Three ideas, nine builds: GPT-6.1 Sol, Claude Opus 5.5 and Gemini Argon 4

Posted on Oct 20 • Originally published at billiem.uk

The bell-ringing prototype I preferred was not the technical reviewer’s winner. Claude’s cascading timing challenge felt nicer to me; the review favoured Sol’s more reliable connection between the blows entered and the feedback returned.

That disagreement is a useful way into these nine builds. GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Argon each implemented the same three browser briefs: a gift sleeve, a bell-ringing practice room and a colour-context experiment. The differences are visible, but the most convincing interface is only one part of the result.

All nine are public comparison prototypes. The original comparison on billiem.uk contains the fuller idea round and method; the links here open the working versions.

In the ringing brief, six synthesised bells play in rows. You tap one bell’s part, follow a visible path, then try to manage without the guide. That leaves quite different things for a builder to emphasise: the timing interaction, the teaching or the record of what you actually struck.

Claude’s Ring had the movement I liked. The pattern cascades down the screen into a timing challenge. Claude’s interfaces and animations felt the cleanest in this batch, and this was a particularly clear example of the feel I preferred. The still below shows a real running stage; judging its movement needs the live version.

Claude’s Plain Hunt stage during an automated capture run with no rope taps. The missed marks are genuine; this still shows the cascading layout, not successful playing or a timing assessment.

The Sol implementation is a smaller practice room. Its advantage in the Codex review was accounting: entered blows, fading assistance and final feedback stayed connected. Claude could react to extra pulls during play and then omit them from the final account. A pleasing challenge and an honest result are separate things to check.

Sol’s Plain Hunt practice room before starting: setup, the full blue-line path and the strike button share one panel.

Gemini’s Ring takes the broader-console route. I found Gemini’s interfaces incredibly noisy, but the review’s objections went further: reversed dodge-direction teaching labels and an idealised blue line presented as the visitor’s performance, even after misses. Those are correctness problems, not a preference for one palette over another.

Gemini’s ready-state method diagram layers notation, view choices, a guide meter and two labelled bell paths. No attempt has started.

The mapped Ring review explains those findings. Its browser and synthetic-performance checks do not establish that a person learned to ring, or that the sound is good. My reaction here concerns the interface and timing feel.

For the gift sleeve, the relationship between flat paper and finished shape matters. The brief also includes a message revealed on opening. A folding animation can show the object taking shape and make that reveal part of the experience.

Claude’s Fold was far ahead for me. It wasn’t close. The review also preferred its connection between object dimensions, net, opening box and readable lid message. Here, the design preference and implementation review agreed.

Claude’s box after moving the fold slider to the end and choosing Open it. The default message is visible inside the lid; this is the browser preview, not a made paper box.

Sol’s Fold is an approachable paper workshop, but its animation omits the message and actual slot insertion. It communicates the sleeve more fully than the personal opening moment.

Sol’s preview keeps the sleeve, fold control and print step together. Shown with the default dimensions; the digital view does not establish physical fit.

Gemini’s Fold offers linked geometric views, dimensions and plenty of controls. There is real material to inspect; it also gives the simple gift-wrapping activity a much noisier surface.

Gemini’s default preview pairs the labelled folding geometry with its sheet dieline. The small print is shown for layout context, not as a usable print template.

The Fold review found practical defects in every version: stale browser printing in Claude, tiny accepted message text in Sol, and calibration or margin faults affecting Gemini’s pattern. No paper was printed and folded in that review. The best box experience is still not a physically proven box.

Blue Without Blue proposed a vivid blue-looking cloak that would reveal neutral grey through a pinhole. None of the three made the grey patch look blue to me.

I probably preferred Sol’s Blue, but only slightly. The picture, pinhole and paint activities are clear. It also adds unnecessary text and explanation. That is a tendency I have noticed with GPT models when given free rein beyond this batch too; it remains my reported impression, not a finding measured across models here.

Sol’s default portrait study and pinhole controls, with its explicit limit on the promised colour shift. The capture records the interface, not the viewer’s perception.

Claude’s Blue felt tighter and cleaner to me. Preferring Sol overall was a close choice, with pros and cons. The technical review found useful exploration in Claude’s workbench, alongside overconfident claims about greys and effects that would read as blue.

Claude’s default viewing-card workbench, with the portrait, isolation controls and sampled RGB values. The card is raised; the readout does not measure a viewer’s experience.

Gemini’s Blue was unpleasant and noisy to look at. Its instrument-style presentation also surrounds the picture with asserted perception values. The Blue review found no observer measurements or calibrated appearance model supporting its cortical-perception colours and induction score. The displayed 100/100 is the interface’s claim.

Gemini’s default portrait and instrument-style readouts. The blue swatch and 100/100 induction index are claims made by the interface, not measured human perception.

Sol was the review’s preferred illustrative studio. That did not make its grey look blue for me. Working controls, preserved pixels and a preferred layout leave the subjective payoff unresolved.

Before implementation, each model proposed five ideas. These sets were broader than a neat “tools versus games versus science” division:

Sol nominated its water simulation; I chose Fold instead. Claude’s Ring and Gemini’s Blue nominations matched my selections. Then all three built each selected idea. Fiction, physical making and learnable systems overlap between the sets; fifteen pitches do not establish permanent model personalities. Their scientific and historical promises are proposals, not verified outcomes.

The implementations came from nine isolated fresh chats at High, using native Sol and authenticated subscription CLIs for Claude and Gemini. They shared each topic’s brief and one ordinary attempt without a fixed duration, while tools, harnesses, system instructions and compute differed. Quota interruptions, same-conversation recovery, missing packaging supplied by Sol and small hosting compatibility edits are documented in the original. Known scoring, printing and scientific faults were retained. Seven versions received private or unfinished recommendations from the anonymous Sol reviewers; all nine were published as comparison prototypes.

Claude’s box remains the decisive one for me. Ring leaves a useful disagreement about feel and reliable feedback. Blue leaves three interpretations of a promised effect I did not see.

Want to talk about something I’ve written or built? Get in touch.

This article was adapted with AI assistance from an original article on billiem.uk. The original article was reviewed before publication.