checkable fact in the clip never gets spoken. In the voice-mode frames the composer carries a model chip reading GPT-6 Astra High. Later, over the finished hero section, the same chip reads GPT-6 Astra Ultra, with Ultra colored magenta instead of gray. Nothing in the audio mentions either tier.

Both halves of that string check out against OpenAI's own documentation. The ChatGPT models page lists GPT-6 Astra first and describes it as "Our most capable model for complex work across code, apps, and research." The same page enumerates the effort options as Light, Medium, High, Extra High, Max, and Ultra, and explains that "Ultra mode goes beyond a single-agent run. It uses subagents to accelerate complex work." Wikipedia dates the model to a limited preview on September 3, 2026 and general availability on September 4, 2026. This post went up on September 25, three weeks into the model's life.

One gap survives verification. The models page states explicitly that "GPT-6 Luna supports reasoning efforts up to Max, but not Ultra," and it does not make a matching positive statement that Astra supports Ultra. The screenshot is the evidence that the pairing is real. The documentation only fails to contradict it.

The skill he invokes by voice is a real sample skill in the openai/codex repository, and its description matches his use case word for word

The dictated prompt on screen reads:

Use your image-generation skill and create six mockups for a hero section that we are going to turn into a 3D asset using Blender.

Whisper heard "image generation skill" and "mock-ups." The UI shows image-generation skill hyphenated and mockups closed up. The difference is cosmetic, and it matters only because the hyphenated form is closer to the actual artifact name.

OpenAI's skills documentation defines Agent Skills as packages that "package instructions, resources, and optional scripts so either product can follow a workflow reliably," built on the open agent skills standard. It lists two bundled system skills, Skill Creator and Skill Installer, and no image generation skill. That absence is misleading. The openai/codex repository carries a bundled sample skill named imagegen at codex-rs/skills/src/assets/samples/imagegen/SKILL.md, described as: "Generate or edit raster images when the task benefits from AI-created bitmap visuals such as photos, illustrations, textures, sprites, mockups, or transparent-background cutouts."

The word "mockups" sits in the skill's own description. He asked for mockups, and the skill exists to produce mockups. The docs also explain that skills fire two ways: explicitly, by typing @ in ChatGPT or $ in Codex, or implicitly, when the model matches the task to a skill description. Speaking "use your image-generation skill" out loud is neither of the explicit forms. It relies on the implicit path, which is why he can name a skill without touching the keyboard.

OpenAI ships the voice layer and the approval layer; the Blender bridge comes from somewhere the video never shows

Three of the four moving parts are first-party and documented.

The voice surface is ChatGPT Voice in the desktop app. OpenAI's docs say it supports "natural turn-taking" so you can "interrupt ChatGPT during a response, ask a follow-up, or change direction," that it runs on the GPT-Live model, and that it is available on Plus, Pro, Business, Edu, and Enterprise. The docs confirm it can "start a Codex task" and "check active tasks."

The Approve for me chip in the composer is also documented. OpenAI's auto-review page describes it as a permissions mode where "selecting an approved Daybreak model automatically switches the permissions control to Approve for me when that mode is available for your account and allowed by organization policy." Auto-review swaps a human approver for a separate reviewer agent at the sandbox boundary, keeping the same network and filesystem limits. The page calls it "a reviewer swap, not a permission grant." Turning that on is what makes a long unattended agent run practical, and it is the setting that quietly does the most work in this demo.

The Blender step has no first-party equivalent. He says:

So I had Codex use Blender for me. I've never used Blender before, but it handled everything.

Blender appears in the frames as a running desktop session with a populated outliner, so something drove it. OpenAI publishes no Blender integration. Every Codex-to-Blender path that turns up in search is a community MCP server or a third-party agent skill. The clip cuts from the voice surface straight to a finished 1,792-tile scene and never shows the connection. A viewer who tries to reproduce this stalls at exactly that cut.

