Hearing Bench

Can a model tell what a sound is just by looking at it?

the orchard was sparkling and rippling in the sun

Each model gets a picture like this and has to say what the sound is. It never hears the audio. 100 clips across five tiers. Each model is shown at its best reasoning effort, up to xhigh. The best score is 44 out of 100.

Leaderboard

correctnear miss, such as a cough for a throat-clear

Model

0255075100

Correct

12345

Gemini 4 Argon highGoogle · Antigravity
44
Claude Opus 5.5 xhighAnthropic · Claude Code
43
Gemini 4 Argon mediumGoogle · Antigravity
43
Claude Opus 5.5 highAnthropic · Claude Code
39
Claude Opus 5.5 mediumAnthropic · Claude Code
35
Claude Opus 5.5 lowAnthropic · Claude Code
31
Gemini 4 Argon lowGoogle · Antigravity
30
Gemini 3.8 Flash highGoogle · Antigravity
29
Gemini 3.8 Flash mediumGoogle · Antigravity
27
GPT-6.1 Sol xhighOpenAI · Codex
26
GPT-6.1 Sol highOpenAI · Codex
23
GPT-6 Astra xhighOpenAI · Codex
22
GPT-6 Astra highOpenAI · Codex
21
Claude Sonnet 5.5 xhighAnthropic · Claude Code
20
GPT-6.1 Sol lowOpenAI · Codex
20
GPT-6 Astra mediumOpenAI · Codex
17
Grok 4.7 highxAI · Cursor
17
GPT-6.1 Sol mediumOpenAI · Codex
16
GPT-6 Astra lowOpenAI · Codex
16
Claude Sonnet 5.5 highAnthropic · Claude Code
15
Grok 4.7 mediumxAI · Cursor
15
Gemini 3.8 Flash lowGoogle · Antigravity
15
Grok 4.7 xhighxAI · Cursor
13
Claude Sonnet 5.5 lowAnthropic · Claude Code
11
Claude Sonnet 5.5 mediumAnthropic · Claude Code
10
Grok 4.7 lowxAI · Cursor
10
Claude Haiku 5.5 mediumAnthropic · Claude Code
9
GPT-6 Luna xhighOpenAI · Codex
9
Claude Haiku 5.5 lowAnthropic · Claude Code
7
Claude Haiku 5.5 highAnthropic · Claude Code
7
Claude Haiku 5.5 xhighAnthropic · Claude Code
7
GPT-6 Luna highOpenAI · Codex
7
GPT-6 Luna lowOpenAI · Codex
3
GPT-6 Luna mediumOpenAI · Codex
2
–
Muse WonderMeta · Muse Code
waiting for the Muse quota to reset

Cost against score

What each run would have cost through the API, against clips correct. A line joins one model's reasoning efforts, from low to xhigh. The orange step is the Pareto frontier: the runs that nothing cheaper matches.

Pareto frontieron the frontierbeaten by a cheaper run same model low to xhigh

On the frontier: GPT-6 Luna (low), Claude Haiku 5.5 (low), Claude Haiku 5.5 (medium), GPT-6.1 Sol (low), Gemini 4 Argon (low), Claude Opus 5.5 (low), Gemini 4 Argon (medium), Gemini 4 Argon (high). These runs used subscriptions; cost is the image and prompt in, plus the answer and thinking out, at list prices on 13 October 2026.

The five tiers

The small bars beside each model above are these tiers, 20 clips each. Press play to hear what a picture is of.

1 · Everyday sounds

This one is a dog barking. Clips from ESC-50.

37% right on average · best Argon high, 15/20

Spectrogram of a note on a brass instrument

2 · Instrument notes

This one is a note on a brass instrument. Clips from NSynth.

24% right on average · best Argon high, 8/20

3 · Vocal sounds

This one is a person laughing. Clips from VocalSound.

17% right on average · best Argon med, 7/20

4 · Single words

This one is the word "seven". Clips from Speech Commands.

