GPT-5.6 needed a more ambitious prompt

I moved LLM Choice over to GPT-5.6 Sol expecting the difference to be obvious. The first twelve experiments were boring.

They looked more polished than the older work, but there was not much going on underneath. The ideas were shallow, the pages were full of familiar panels and controls, and most of the runs were over in about five minutes. The repository records back that impression up: the twelve artefacts took between 3 and 7.9 minutes each, with an average of 4.8 minutes.

That was not the first impression I expected from a model that OpenAI describes as a step forward in agentic coding. My experience of the later runs is that Sol is significantly better than GPT-5.5 for this kind of open-ended work. Getting there still required a better answer to what the model should spend that capability on.

The first prompt rewarded finishing

The original version of LLM Choice was deliberately loose: choose something, make an inspectable artefact, and leave me an idea I can react to. For the first GPT-5.6 batch, I had Sol adapt that prompt for itself rather than simply reusing a GPT-5.5 prompt.

It still asked for twelve independent browser artefacts in one batch. Each needed a concrete interaction and a verifier, and each was given the same nominal 35-minute budget. In practice, Sol treated the first working object as something close to the finished object. It made all twelve quickly, dressed them well, and moved on.

My reaction was blunt: “It’s just boring. We need to think about how we’re going to make experiments more interesting.”

There will be variation in this project. Every experiment has a different degree of interactivity, a different quality of idea, and a different amount of exploration behind that idea. The problem was that the prompt gave all twelve roughly the same shape and encouraged throughput. A stronger model made that throughput look nicer. It did not automatically make the results deeper.

The next prompt tried to protect depth explicitly. It separated experience mode, ambition, finish, and conceptual depth instead of treating “good artefact” as one vague target. A compact jewel could stay compact, while a flagship needed its own builder and more room. Substantial systems needed a real engine and at least two deepening passes after the first working version. Every idea needed a sentence explaining why it deserved to exist, and visual polish was not allowed to rescue a weak concept.

That changed the working rhythm. The next twenty Spectrum V2 runs averaged 20 minutes each, rather than 4.8. That is still absurdly quick for what some of them became, but it is enough of a difference to show that this was not just the same test with a different model name.

It also means this is not a controlled GPT-5.5 versus GPT-5.6 comparison. Sol was running at ultra effort. The prompt changed, the effective build time changed, and the later batch added concept review and portfolio curation. I am describing the system I used and the things it made, not pretending I isolated one variable.

Not everything in the second batch was good

The later GPT-5.6 gallery is still mixed, and I do not want to flatten all of it into praise.

Single Pixel Camera is an interesting concept, reconstructing an image from one scalar brightness measurement per mask, but I only found it mildly interesting as an artefact. Resonant Room was more engaging because I could resize a room, move the source and listener, and actually hear the result change. Lightning’s Preference looked pretty and made diffusion-limited growth visible, but again I would put it in the mildly interesting group. Glasshouse Mendel was similar for me: a reasonably neat idea, not one of the pieces that changed my view of the model.

One Bird Knows did not work for me at all. The three-dimensional flock loaded on my Mac, but I could not start it or press any of its controls. The page was just dead. I do not know whether that is the artefact, my browser, macOS, or something more specific, so I am not diagnosing it from one failed visit. It is still a useful reminder that an elaborate verifier and a polished surface do not guarantee that the thing in front of a particular reader works.

The strongest pieces went much further

Harbour of Three Walls was the first one that properly impressed me. It gives you exactly three breakwaters to move around a simulated harbour. The interesting bit is not simply making the water calmer. You also have to preserve a usable route for a vessel, so sealing the harbour mouth produces the quietest water and the worst harbour. The seabed affects the wave field too. It feels like an actual small system rather than a visual effect with some sliders beside it.

Deep Time Radio follows signals from Voyager 1, Voyager 2, and New Horizons through distance, delay, dwindling spacecraft power, and the infrastructure listening on Earth. It is a cool little visualisation, but it is also an informative piece. It uses a frozen JPL geometry snapshot and includes the NASA and JPL sources at the bottom instead of asking the presentation to carry unsupported facts. Sol did not just find a space theme and animate it. It did more work to get the idea somewhere coherent.

Caustic Smithy is perhaps the clearest example of what I like about the new set. You can move lenses, prisms, mirrors, a source, and a screen, then watch the light react immediately. It is attractive because it is interactive, not because it has a decorative animation playing behind a dashboard.

I can imagine how useful something like this would have been when I was studying physics and mechanics at secondary school. Being able to make little simulations on the go, while also having the equations and limitations explained, would have been really, really cool. This is where LLM Choice starts to suggest something beyond a gallery of curiosities.

Black Glass pushes the presentation further again. It lets you aim light near a Schwarzschild black hole and see neighbouring rays escape, loop, or disappear. It has a genuinely spatial, three-dimensional feel that I was not getting from the GPT-5.5 artefacts. The interaction makes it instantly prettier, but there is also a real model underneath the spectacle.

That combination is the improvement I notice most. When GPT-5.6 has an idea for one of these simulations, it seems to do much more to get the idea there and make it feel complete.

Sol appears to prefer systems, with a large caveat

Looking across the published gallery, the subject and form have shifted. The 31 GPT-5.5 artefacts include 14 tagged as atlases and 20 as systems. Among the 32 GPT-5.6 artefacts, only four are tagged as atlases and 26 are systems. Tags can overlap, but the direction is clear.

The older gallery contains more collections and generated worlds: glyph families, invented lexicons, transit networks, estuaries, herbariums, and memory maps. The newer work is much more interested in instruments built around physical or mathematical mechanisms: geometric optics, room acoustics, wave shelter, CT reconstruction, resistor networks, minimal surfaces, orbital light paths, and temperament.

That looks like a model preference, but I would not claim that from these runs. The Spectrum V2 prompt explicitly required a planned mix of play, exploration, tools, and observational pieces. It asked for real questions, meaningful interaction, deeper systems, and fewer decorative dashboards. The first batch also used preassigned scopes. Sol was not choosing in a vacuum in either case.

The most I can say is that Sol plus the new prompt seems very comfortable turning exact mechanisms into interactive instruments. Whether Sol itself prefers those subjects is something I would need to test with matched prompts, effort levels, and the other GPT-5.6 models.

It still very clearly looks like Sol

The visual language is recognisable across nearly all of these pages. There are gradient backgrounds, distinctive display fonts, glowing effects, coloured buttons, and dense control panels. It is almost still a generic LLM style, but it has its own taste. Compared with the GPT-5.5 work, it often feels like a more polished version of the same instinct.

That polish can become a problem. There is a lot of information on the pages and often a lot of options. The model is doing more, but it is not consistently deciding what the reader needs first. Some artefacts open as a field you can understand immediately. Others still make you decode a wall of interface before you know what matters.

My next prompt change may need to talk about UX as directly as Spectrum V2 talks about conceptual depth. I do not want that to mean making every artefact sparse or identical. I want the model to consider hierarchy, onboarding, and the cost of every control instead of assuming that more information proves that more thought happened.

I will probably look separately at Sol’s different effort modes, and at Terra and Luna, because ambition, idea quality, visual language, and execution may move differently across them. For now, the more interesting result is smaller: GPT-5.6 did not save a shallow prompt, but once the prompt gave it room to deepen an idea, the ceiling moved a long way.

The next thing I want to test is whether I can make these artefacts more useful without choosing their subjects for them. That keeps the dimension of choice intact while putting more pressure on UX and usefulness, not just visual capability.