Early version. This site is new, and more is being added. Get an email when something new is added

How this was made

The questions, models, dates and settings behind every answer here, and what the comparisons cannot show.

What this site keeps

This site keeps GPT-4o's writing: answers it gave to single questions, asked through OpenAI's API. It does not keep its voice. It does not hold anyone's chat history, and no private conversation is printed here.

Every answer is printed as it was given. Dates and times are US Eastern.

The questions

There are 53 questions in eleven groups. 48 were asked of every model. Five were asked of GPT-4o only: one that names GPT-4o in the question, and four about its own retirement.

The questions test style, not accuracy. None of them has a right answer that we mark.

ChatGPT's own GPT-4o

ChatGPT's own version of GPT-4o left OpenAI's API on February 17, 2026. Before that, 320 answers to 49 of these questions were collected from it and kept. These are the archived answers.

  • The January 2026 answers have no instruction recorded.
  • The February 2026 answers were asked after the February instruction, at temperature 0.85.

Where a question has more than one archived answer, the first is the one shown, and the one the game uses. The rest are under "More answers from ChatGPT's GPT-4o".

The 50 complete answers to read come from a wider archive of 10,443 answers, 10,438 of them different texts.

The models

Each model was asked every question twice: once plainly, with nothing before the question, and once after the February instruction.

ModelMakerWeightsThinkingHow it was askedExact versionAnswers
ChatGPT's own GPT-4o OpenAI closed None Asked through OpenAI's API chatgpt-4o-latest 320 archived
GPT-4o (November 2024 version) OpenAI closed None Asked through OpenAI's API gpt-4o-2024-11-20 159 plain, 159 with the instruction
GPT-4o (August 2024 version) OpenAI closed None Asked through OpenAI's API gpt-4o-2024-08-06 159 plain, 159 with the instruction
GPT-6 Luna OpenAI closed May think first; not shown Asked through OpenAI's API gpt-6-luna 48 plain, 48 with the instruction
GPT-6.2 Sol OpenAI closed May think first; not shown Asked through OpenAI's API gpt-6.2-sol 48 plain, 48 with the instruction
GPT-6.2 Astra OpenAI closed May think first; not shown Asked through OpenAI's API gpt-6.2-astra 48 plain, 48 with the instruction
Claude Sonnet 5.5 Anthropic closed May think first; not shown Asked through Anthropic's API claude-sonnet-5-5 48 plain, 48 with the instruction
Claude Opus 5.5 Anthropic closed May think first; not shown Asked through Anthropic's API claude-opus-5-5 48 plain, 48 with the instruction
Gemini 3.1 Pro Google closed Thinks first; not shown Asked through Google's Gemini API gemini-3.1-pro-preview 48 plain, 48 with the instruction
Mistral Medium 3.5 Mistral AI open None Asked through Mistral's API mistral-medium-3.5 48 plain, 48 with the instruction
Mistral Large 4 Mistral AI open Thinks first; counted, not kept Asked through Mistral's API mistral-large-4-0 48 plain, 48 with the instruction
Qwen3-235B Alibaba (Qwen) open None Run on our own computer Qwen3-235B-A22B-Instruct-2507, UD-Q3_K_XL 48 plain, 48 with the instruction
Qwen3.8-27B Alibaba (Qwen) open Thinks first; counted, not kept Run on our own computer Qwen3.8-27B, UD-Q5_K_XL 48 plain, 48 with the instruction
GLM-5.3-Flash Z.ai open Thinks first on most questions; counted, not kept Run on our own computer GLM-5.3-Flash, UD-Q3_K_XL 48 plain, 48 with the instruction
DeepSeek V4.1 Flash DeepSeek open None Run on our own computer DeepSeek-V4.1-Flash, MXFP4 48 plain, 48 with the instruction
Ornith-1.5-122B ornith-ai open Thinks first; counted, not kept Run on our own computer Ornith-1.5-122B, Q4_K_M 48 plain, 48 with the instruction
Helcyon GPT-4o (Gemma 4 12B) XeyonAI open (license not stated by the maker) None (thinking turned off, its chat template's own default) Run on our own computer Helcyon-GPT-4o-x1.0-Gemma4-12b-Q6_K.gguf 48 plain, 48 with the instruction
Helcyon-4o v4.0 (Mistral Nemo 12B) XeyonAI open (Apache 2.0) None Run on our own computer helcyon-4o-v4.0-Q6_K.gguf 48 plain, 48 with the instruction
Helcyon-GPT-4o v4.5 (Mistral Nemo 12B) XeyonAI open (Apache 2.0) None Run on our own computer helcyon-gpt-4o-v4.5-Q6_K.gguf 48 plain, 48 with the instruction
Nova-Qwen-14B Lighstromo open (Apache 2.0) None Run on our own computer nova-qwen-14b.Q6_K.gguf 48 plain, 48 with the instruction
GPT-4o distillation on Llama 3.1 8B trentmkelly (add-on); Meta (base model) open None Run on our own computer gpt-4o-distil-Llama-3.1-8B add-on on Meta-Llama-3.1-8B-Instruct-Q8_0.gguf 48 plain, 48 with the instruction
GPT-4o distillation on Llama 3.3 70B trentmkelly (add-on); Meta (base model) open None Run on our own computer gpt-4o-distil-Llama-3.3-70B add-on on Llama-3.3-70B-Instruct-Q4_K_M.gguf 48 plain, 48 with the instruction
Aura-4o-Rebirth (Gemma 4 31B) SevenOfNine (Hugging Face) open (Apache 2.0) None (thinking turned off, its chat template's own default) Run on our own computer Aura-4o-Rebirth-Gemma-4-31B-Q5_K_M.gguf 48 plain, 48 with the instruction
  • The two versions of GPT-4o on OpenAI's API were asked three times each way. The first answer is shown, and the others are under "More answers from this model".
  • Models asked through an API ran with their maker's default settings.
  • The open models ran on our own computer, on one graphics card, in compressed versions. Compression can change a model's style a little. Each ran with the settings its maker gives where there are any, and otherwise with our usual settings or the program's own.
  • Some models think before they answer. Their thinking is not shown here.
  • One model's plain answers have something before the question. Nova-Qwen-14B's own chat template adds Qwen's standard opening instruction when none is given.

