Claude Sonnet 5.5: from a tiny emoji to a 3D print

What Sonnet 5.5 brings

Anthropic released Claude Sonnet 5.5 on 28 September 2026. The company positions it for clearly defined everyday work and quick revisions, with Opus 5.5 remaining its stronger option for complex, open-ended tasks.

The headline claims are over 30% faster output and up to 30% lower cost per task than Sonnet 5, because fewer tokens are needed. API rates stay at $2 per million input tokens and $10 per million output tokens. These are Anthropic’s measurements; a cheaper completed task is different from a reduced token price.

Sources: Anthropic [1] · Claude / Anthropic [2]

A 60 × 46 pixel emoji, then a physical object

In Anthropic’s showcase, @LEJ_limited gave Sonnet 5.5 a 60 × 46 pixel emoji and asked for a printable 3D model. According to the post, the object was on the creator’s desk a few hours later. It is the most immediately understandable example in the thread: the output has crossed from a screen into a physical form.

A tiny reference leaves plenty for the maker to interpret. This example is interesting for that translation of shape, rather than as proof that a model can manufacture anything. The post does not provide the complete modelling and printing process.

Sources: Claude / @LEJ_limited [6]

A forest whose image and sound are code

@measure_plan’s example renders light through a forest canopy in real time. Anthropic identifies both the visuals and the sound as written in code. That detail changes how to read the clip: you are looking at a programmed audiovisual scene, rather than evidence that Sonnet is a video-generation model.

For visual artists, creative coding offers a different kind of material to work with. A scene can be expressed through rules and parameters. The showcase gives that idea a clear subject: light, atmosphere and sound occupying the same environment.

Sources: Claude / @measure_plan [4]

The same subject, two generations of Sonnet

The thread opens with @_re_pete’s fall foliage simulator, labelled as a comparison between Sonnet 5 and Sonnet 5.5. It is a useful format because both versions deal with a recognisable subject instead of unrelated prompts.

Watch the two results for composition, the treatment of leaves and the way the scene moves. Those visible decisions explain a creative result more clearly than a broad claim that a model has “better taste.”

Sources: Claude / @_re_pete [3]

Dinosaur history, animated in JavaScript

Kevin Ngo’s example tells the history of dinosaurs in under a minute. Anthropic describes it as JavaScript animation made from the same prompt with Sonnet 5 and Sonnet 5.5. Here the comparison is about a sequence: the model has to arrange images and motion over time.

It is a compact example of coded storytelling. The clip lets you compare visual presentation and pacing; historical accuracy would need its own source check.

Sources: Claude / Kevin Ngo [5]

A sketchbook with working controls

Another Kevin Ngo example turns a hand-drawn graph-paper aesthetic into a functioning flowchart tool. The post describes dragging shapes, connecting them and sketching on top. The style belongs to an interface you can use, rather than a static mock-up.

The graph-paper treatment suits the task: shapes, connections and freehand marks share the same visual language. It gives the diagram editor a clear identity while keeping its controls visible.

Sources: Claude / Kevin Ngo [7]

Brick Studio: a build with a clock attached

CodeRabbit asked Sonnet 5.5 and Opus 5.5 to build Brick Studio: a brick-model designer with an Alpine Chalet example and a 45-second showcase video. Sonnet finished in 29 minutes 27 seconds; Opus took 44 minutes 50 seconds. CodeRabbit described the results as close, with Opus slightly higher in fidelity.

The report also records a shared-folder complication and says neither model received the reference screenshot. That makes this a documented single build, not a clean verdict on which model always wins. It is still valuable to have an actual deliverable beside the timings.

Sources: Hendrik Krack / CodeRabbit [8]

The test that did not produce an image

Simon Willison’s launch-day experiments add a useful counterpoint. A request for an SVG pelican on a bicycle exhausted 128,000 thinking tokens at Max effort without producing the SVG. At Xhigh, it returned the SVG in 41 seconds.

He also published a WebGL version generated through Claude’s free tier; the interactive page is linked below. These are different requests, with different outcomes. Their value is precisely that the author records both the successful result and the failed attempt.

A higher effort setting means more room to reason, not a promise of a better image. For someone paying for repeated attempts, that distinction can matter as much as the best-looking demo.

Sources: Simon Willison [9] · Simon Willison [10]

What I would take from these examples

The strongest story here is the range of things people are making: a printable shape, an audiovisual scene, an animated explanation and a usable design tool. Many of these results are built through code, even when the result looks nothing like software.

Anthropic’s thread is a curated showcase. It can help you find an idea worth exploring; it cannot tell you how many attempts your own project will need. Choose a result you actually want to make, then judge the working version. The useful question is what survives beyond the clip.