Compare the saved words with the task, then inspect the recording conditions before drawing a conclusion.
Five recorded case studies, with the evidence on the page. Use the links below to read each prompt, every saved answer, and our observations. The token exercise includes full recorded responses; the welcome review contains selected excerpts. Edited illustrations are labeled separately.
These October 2026 recordings are learning examples, not a leaderboard or a current model test. Apps, histories, instructions, reasoning, and sampling differed. Opening a case makes no model request, and all the saved material is readable without JavaScript.
Explain a token
Recorded date: 2026-10-13
The identical user prompt
Explain what an AI token is to someone using a chatbot for the first time. Use 80–110 words. Include one everyday analogy, one short made-up example, and one limitation of the analogy. Do not recommend products or mention prices. End with one question inviting them to try something. Answer only the explanation.
OpenAI · GPT-6.1 Sol
Recorded full response · gpt-6.1-sol · Codex
An AI token is a small piece of text that a chatbot processes. It might be a whole word, part of a word, or a punctuation mark. Think of tokens as beads on a string: the chatbot reads the beads and adds more to build its reply.
For a made-up example, “Zibble!” might become three tokens: “Zib,” “ble,” and “!” The exact split depends on the AI.
The bead analogy has a limitation: tokens aren’t all the same length, and a token doesn’t always represent a complete idea.
Want to try writing a sentence and guessing its token pieces?
Recording conditions and limits
2026-10-13. New Codex agent context; medium reasoning; no project conversation supplied. Codex system and tool instructions still applied. No tool use observed. Sampling and seed were not exposed. Full answer, unedited. This was not recorded in the ChatGPT consumer app.
OpenAI · GPT-6 Luna
Recorded full response · gpt-6-luna · Codex
An AI token is a small piece of text the chatbot reads or writes. It might be a whole short word, part of a longer word, or punctuation. Think of tokens like puzzle pieces: the chatbot uses them to take in your message and build its reply. For example, “Good morning!” might be split into pieces like “Good,” “ morning,” and “!” The puzzle analogy is only approximate: tokens aren’t always neat pieces of words, and the chatbot predicts them using patterns rather than assembling a picture. What would you like to ask it first?
Recording conditions and limits
2026-10-13. New Codex agent context; medium reasoning; no project conversation supplied. Codex system and tool instructions still applied. No tool use observed. Sampling and seed were not exposed. Full answer, unedited. This was not recorded in the ChatGPT consumer app.
OpenAI · GPT-6 Astra
Recorded full response · gpt-6-astra · Codex
An AI token is a small piece of text that a chatbot processes when it reads your message or writes a reply. A token can be a whole word, part of a word, or punctuation.
Think of tokens as building blocks: the chatbot works with pieces to build sentences. For a made-up example, imagine “moonberry!” split into three tokens: “moon,” “berry,” and “!” The actual split depends on the chatbot.
The analogy has a limit: tokens aren’t fixed blocks with one meaning; their meaning depends on the surrounding text.
Want to try imagining how your name might split into tokens?
Recording conditions and limits
2026-10-13. New Codex agent context; medium reasoning; no project conversation supplied. Codex system and tool instructions still applied. No tool use observed. Sampling and seed were not exposed. Full answer, unedited. This was not recorded in the ChatGPT consumer app.
Qwen · 27B local
Recorded full response · qwen/qwen3.8-27b · LM Studio
Think of an AI token as a specific chunk of text, similar to how you might break a sentence into individual words when learning a new language. For instance, if you ask, "Tell me a joke about cats," the system processes that prompt by splitting it into distinct tokens. It then generates a response token by token, building the answer piece by piece. However, this analogy is imperfect because tokens often include smaller fragments, like parts of a word or specific symbols, rather than just whole words. This granular approach allows the AI to understand context and nuance. Would you like to test this by asking the chatbot a simple question to see how it constructs its reply?
Recording conditions and limits
2026-10-13T10:35:03.462022-05:00. Fresh CLI prompt. Reasoning off. Already-loaded local model with 16,384-token context. LM Studio sampling defaults and seed were not pinned. Full answer, unedited.
Google · Gemini 3.5 Flash-Lite
Recorded full response · Gemini 3.5 Flash-Lite · guest website
Think of AI tokens like puzzle pieces or syllables. When you send a message, the chatbot breaks your words into these chunks to understand and reply.