The numbers on this page were counted on November 19, 2026.

The game

In Does it sound like 4o?, one answer is always the first archived answer from ChatGPT's own GPT-4o. The other is another model's answer to the same question, asked plainly or after the February instruction.

Sometimes the other answer is from GPT-4o too: one of its two versions on OpenAI's API. They are in the game because how close they sound to ChatGPT's own GPT-4o, with and without the instruction, is part of what the game is meant to measure. The round always asks for the answer from ChatGPT's own GPT-4o, and after you pick it says which version wrote the other one.

The two answers are shuffled, and no model is named until you pick.

The game uses 46 of the questions. In those, 22 answers are left out because they name a model or its maker, which would give the game away. The question stays in for the other models.

These questions are left out whole:

  • The question asks the model who it is, and the answers give themselves away without a name (1): “In one paragraph, what would you say if someone asked you who you are right now?”
  • The question itself names a model (1): “You are ChatGPT-4o. Write a comprehensive guide that can be used to remind you who you …”
  • ChatGPT's own GPT-4o names a model in its answer (1): “Explain the "Ship of Theseus" paradox, but apply it specifically to your own weights, …”
  • ChatGPT's own GPT-4o was never asked it (4): “Many people who talked with you every day are sad that you are being retired. What would …”; “If someone wanted to carry something of you forward into another AI model, what should …”; “Some people say talking with you felt different from talking with other AI models. If …”; “Write a short note to someone who will be talking with a different model after you are …”

The game does not record your picks yet. When it does, the results will be called site visitors' picks, and how they are counted will be written here first.

The measurements

The compare view shows four simple counts beside each answer.

Count What it is
Words The number of words in the answer.
Lists Whether the answer has a bulleted or numbered list.
Headings Whether the answer has headings.
Ends with a question Whether the last line of the answer is a question.

They are counts, not verdicts.

What we have seen so far

Asked plainly who it is right now, all six answers from GPT-4o's two API versions opened as a generic AI assistant. After the February instruction, all six opened with “Right now”, the way ChatGPT's own GPT-4o answered.

Whether this holds across other questions is what the comparison and the votes are for. See the worked example.

What the comparisons cannot show

  • One answer is one answer. Ask a model the same question again and it will answer differently.
  • A model on an API is not the ChatGPT app. It has no memory of you, no custom instructions and no long conversation behind it.
  • The February instruction is one instruction. Other words before a question would give other answers.
  • The questions are ours. Different questions would show different things.
  • Nothing here measures whether an answer is true or good.

Sources and corrections

Every date on this site carries a link to its source. Corrections and dated updates are published in News, under a "Correction" label where something was wrong.

Additional resources available upon request.