The site he "deployed" is running on localhost, and the transcript claims a step the frames do not contain

Three separate frames show the browser address bar reading 127.0.0.1:4175. That is a loopback address on a port in the range Vite hands out for local dev servers. The site rendering there is the finished APERTURE page with the working 02 / CONNECTED scroll section.

The transcript says:

I refined it. I deployed it. I created everything through Codex's new voice chat.

No frame shows a deployment. No frame shows a public hostname, a build log, a hosting dashboard, or a non-loopback URL. Every shot of the running site is local. The refinement claim is supported, since the hero sculpture visibly changes state across frames from intact torus to scattered tiles to reassembled infinity loop under the Reset and Reassemble ↗ controls. The deployment claim rests on the narration alone.

The Blender evidence is stronger and more specific than the narration suggests. The outliner label reads Magnetic infinity / 1792 tiles, and the individual objects are named Tile_000_00 upward in a two-index scheme, which is the naming a script produces rather than a human clicking through the UI. The timeline runs Start 1 to End 250, Blender's default 250-frame range left untouched. Those details corroborate that code generated the scene. They say nothing about who wrote the code or through what channel.

The hashtag is the disclosure, and the engagement numbers are small for a partner post

#ChatGPT_Partner is the entire caption apart from the @ChatGPT mention. It is a paid-partnership disclosure, required of US influencers by the Federal Trade Commission when they have a material connection to an advertiser.

The tag belongs to an active campaign. HypeAuditor data cited by AndroidHeadlines counts sponsored Instagram posts carrying #chatgptpartner or #chatgpt_partner rising from 61 in June to 122 in July to 141 in August 2026. The same report says OpenAI hired Charles Porch, formerly Instagram's head of partnerships, as VP of Global Creative Partnerships, and engaged the agency Viral Nation to run creator campaigns at scale. Most of that campaign targets parenting, fitness, travel, and education. A Codex and Blender workflow is an outlier in it.

At capture the post held 2,344 views against 91 likes, 63 saves, 16 reposts, and 1 comment. The save-to-like ratio of roughly 0.69 is high, which fits a tutorial people intend to return to. The absolute numbers are modest for a sponsored placement, which is worth holding in mind when reading the closing line, "So if you haven't tried out Astra yet, now is the time."

Key Takeaways

  • The model chip in the composer reads GPT-6 Astra High early in the clip and GPT-6 Astra Ultra later. OpenAI's models page confirms Astra exists, confirms Ultra is an effort tier that "uses subagents to accelerate complex work," and confirms Luna is excluded from Ultra.
  • The spoken instruction "use your image-generation skill" maps to imagegen, a sample skill bundled in the openai/codex repository whose description explicitly covers "mockups."
  • Approve for me is the most consequential setting visible and the one never mentioned in the audio. OpenAI documents it as an auto-review mode that replaces human approval with a reviewer agent at the sandbox boundary.
  • ChatGPT Voice is documented as able to "start a Codex task" and "check active tasks." That is narrower than the clip's claim of building, refining, and deploying a site end to end without a keyboard.
  • The finished site is served from 127.0.0.1:4175 in every frame that shows it. "I deployed it" has no visual support.
  • Unverified: how Codex reached Blender. OpenAI publishes no Blender integration, the frames never show the connection, and the only candidates found are community MCP servers and third-party agent skills. Which one he used, and whether the model wrote Python against Blender's API or drove a tool bridge, cannot be determined from the video.
  • Unverified: whether GPT-6 Astra officially supports the Ultra effort tier. The docs name Ultra as a tier and name Luna as excluded from it, without stating Astra's support either way. The on-screen chip is the only evidence.
  • Unverified: what @chase_ai_ was paid or what deliverables the partnership specified. The disclosure hashtag establishes a paid relationship and nothing about its terms.

Resources

Published September 25, 2026. Writeup generated from a favorited TikTok.