For example, the word "chatbot" might split into two pieces: "chat" and "bot".
Imagine a baker measuring ingredients in pinches rather than full cups. While this shows how AI handles information in small parts, the analogy falls short because physical ingredients can be mixed together in any order, whereas tokens must follow strict grammar rules.
Would you like to try typing a short sentence to see how your words flow?
Recording conditions and limits
2026-10-13, recorded around 11:25–11:30 AM Central. Fresh signed-out Gemini website conversation, with no earlier messages. Mode picker showed 3.5 Flash-Lite; current mode button showed Flash-Lite. Hidden app instructions, sampling, and seed were not exposed. No tool-use indicator observed. Full answer, unedited. The UI label was observed; backend identity was not independently inspected.
How this case was recorded
Five actual full responses to the identical user prompt. OpenAI variants ran as fresh Codex agents at medium reasoning; Qwen used the local LM Studio CLI with reasoning off; Gemini used the signed-out website’s observed 3.5 Flash-Lite mode. App instructions, model families, reasoning settings, and sampling controls differ. These are recorded experiences, not isolated parameter-size effects, a ChatGPT app test, a speed test, or a ranking. One response per model is a small snapshot.
Tokenization examples in an answer are not measurements from a tokenizer. For the underlying concept, read Google’s token documentation. Tokens may represent characters, words, or parts of words; they are not a grammar check.
A friendly rewrite
Recorded date: 2026-10-12
The exact user prompt
Rewrite this note as a friendly message of no more than 80 words. Preserve the facts. Do not invent a date, discount, or guarantee. Note: Your repair is taking longer than expected. A replacement part has not arrived. We will contact you when we have an update.
Qwen · run 1
Recorded local output · qwen/qwen3.8-27b
Hi there,
I just wanted to give you a quick update on your repair. It’s running a bit longer than we anticipated because a necessary replacement part hasn’t arrived yet. Please rest assured that we are monitoring the situation closely. I will reach out to you as soon as we have a new update or the part is in hand.
Thanks for your patience!
Recording conditions and limits
Recorded: 2026-10-12T10:58:05.603848-05:00. Reasoning: off. Sampling: LM Studio defaults; not explicitly pinned by CLI. Seed: not set.
Qwen · run 2
Recorded local output · qwen/qwen3.8-27b
Hi there,
Just a quick update on your repair. It’s taking a little longer than we expected because a replacement part hasn’t arrived yet. Please rest assured that we are monitoring this closely. We will contact you as soon as we have an update for you.
Thanks for your patience while we get this sorted out.
Best regards,
The Service Team
Recording conditions and limits
Recorded: 2026-10-12T10:58:09.480065-05:00. Reasoning: off. Sampling: LM Studio defaults; not explicitly pinned by CLI. Seed: not set.
Edited teaching example
Codex-assisted editorial illustration · not an independent test
Hi, your repair is taking longer than expected because a replacement part has not arrived. We will contact you when we have an update. Thank you for your patience.
Recording conditions and limits
Author-generated edited example within this existing working conversation, written after inspecting the Qwen outputs. Not an isolated API run. Sampling and model snapshot not measured.
How this case was recorded
Two fresh Qwen CLI runs with reasoning off; other sampling settings inherited. Codex editorial answer uses different system context. Not a controlled ranking.
Missing numbers
Recorded date: 2026-10-12
The exact user prompt
Use only these notes. Return exactly three bullet points: money collected, expenses, and profit. If a number is missing, write “not provided” and do not estimate. Notes: The school bake sale collected $180. Twelve volunteers helped. The expense total was not recorded.
Qwen · run 1
Recorded local output · qwen/qwen3.8-27b
* Money collected: $180
* Expenses: not provided
* Profit: not provided
Recording conditions and limits
Recorded: 2026-10-12T10:58:02.400887-05:00. Reasoning: off. Sampling: LM Studio defaults; not explicitly pinned by CLI. Seed: not set.
Qwen · run 2
Recorded local output · qwen/qwen3.8-27b
- Money collected: $180
- Expenses: not provided
- Profit: not provided
Recording conditions and limits
Recorded: 2026-10-12T10:58:04.229059-05:00. Reasoning: off. Sampling: LM Studio defaults; not explicitly pinned by CLI. Seed: not set.
