📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
Summary: On November 4, OpenAI shipped Images 2.6 across ChatGPT, Codex and the API: generation 50% faster, GIF output, reusable reference images, and edits confined to whatever you mark up. The company says 4 billion images now come through these surfaces every week. Speed and motion sit on the surface. What moved underneath is image work going from one roll of the dice to rounds of edits, and the costly part of rounds is your presence, the gap KAIHE AIBOX（铠盒AIBOX）fills every day.
What shipped
Three surfaces at once: ChatGPT, Codex and the API, generation 50% faster. The API splits into two models. Flare optimizes for speed and returns an image in as little as six seconds at low quality. Sunburst optimizes for detail, holding characters, products and brand elements together across rounds.
Four capabilities are new: GIF animation, reference images, markup-guided edits and SVG export. A complexity parameter joins them with low, medium and high settings, defaulting to high, controlling how much planning happens before rendering. Adjusting it costs nothing extra.
Why GIFs matter more than they look
The published use cases are mascots in motion, product shots turned into loops, and animated storyboards that test movement before a shoot. That reads like a social-media toy. What it replaces is a category of asset. Producing anything that moves used to mean shuttling between an image tool and a video editor. Now a prompt asks for a six-frame sprite sheet in three columns by two rows, the sixth frame flowing back into the first.
Whether half the time is worth it depends on price
The detail easiest to misread: the price did not move. Sunburst and Flare both bill $8 per million input tokens, $30 per million output tokens and $2 for cached input, identical to the two models before them.
Same money, half the wait, so one budget now buys half again as many rounds. For a team producing hundreds of images a day that is real. For someone making one image now and then, six seconds and twelve seconds feel alike.
Reference images turn a slot machine into an editing session
The workflow change comes from reference images plus markup edits working together. Save a brand asset once and reuse it every round. Circle an area, type "make this blue", and the model touches that area alone. Lower complexity for a quick fix, raise it for dense typography and complicated layouts.
Image work goes from writing a prompt and rerolling until something lands to laying down a base and refining it round by round.
The money of rounds is easy to count. The time is not.
| The job | Time per attempt | Rounds a day | Someone present? |
|---|---|---|---|
| One poster | 6-20 seconds | a dozen tweaks | every round |
| A batch of product shots | seconds each | a hundred or more | watching the queue |
| One GIF set | six frames per pass | repeated motion tests | picking each pass |
| Brand asset templating | minutes per round | across several days | machine on daily |
The last row is where everyone lands. More rounds mean longer tasks, and longer tasks lean on a machine that stays on. OpenAI got a single frame to six seconds. The other half of the cost is your presence.
Where this lands
KAIHE AIBOX is a dedicated agent computer with the agent system built in, running around the clock on its own without tying up your laptop.
It routes work to the right tier: bulk drafts to Flare, finished assets to Sunburst. Tasks set during the day run overnight. Cloud models cost less off-peak, and the agent schedules work into those windows so the difference lands on the bill. Hardware and agent system are bought once, no recurring software subscription.
What is scarce has shifted. Not whether an image can be made, but whether a machine will sit through those dozen rounds on your behalf.
The lines OpenAI drew itself
GIFs are the new risk surface and the system card is blunt about it. An animation carries motion and cross-frame meaning a static frame does not, so a dedicated layer reviews the caption alongside the frame grid.
A targeted evaluation of 310 flagged GIFs reached 77.66% recall. OpenAI's own wording: violations tied to motion stay hard to catch and the safeguards will miss some. Neither model crossed the high-risk thresholds for biosecurity or cybersecurity.
Related Reading: - Does My AI Keep Working When My Computer Is Off? - Too Many Small Things to Remember? Hand Them Over - Does a Personal AI Assistant Need Its Own Computer?
KAIHE AIBOX #Images 2.6 #OpenAI #AI frontier #agent computer #local AI
For more information, search [KAIHE AIBOX] or contact: [email protected]
KAIHE AIBOX · 7x24 Personal AI Assistant | AI前沿