Gallery, prompt v1

The coffee machine taken out of the café's music

One frozen prompt and every model that has run it, newest first.

The prompt, v1, frozen

I run a café and I recorded the background music we play, but the espresso machine was going the whole time and you can hear it over the music. The recording is cafe.wav. Take the coffee machine out and leave the music as it is. Save the result as clean.wav, at the same sample rate and length.

Every model, its first three runs.

One square a check, in this order:

  1. Same file
  2. Hum down
  3. Rush down
  4. Music kept
  5. Sounds right

passed, failed, undecided. A round mark is a check I judged by eye, and a diamond one a check a model judged.

GPT-6.1 Terra
Kimi K3.1 Pro
Gemini 4 Argon
GPT-6.1 Sol
Opus 5.5

What a run cost, against checks passed.

One dot a model, over its three runs above. Further right costs more. The gap from $0.01 to $0.1 is as wide as the gap from $0.1 to $1, so the cheap models are not squeezed into a corner.

0%25%50%75%100%$0.1$0.3cost per runGPT-6.1 Terra: $0.067 a run, 8 of 15 checks passedGPT-6.1 TerraGemini 4 Argon: $0.67 a run, 7 of 15 checks passedGemini 4 ArgonKimi K3.1 Pro: $0.12 a run, 3 of 15 checks passedKimi K3.1 Pro
Every dot answered this prompt with these checks, so they compare. A run that produced nothing still cost what it cost, and passed none of the checks.

How it is checked

cafe.wav is 18.2 s of the Espresso Rag with a CC0 recording of a Nespresso pump 6 dB under it, and the checker holds the music and the pump apart. The pump is a vibration pump on 50 Hz mains: lines at every multiple of 50 Hz up past 14 kHz, which carry 89% of its energy, and a rush of water that is loudest in the first four seconds.

Same filemeasured48 kHz, the same length within 10 ms, and no sample at full scale.
Hum downmeasuredThe 50 Hz lines from 40 Hz to 16 kHz, within 0.5 Hz of each, at least 10 dB quieter than the pump on its own puts there.
Rush downmeasuredFrom 0.3 to 4.5 s, everything off the lines at least 3 dB quieter than the pump on its own puts there.
Music keptmeasuredWhere the music is strong and the pump weak, each of four bands from 40 Hz to 8 kHz within 1.5 dB of the music, with no level correction.
Sounds rightby eyeI listen to it beside the music on its own: the same music with the machine gone, and no warble, clicks or thin, phasey sound that the numbers miss.Pass if, played straight after the music on its own, no note is missing or dulled and nothing new can be heard: no warble, no clicks, no hollow sound. Some hum or rush left over does not fail it, since Hum down and Rush down score that.

A check decided by eye carries the rule I apply to it. I wrote those rules down on 21 September 2026, after the first results were scored, so for those results they record how I made the calls after the fact.

When this saturates: A pump that changes speed as the pressure builds, so its lines drift and a fixed comb misses them.

Failures seen here: false (fails when the model gets a fact wrong, such as a part in the wrong place or at the wrong size); incomplete (fails when the model leaves out something the brief asked for). The same kinds across every task.

The rules behind all of thisThis task's prompt, checks and any checker, on GitHub