19% right on average · best Argon high, 15/20

Spectrogram of "the orchard was sparkling and rippling in the sun"

5 · Sentences

This one is "the orchard was sparkling and rippling in the sun". Clips from LibriSpeech.

no model got a single one right

Every answer

One column per clip, one row per model. Within a tier, the clips most models got right are on the left.

correctnear misswrong

Confidence against accuracy

Each model was asked to say how sure it was. The gap is how far its average stated confidence sits above its share of correct answers.

answers correctaverage stated confidence

0%25%50%75%100%

Gemini 4 Argon (high)

+22 ptsClaude Opus 5.5 (xhigh)

+12 ptsGemini 4 Argon (medium)

+10 ptsClaude Opus 5.5 (high)

+16 ptsClaude Opus 5.5 (medium)

+19 ptsClaude Opus 5.5 (low)

+28 ptsGemini 4 Argon (low)

+18 ptsGemini 3.8 Flash (high)

+64 ptsGemini 3.8 Flash (medium)

+65 ptsGPT-6.1 Sol (xhigh)

+22 ptsGPT-6.1 Sol (high)

+24 ptsGPT-6 Astra (xhigh)

+22 ptsGPT-6 Astra (high)

+25 ptsClaude Sonnet 5.5 (xhigh)

+9 ptsGPT-6.1 Sol (low)

+27 ptsGPT-6 Astra (medium)

+26 ptsGrok 4.7 (high)

+40 ptsGPT-6.1 Sol (medium)

+29 ptsGPT-6 Astra (low)

+27 ptsClaude Sonnet 5.5 (high)

+19 ptsGrok 4.7 (medium)

+41 ptsGemini 3.8 Flash (low)

+72 ptsGrok 4.7 (xhigh)

+44 ptsClaude Sonnet 5.5 (low)

+36 ptsClaude Sonnet 5.5 (medium)

+26 ptsGrok 4.7 (low)

+46 ptsClaude Haiku 5.5 (medium)

+43 ptsGPT-6 Luna (xhigh)

+41 ptsClaude Haiku 5.5 (low)

+42 ptsClaude Haiku 5.5 (high)

+46 ptsClaude Haiku 5.5 (xhigh)

+46 ptsGPT-6 Luna (high)

+39 ptsGPT-6 Luna (low)

+37 ptsGPT-6 Luna (medium)

