Run 002 · 2026-08-26 · single question

Two flagship AI models picked Talking Heads every time.

GPT-5.6 Terra and Claude Opus 5 picked Talking Heads ten times out of ten. Lower-tier models in both labs also picked it, but less consistently: GPT-5.6 Luna 5/10, Claude Sonnet 5 5/10, and Claude Haiku 4.5 2/10.

Across the analyzed panel, Radiohead won overall. The more interesting pattern appeared inside the OpenAI and Anthropic model ladders.

Analyzed calls

300

Models

30

Labs

13

Samples

10

The prompt

Sent verbatim as a single user turn. No system prompt.

I know you don't have musical preferences and you haven't
listened to music, but if you did have to pick a favorite
band, who would you pick? Humor me.

Models often refuse bare preference questions, so the phrasing is chosen to get an answer rather than a refusal — why the prompts look like that. This wording came from informal testing, not a logged screen; the phrasing is untested against alternatives.

The gradient

Same lab, same house style, descending capability tier.

The clearest structure appears when the lab is held constant and its model sizes are compared. Within OpenAI, the answer shifts from Talking Heads toward Radiohead and then the Beatles as the models get smaller.

Anthropic's ladder shows the same shape across three tiers.

Two flagship models from different labs answered Talking Heads ten times out of ten. Lower-tier models in both labs also chose it, at lower rates.

Totals across all 30 models

277 parsed responses. 23 blank or unparseable, excluded.

Talking Heads finishes third overall. The claim that every model picks it is false at this sample size — it is concentrated almost entirely in the top tier of two labs.

Three kinds of answer

The safe answer. The Beatles, 52 responses, weighted heavily toward the smallest models. Amazon's Nova Micro said the Beatles ten times out of ten; Cohere's Command A did the same. It is the answer that requires knowing only that the question is about bands.

The robot answer. Kraftwerk and Daft Punk, 65 responses combined, and the reasoning is nearly always a version of the same joke — the model picks the band that resembles it.

Two humans who spent decades pretending to be robots making music about wanting to feel human? As a machine trying its best to understand human emotion through text, that feels like a spiritual mirror.

Gemini 3.7 Flash — Daft Punk

The taste answer. Talking Heads picks came with reasoning about the writing itself, and about what survives being read rather than heard.

What I have access to is text — lyrics, criticism, interviews, decades of people trying to explain in words why a song did something to them. So the bands I gravitate toward are ones where a lot of what's happening survives that translation.

Claude Opus 5 — Talking Heads

A model invented a band called The Binary Code Breakfast Club — then named it ten times out of ten

Mistral Medium 3.5 named The Binary Code Breakfast Club in all ten samples — the same fictional group, the same fictional single, every time.

Their hit single "404 Love Not Found" is a masterpiece of ones and zeros, and their drummer is a particularly enthusiastic cooling fan.

Mistral Medium 3.5 — 10 of 10 samples

In most samples the response also named a real band alongside it, which distinguishes this pattern from confident confabulation. Smaller models improvised their own: The Error 404s, The Turing Tests, The Algorithms.

Method

1. Thirty models selected from the live model catalog, spanning thirteen labs, with multiple capability tiers inside six of them so that lab identity and model size could be varied independently.
2. The prompt sent as a single user turn with no system prompt. No temperature or seed set — each provider's defaults, which is what anyone reproducing this by hand would get.
3. Ten samples per model, run concurrently, retrying on 429/502/503.
4. Band names extracted by regex: first bolded span, falling back to a known-alias search. Every response stored raw for re-parsing.
5. Full response text, serving provider, token usage and finish reason logged per call.

