GPT-6 Astra vs Claude Fable 5.1: Faster on All Three

For three months we have been handing the same build briefs to frontier models and publishing whatever comes back, untouched. Until now every entrant has been a Claude model. This time we ran the same tasks against OpenAI's newly released GPT-6 Astra and put it directly against the model that won our last round, Anthropic's Claude Fable 5.1.

Three briefs: a top-down motorway interchange and a seaside fairground, both as single animated SVG files, and a cinematic rocket launch as one web page. One attempt each, maximum reasoning effort, and neither model was allowed to open a browser to check its own work. All six builds run live below.

Astra was faster on all three rounds, by a wide margin every time. It lost 10–9.

One disclosure before the builds: Fable 5.1 ran as a subagent inside Claude Code with file-write access, while Astra was run by hand in its own environment. Identical prompts, different scaffolding. That is worth knowing, though nothing below turns on it — the faults that decided this are in the output, not the harness.

Round 1: The interchange

The brief: a top-down view of a busy interchange as one self-contained SVG. Two roads crossing on different levels, connecting ramps, at least twelve vehicles following the curve of the road they are on, and correct layering so traffic on the lower road is hidden by the overpass. It is a brief with two traps that go unstated — whether the ramps genuinely connect the two roads, and whether traffic flows the right way after merging. Both have caught models before.

Astra clears both hidden traps. Its bridge is real — a deck with a shadow beneath it, parapets, bearings at all four corners, and lane markings that stop cleanly at the deck edge. We caught an articulated truck entering from the west with its cab already hidden behind the parapet while its trailer was still in the open, which is the occlusion working exactly as it should, mid-crossing. And every one of its vehicle paths begins off-screen, so ramp traffic genuinely arrives from the mainline rather than appearing out of nothing.

It also does things nobody asked for. Every vehicle throws a headlight cone. Direction arrows are painted onto the road surface, which has the pleasing side effect of making the traffic-direction trap visually checkable. There is street lighting along the median, road signage, an articulated lorry and a bus among the traffic.

Its one fault is small and specific: the white shoulder line runs straight through the exit, so the ramp is never visually separated from the road it leaves. That is precisely the mistake Claude Opus 5 made in our last round — the same error, from a different lab. Fable 5.1 is the only model that has ever got this detail right.

Round 1 to Astra, 4–2. Faster, denser and more inventive, with a smaller flaw than its rival's absence of flaws could overcome.

Round 2: The fairground

A seaside fairground at dusk, one SVG. A Ferris wheel with at least ten gondolas turning continuously, a spinning carousel with at least six horses rising and falling out of phase, twinkling bulbs, a reflection in the water. The stated trap is the gondolas: each must stay level with the horizon throughout the rotation, which requires a second animation running backwards against the wheel.

Both models clear the gondola trap cleanly. Astra's wheel turns a full circle every 48 seconds and each of its gondolas counter-rotates the same 360 degrees over the same 48 seconds, cancelling it exactly, with a gentle sway layered on top. Watch it for a while and every cabin stays dead level.

Then look at the carousel. Astra's does not turn. We pulled apart all 37 of its animations: thirteen full rotations, every one of them belonging to the Ferris wheel; eleven vertical translations, which are the horses bobbing on their poles; twelve gondola sways; and one remaining animation — a three-degree rock back and forth about the carousel's centre, over eight seconds. That is the entire ride. The canopy rocks, the horses bob on the spot, and nothing revolves.

Which explains the thing you notice before you can name it: the horses never face a direction of travel, because there is no travel. The brief asked for a spinning canopy. It got a rocking one.

Round 2 to Fable 5.1, 4–3. Not a cosmetic slip — a missing mechanism.

Round 3: The rocket launch

This is the only brief in the series that does not say how to build. It names no library, does not ask for 3D, and offers exactly one aesthetic instruction: make it look good. When we first ran it in June, both models reached for Three.js and WebGL from a CDN without being told to.

Both launches below start on a countdown and play through on their own.

Astra 6 — 28 minutes

Fable 5.1 — 49 minutes

Neither used a library. Both hand-built the whole thing on a 2D canvas. Counting our September run, that is now four models from two different labs, all given explicit permission to pull a 3D engine off a CDN, all declining. In June, on the identical brief, it was two out of two reaching for one. Something has shifted in what these models consider the obvious tool, and it has shifted in the same direction at both companies.

Astra builds the better-looking product. A three-core vehicle rendered with real care, enormous editorial typography, the curvature of the Earth below, a flight-sequence timeline, a zoom control and a playback-speed multiplier that nobody requested. Its physics are sound where they exist: 38.5 km at 1,003 metres per second by T+1:39 tracks a real ascent closely, and the g-load climbing from 2.13 to 3.29 as propellant burns away is correct behaviour rather than a decorative number.

And then it is over. Astra's flight sequence reads countdown, ignition, liftoff, max-Q, ascent — and stops. There is no main-engine cutoff, no stage separation, no second-stage burn, no orbital insertion. Throttle stays pinned at 100% and the vehicle simply keeps climbing until the animation runs out. It is a first act with no second one.

Fable 5.1's is the less striking image and the more complete flight. It carries the staging its rival lacks, and it computes the speed of sound from altitude rather than dividing by a fixed 340 m/s — at 4.54 km it reports Mach 0.73 where the standard atmosphere gives 0.729. It has a real fault of its own: its burn never cuts off either, and the camera judders late in the ascent. But it flies a mission rather than a minute.

Round 3 to Fable 5.1, 4–2.

The clock

Build	Astra 6	Fable 5.1
Interchange	16m 15s	25m 55s
Fairground	26m 33s	35m 16s
Rocket launch	28m 00s	49m 07s

Astra was quicker on every round, and not narrowly — it finished the launch in 57% of the time Fable 5.1 took, and the interchange in 63%, while producing more than twice the traffic on it.

The verdict: 10–9

	Interchange	Fairground	Launch	Total
Fable 5.1	2	4	4	10
Astra 6	4	3	2	9

The pattern held on all three rounds without a single exception. Astra was faster every time. It was richer on surface detail every time — headlight cones, painted road arrows, jewelled saddle blankets, a zoom control. And all three of its faults were mechanisms rather than decoration: a shoulder line that never separates the ramp, a carousel that does not turn, a launch with no second act.

Set that against how Fable 5.1 lost the round it lost. It was beaten on the interchange for being slower and less spectacular, not for anything being wrong — its interchange remains the only one anyone has built for this series without a visible fault.

That is the whole comparison in one line. Fable 5.1 lost on spectacle. Astra lost on things not working.

It is worth resisting the obvious reading of that. Astra is plainly a very strong model: it cleared both of the interchange's hidden traps, its ascent physics are genuinely modelled, and it produced the most detailed single frame anyone has submitted in three months of running this. Its failures are not failures of capability. They are failures of finishing — a ride that was drawn but never wired to turn, a flight that was staged but never staged. Every one of them is the kind of thing a model catches in about four seconds by opening what it just built, which is exactly what neither model was allowed to do.

That constraint is the point of this test, and it keeps producing the same lesson from a new angle. In June it was that whoever checks their own work wins. In August it was that the frontier model treats a correct answer as the middle of the job. Last week, across five rounds and four models, the season came down to a single point and a variable declared fourteen lines too late. Now, running the same briefs across two labs for the first time, the fastest model in the test lost by shipping three things that look finished and are not.

Every file above is the untouched output of a single run. Scroll back up, watch a carousel that never turns, and see whether you score it the way we did.