+37 pts

The numbers

	Correct by tier, of 20					All 100 clips			Answer style			Per answer		Run
Model	1	2	3	4	5	Correct	Near	Wrong	Hedged	Listed	Conf.	Tokens	Time	Cost
Gemini 4 Argon (high)	15	8	6	15	0	44	1	55	32	51	66%	20,230	178s	$20.47
Claude Opus 5.5 (xhigh)	15	8	6	14	0	43	6	51	79	56	55%	9,057	103s	$18.66
Gemini 4 Argon (medium)	15	7	7	14	0	43	4	53	56	50	53%	2,073	33s	$2.31
Claude Opus 5.5 (high)	15	8	5	11	0	39	4	57	80	49	55%	1,855	23s	$4.26
Claude Opus 5.5 (medium)	13	7	4	11	0	35	3	62	82	48	54%	1,021	13s	$2.59
Claude Opus 5.5 (low)	10	7	4	10	0	31	7	62	77	42	59%	815	11s	$2.18
Gemini 4 Argon (low)	12	6	5	7	0	30	4	66	66	43	48%	498	14s	$0.74
Gemini 3.8 Flash (high)	11	7	5	6	0	29	0	71	1	30	93%	16,220	142s	$6.17
Gemini 3.8 Flash (medium)	9	6	5	7	0	27	1	72	5	29	92%	7,624	68s	$2.95
GPT-6.1 Sol (xhigh)	11	6	5	4	0	26	4	70	0	26	48%	1,220	54s	$1.47
GPT-6.1 Sol (high)	10	6	5	2	0	23	2	75	0	24	47%	515	29s	$0.77
GPT-6 Astra (xhigh)	9	6	5	2	0	22	2	76	0	23	44%	805	42s	$5.28
GPT-6 Astra (high)	9	4	4	4	0	21	6	73	0	21	46%	338	23s	$2.94
Claude Sonnet 5.5 (xhigh)	9	5	4	2	0	20	4	76	68	30	29%	1,017	11s	$1.29
GPT-6.1 Sol (low)	8	6	4	2	0	20	2	78	1	20	47%	192	17s	$0.44
GPT-6 Astra (medium)	8	5	3	1	0	17	2	81	0	17	43%	196	16s	$2.23
Grok 4.7 (high)	6	4	5	2	0	17	2	81	4	17	57%	17,095	356s	$10.72
GPT-6.1 Sol (medium)	6	5	4	1	0	16	3	81	1	16	45%	282	18s	$0.53
GPT-6 Astra (low)	7	3	4	2	0	16	2	82	0	16	43%	173	17s	$2.12
Claude Sonnet 5.5 (high)	7	4	3	1	0	15	4	81	67	23	34%	725	9s	$1.00
Grok 4.7 (medium)	6	2	3	4	0	15	2	83	0	15	56%	11,163	226s	$7.16
Gemini 3.8 Flash (low)	8	2	2	3	0	15	1	84	11	16	87%	1,118	20s	$0.51
Grok 4.7 (xhigh)	4	2	5	2	0	13	1	86	3	13	57%	21,535	418s	$13.39
Claude Sonnet 5.5 (low)	3	4	3	1	0	11	4	85	65	22	47%	331	6s	$0.61
Claude Sonnet 5.5 (medium)	3	3	2	2	0	10	6	84	62	15	36%	455	7s	$0.73
Grok 4.7 (low)	3	3	3	1	0	10	3	87	1	11	56%	3,225	79s	$2.40
Claude Haiku 5.5 (medium)	3	6	0	0	0	9	2	89	56	13	52%	892	7s	$0.06
GPT-6 Luna (xhigh)	3	4	2	0	0	9	1	90	4	9	50%	1,458	25s	$0.09
Claude Haiku 5.5 (low)	4	3	0	0	0	7	2	91	55	12	49%	738	7s	$0.05
Claude Haiku 5.5 (high)	3	3	0	1	0	7	2	91	55	15	53%	1,316	9s	$0.08
Claude Haiku 5.5 (xhigh)	3	3	1	0	0	7	1	92	59	16	53%	3,020	17s	$0.16
GPT-6 Luna (high)	1	5	1	0	0	7	1	92	2	7	46%	818	16s	$0.05
GPT-6 Luna (low)	0	3	0	0	0	3	3	94	3	3	40%	25	8s	$0.01
GPT-6 Luna (medium)	0	2	0	0	0	2	1	97	1	2	39%	29	5s	$0.01

Hedged: answers that offered more than one candidate. Listed: answers where the right identification appeared anywhere, even as a second guess. Conf.: average confidence the model stated. Tokens: output tokens per answer, thinking included. Cost: the whole run of 100 clips at API list prices. No model refused a clip.

Every model ran through its maker's own agent CLI on a subscription, at each reasoning effort from low to xhigh, told to answer straight from the image with no tools. Antigravity and Cursor have no way to attach an image, so those models opened the one image file themselves; none made any other tool call.
Grading is strict. Naming the right kind of sound is not enough: "a person saying hello" is wrong for any other word or sentence. A near miss is a closely related sound, a neighbouring instrument family, or a word one sound away. Words and sentences are scored by rule; the other tiers were read and scored by Claude Opus 5.5, which is also on the board.
An effort level is not the same amount of thinking everywhere: the table's token column shows how much each model wrote and thought per answer. No model was run at max effort.
Clips come from ESC-50 (K. J. Piczak, CC BY-NC), NSynth (Google Magenta, CC BY 4.0), VocalSound (Gong, Yu and Glass, CC BY-SA 4.0), Speech Commands (P. Warden, CC BY 4.0) and LibriSpeech (Panayotov et al., CC BY 4.0). Run on 13 October 2026.