September 2026 update

In September we extended our education allowance to universities all over the world. Students and teachers at more than 10,000 universities and colleges, in 200 countries, now start with 300 free Nano Banana Pro images, ten times what a normal free account gets. We also changed how you talk to the models. A prompt can now point at the exact reference image it means, so one sentence can take the brick from one photo and the window frames from another. And regional edits now zoom in on the part of the image you paint.

September in numbers

• More than 10,000 universities and colleges, in 200 countries. Students and teachers at every one of them now get our education allowance.
• 300 free Nano Banana Pro images for each of them to start with, ten times a normal free account.
• 21 AI models in production. GPT Image 2.5 came in. GPT Image 2 and FLUX.2 Dev went out.

Shipped

Free for education, worldwide. Until September our education allowance was reserved for our partner universities. Now it covers every university and college on the open, worldwide list of university email domains, more than 10,000 of them. Sign up with your university address and your account starts with 300 Nano Banana Pro images instead of 30. Students and teachers get the same, and each address can claim it once.

Pointing at the reference you mean. You could always attach several reference images to a prompt, but not tell the model which one you meant, so it had to guess. Now you type @, pick an image, and it appears in your prompt as a small badge. One sentence can take the brick from @image2 and the window frames from @image3. It works in FLUX.2, GPT Image 2.5 and both Nano Banana models, as it already did in our video models.

Regional editing that zooms in. Until now, when you painted over a small part of a large image, the model got the whole picture. Now it gets a close-up of the area you painted, so it can put finer detail into the part you are changing. If the edit depends on something far away in the picture, choose Full image instead, or Both to compare the two. Before anything is spent, a preview shows what the model will see and an estimate of the cost. Cancel, and nothing is charged.

GPT Image 2.5. It replaces GPT Image 2, in a quicker version and a more detailed one, both at one price. The material ID pass from August, which lets you click a wall or a window in any image to select it, now runs on the detailed version. On our test images it was more accurate, at about a third of the cost.

Prices that follow the resolution. For several models, a render is now priced by its actual resolution instead of an estimated average. Extra reference images count too, but only the ones a model actually receives.

Two people, one layer. Five people have been able to work on one canvas since July. In September we fixed three ways that could lose work. Some changes, like a video trim, only reached the others after a reload. An edit you had just made could jump back when someone else’s arrived. And undoing a deleted layer could lose some of its settings. Each property of a layer now syncs on its own, so two people changing different things on the same layer both keep their change, and a fresh edit to a note, a group or a widget no longer jumps back.

A safer way to clean up files. Files are no longer deleted by the editor in your browser. Instead, one job on our servers removes only the files that nothing in INTER\STITIAL refers to anymore, and deletes them for good only after more than six weeks, during which anything still needed is put back. For now it only reports what it would remove.

The Visualisations page is gone. It listed every file you had generated, separate from the canvases you used them on. What you work with now lives on your canvases and in your asset library.

Watching what works on social media. For our own marketing, we built a system that follows more than 60 accounts on Instagram, YouTube and TikTok, from other visualization tools to architecture studios and design media, and flags the posts that do far better than that account usually does. Gemini watches each one and notes why it worked, and every Monday we get a list to study. It has been running since mid-September, too short to say what it has taught us.

On the road

RenderTube at TUM. The RenderTube course at TUM’s Chair of Architectural Informatics has come to an end. Architecture students took their own designs to finished images with INTER\STITIAL, and we ran workshops and sat on their reviews along the way. Below is Felix Wichmann’s CoolSchool, an extension to the Türkenschule on Munich’s Schellingstraße that stays cool without air conditioning.

Felix began every image with a plain grey view of his ARCHICAD model. The shapes came only from the model, and the prompt only described material, color, light and people. When he described shapes, the AI redesigned the building. He also noticed that every extra pass through the AI made an image softer, so he kept refining the prompt until a single pass was enough.

The Regelthesis at TUM. This winter semester, TUM’s master’s students who write their thesis without a research topic of their own all work on one shared design brief, the Regelthesis. They are using INTER\STITIAL for it.

B|E Connect, Munich. On 28 September we had our own table at B|E Connect, an evening for construction tech that the Built Environment team at TUM Venture Labs held alongside Bits & Pretzels.

Learned

Finishing takes longer than building. Most of what we listed as next in August is built: saving and sharing pipelines, the agent, placing a building in its real surroundings, and sending the 3D model out of CAD. None of it reached you in September. These are big changes to how you work, and a large part of the month went into them. They go out once they hold up in testing.

Next

• Saving and sharing pipelines. Wire up a workflow once, save it to your library, and send it to a colleague by link.
• Putting a building in its real surroundings. Place your model at its real address, with the sun where it will be on the day you choose.
• Models straight out of CAD. The plugins send the building itself as a 3D model, ready to place on its site. For ARCHICAD this is built.
• The agent, in INTER\STITIAL first. Tell it what you want, and it arranges the canvas, builds pipelines and asks before it spends credits on a render.

Where this is going

One assistant, wherever you work. The agent will work the same way in INTER\STITIAL and inside Revit, ARCHICAD and Rhino. You say what you want in plain words. It looks at what you have open and selected, tells you what it is about to change, and then does the work: a set of facade studies on the canvas, or timber on the facade of the model itself. Everything it did can be undone in one step, and it asks before it spends credits. A studio can teach it its own routines and share them with the whole team.

A full loop between CAD and INTER\STITIAL. Once the building itself comes over as a 3D model, the loop can close. Furniture, fittings and other 3D assets you make in INTER\STITIAL go straight into your Revit or ARCHICAD model. What you decide on the canvas, a new material or a different object, the agent carries back into the model and its drawings. Design and visualization stop being two files that drift apart.

---

INTER\STITIAL is an AI-driven architectural visualization platform built in Munich. We send build updates monthly. Contact