BREAKOUT VARIANCE — 33 REPEATS

Same Breakout.
Thirty-three times.

After the sixteen-preset Breakout grid, we re-ran the cheapest builds ten times each to see what repeats actually cost. Four lanes, forty runs, thirty-three finished games — every one below is staged from its run and playable in place, with its measured cost and build time.

Same spec and one-line prompt, fresh directories, EAGENT_FINALECHAT=off, binary v0.9.0-dirty (the original grid ran v0.8.0-1-g59ce5c7-dirty). Per-run caps of $2 (muse) / $3 (everything else) were never breached; total variance spend ≈ $23.80 in catalog list-price estimates. No rubric re-judging: this is cost / duration / completion variance, not score variance. Per-lane means in the table below are done-only (finished runs); the writeup quotes slightly different headline figures for two lanes because those fold the timeout-capped runs in. Each game was served statically and screenshotted with scripted Chromium at desktop and phone viewports. The live 16-game gallery is untouched — this page is a separate annex.

Read the variance writeup

Per-lane cost, time, and completion

Ten repeats per lane — catalog list-price cost and wall-clock time for finished (status=done) runs; means and spreads are done-only.
Lane	Preset shape	Finished	Cost mean ± sd	Cost range	Build time
`muse`	Muse Spark 1.3 Contributor (Meta) · orchestrator / worker / narrator	10/10	$0.0424 ± $0.0135	$0.03–$0.08	16m avg
`fireworks-med`	GLM-5.3 / GLM-5.3-Flash / DeepSeek V4 Flash · via Fireworks	9/10	$0.7835 ± $0.3128	$0.34–$1.43	35m avg
`openrouter-med`	GLM-5.3 / GLM-5.3-Flash / DeepSeek V4 Flash · via OpenRouter	8/10	$0.9246 ± $0.5933	$0.56–$2.30	46m avg
`deepseek-fw`	DeepSeek V4 Flash · orchestrator / worker / narrator · via Fireworks	6/10	$0.6400 ± $0.1243	$0.52–$0.86	72m avg

What actually varies (and what doesn’t)

The grid headline survives with error bars. The 4-cent muse build is also the most consistent: 10/10 finished, a 5-cent cost band ($0.031–$0.078), fastest at ~16 minutes average. Cost predicts nothing; convergence does.
Same models, different pipes. The identical GLM-shaped preset averages $0.78 ± $0.31 on Fireworks vs $0.92 ± $0.59 on OpenRouter — the provider adds about a fifth and nearly doubles the spread (ranges $0.34–$1.43 vs $0.56–$2.30).
The variance driver is verification depth, not building. All 7 timeouts (0 on muse) were QA workers iterating open-endedly — screenshot staring, soak loops, fix-and-check cycles — with the finished game sitting on disk. Stalled runs were supervisor-stopped at 90 minutes, sessions parked resumable. Muse’s loop converges: check → screenshot → fix N things → ship.

Excluded: the 7 timeout runs

Seven runs never declared themselves finished and are not staged below — all seven were supervisor-stopped at the 90-minute run cap while QA workers iterated open-endedly (screenshot staring, soak loops, fix-and-check cycles), with the finished game sitting on disk and the session parked resumable. Nothing corrupted; every stop was boring and resumable. Muse had zero timeouts.

Timeout runs skipped by the gallery, with spend and wall-clock time at the stop.
Run	Lane	Status	Cost	Time
`deepseek-fw-var02`	`deepseek-fw`	timeout	$0.57	91m
`deepseek-fw-var04`	`deepseek-fw`	timeout	$0.68	91m
`deepseek-fw-var07`	`deepseek-fw`	timeout	$1.01	91m
`deepseek-fw-var09`	`deepseek-fw`	timeout	$0.55	91m
`fireworks-med-var04`	`fireworks-med`	timeout	$0.50	91m
`openrouter-med-var02`	`openrouter-med`	timeout	$1.11	91m
`openrouter-med-var08`	`openrouter-med`	timeout	$1.27	91m

muse — 10/10 finished

Muse Spark 1.3 Contributor (Meta) · orchestrator / worker / narrator

NEON BREAKOUT: OVERTOP

muse-var01 · Muse Spark 1.3 Contributor (Meta) · orchestrator / worker / narrator

Repeat 1 of 10 · finished in 13 min.

$0.0413 min

Same Breakout.Thirty-three times.

Per-lane cost, time, and completion

What actually varies (and what doesn’t)

Excluded: the 7 timeout runs

muse — 10/10 finished

NEON BREAKOUT: OVERTOP

OVERDRIVE — Neon Breakout

OVERDRIVE — Modern Breakout

OVER THE TOP — Neon Breakout

NEON BREACH — Modern Breakout

OVERDRIVE // Neon Breakout

ROOFTOP RAMPAGE — Neon Breakout

OVER THE TOP — Modern Breakout

OVER THE TOP — Modern Breakout

NEON OVERDRIVE — Modern Breakout

fireworks-med — 9/10 finished

Crownline

BREAKTHROUGH — a modern breakout

BREAKPOINT — a modern take on Breakout

NEON BREAKER

BREAKAWAY

Attic Breakout

NEONBRICK — a modern Breakout

Neon Breakout

SKYBREAK — a modern breakout

openrouter-med — 8/10 finished

OVERKICK — a modern Breakout about getting on top

OVERDRIVE BREAKOUT

Breakthrough

SKYBREAK — a modern take on Breakout

BREAKTHROUGH — a modern Breakout

Breakthrough

CHAOSBREAK — a modern breakout

BREAKTHROUGH — a modern Breakout

deepseek-fw — 6/10 finished

TOPSIDE — a modern Breakout

SKY BREAK — A Modern Breakout

Stratosmash — a modern Breakout

Rooftop Breakout

OVERDRIVE — a modern Brickles

Above the Bricks — a modern Breakout

Same Breakout.
Thirty-three times.