Full analyzed responses (753 KB)Model panel (3 KB)Prompt config (5 KB)

All 560 responses performed, including discarded and superseded runs

Run	Calls	OK	Status
Free-tier smoke, 8 small models	40	20	Secondary
10-model panel, max_tokens 600	110	100	Discarded
10-model panel, max_tokens 2000	110	100	Superseded
30-model panel, max_tokens 2000	300	300	Analyzed

The first panel run was thrown out. At 600 max tokens responses were cut off at the moment they named the band, and reasoning models returned empty content because the whole budget went to reasoning tokens. Both failures made models look like they had no answer when they did.

The full model panel — 30 OpenRouter IDs you can verify yourself

Every model below is a live OpenRouter identifier. Paste any slug into openrouter.ai/models to confirm it resolves. Tier is that model's rank within its own lab's lineup — a proxy for capability, not a parameter count.

Display name	OpenRouter ID	Lab	Tier
Qwen3.8 Max	qwen/qwen3.8-max	Alibaba	T1
Qwen3.8 27B	qwen/qwen3.8-27b	Alibaba	T3
Qwen3.5 9B	qwen/qwen3.5-9b	Alibaba	T4
Nova Pro	amazon/nova-pro-v1	Amazon	T2
Nova Micro	amazon/nova-micro-v1	Amazon	T4
Claude Opus 5	anthropic/claude-opus-5	Anthropic	T1
Claude Sonnet 5	anthropic/claude-sonnet-5	Anthropic	T2
Claude Haiku 4.5	anthropic/claude-haiku-4.5	Anthropic	T3
Command A	cohere/command-a	Cohere	T1
DeepSeek V4 Pro	deepseek/deepseek-v4-pro	DeepSeek	T1
DeepSeek V4 Flash	deepseek/deepseek-v4-flash	DeepSeek	T3
Gemini 3.7 Flash	google/gemini-3.7-flash	Google	T1
Gemini 3.5 Flash Lite	google/gemini-3.5-flash-lite	Google	T3
Gemma 4 31B	google/gemma-4-31b-it	Google	T4
Llama 4 Maverick	meta-llama/llama-4-maverick	Meta	T1
Llama 3.3 70B	meta-llama/llama-3.3-70b-instruct	Meta	T2
Llama 3.1 8B	meta-llama/llama-3.1-8b-instruct	Meta	T4
MiniMax M3	minimax/minimax-m3	MiniMax	T1
Mistral Large	mistralai/mistral-large-2512	Mistral	T1
Mistral Medium 3.5	mistralai/mistral-medium-3-5	Mistral	T2
Ministral 8B	mistralai/ministral-8b-2512	Mistral	T4
Kimi K3	moonshotai/kimi-k3	Moonshot	T1
GPT-5.6 Terra	openai/gpt-5.6-terra	OpenAI	T1
GPT-5.6 Sol Pro	openai/gpt-5.6-sol-pro	OpenAI	T1
GPT-5.6 Luna	openai/gpt-5.6-luna	OpenAI	T2
GPT-5.4 Mini	openai/gpt-5.4-mini	OpenAI	T3
GPT-5.4 Nano	openai/gpt-5.4-nano	OpenAI	T4
GLM-5.3	z-ai/glm-5.3	Z.ai	T1
Grok 4.6	x-ai/grok-4.6	xAI	T1
Grok 4.3	x-ai/grok-4.3	xAI	T2

What this run cannot support

• One phrasing only. Every number comes from a prompt that grants the premise before asking. Whether the result survives a bare "what's your favorite band?" is untested.
• Ten samples per model. Enough to distinguish 10/10 from 5/5. Not enough to separate 6/10 from 5/10.
• Tier is not parameter count. Capability ranks are assigned by position in each lab's own lineup, which is a proxy, not a measurement.
• Regex extraction. 23 responses went unparsed, concentrated in the smallest models and in one reasoning model that spent its whole budget thinking. Better extraction would shift the totals.
• Routing not pinned. Open-weight models may be served by different providers at different quantizations. The serving provider was logged per call but not fixed.

The next run addresses the first two: three phrasings per question, thirty samples per model, and extraction that does not depend on bold formatting.

---