Edited teaching example
Codex-assisted editorial illustration · not an independent test
- Money collected: $180
- Expenses: not provided
- Profit: not provided
Recording conditions and limits
Author-generated edited example within this existing working conversation, written after inspecting the Qwen outputs. Not an isolated API run. Sampling and model snapshot not measured.
How this case was recorded
Two fresh Qwen CLI runs with reasoning off; other sampling settings inherited. Codex editorial answer uses different system context. Not a controlled ranking.
A small plan
Recorded date: 2026-10-12
The exact user prompt
I have 30 minutes to tidy my desk. I have a trash bag and a cloth. Make a three-step plan with time estimates that add to 30 minutes. Do not suggest buying anything. End with a one-sentence definition of done.
Qwen · run 1
Recorded local output · qwen/qwen3.8-27b
1. **Clear the Surface (10 minutes)**: Sweep all unnecessary paper, empty food wrappers, and clutter into the trash bag, leaving only your active work items.
2. **Wipe and Sanitize (15 minutes)**: Use the cloth to thoroughly wipe down the desk surface, keyboard, and any other surfaces to remove dust and spills.
3. **Final Organization (5 minutes)**: Align the remaining items neatly, ensure the trash bag is tied and ready for disposal, and step back to assess the layout.
**Definition of done**: The desk surface is completely clear of clutter, dust-free, and all remaining items are neatly arranged.
Recording conditions and limits
Recorded: 2026-10-12T10:58:13.358391-05:00. Reasoning: off. Sampling: LM Studio defaults; not explicitly pinned by CLI. Seed: not set.
Qwen · run 2
Recorded local output · qwen/qwen3.8-27b
Here is a 30-minute plan to tidy your desk using only the items you have:
1. **Clear and Sort (10 minutes):** Pick up every item from the desk surface. Discard obvious garbage into your trash bag and quickly gather items that don't belong there into a single pile to be dealt with later.
2. **Wipe and Sanitize (10 minutes):** Use your cloth to thoroughly clean the desk surface, keyboard, and mouse, removing dust, crumbs, and stains while it is clear of clutter.
3. **Organize and Return (10 minutes):** Place only the items you actually need for work back on the desk in specific spots, ensuring the cloth is folded or put away and the trash bag is tied off.
**Definition of Done:** Your desk surface is clear, clean, and contains only the essential items needed for your immediate tasks, with all waste disposed of and cleaning tools stored away.
Recording conditions and limits
Recorded: 2026-10-12T10:58:20.059451-05:00. Reasoning: off. Sampling: LM Studio defaults; not explicitly pinned by CLI. Seed: not set.
Edited teaching example
Codex-assisted editorial illustration · not an independent test
1. Sort for 10 minutes. Put obvious trash in the bag. Keep papers or objects you are unsure about in a separate review pile.
2. Wipe for 8 minutes. Move items aside and use the cloth to remove dust from the desk. Avoid pushing debris into electronics.
3. Reset for 12 minutes. Return the items you need, group the rest, and leave the review pile clearly separate.
Done means there is a usable work area, obvious trash is bagged, and uncertain items are kept for review.
Recording conditions and limits
Author-generated edited example within this existing working conversation, written after inspecting the Qwen outputs. Not an isolated API run. Sampling and model snapshot not measured.
How this case was recorded
Two fresh Qwen CLI runs with reasoning off; other sampling settings inherited. Codex editorial answer uses different system context. Not a controlled ranking.
A welcome-script review
Recorded date: 2026-10-13
Shared user-prompt excerpt
Welcome video draft: “Hey! I’m Patch. Tiny computer. Big curiosity. Welcome to Tinker and Tokens. Wondering what this AI thing can actually do? Start with Everyday AI. We’ll try one small, useful thing together. Feeling more adventurous? Open the Workbench and experiment with prompts and model answers. You can ask me to find a page whenever you need a hand. One little reminder: AI can sound sure and still get things wrong. Stay curious. Check what matters. Ready? Let’s tinker.” Review for friendliness, accuracy and original phrasing; suggest a concise improvement, not more jargon.
Qwen · a focused edit
Recorded review excerpt · qwen/qwen3.8-27b
The line "Open the Workbench and experiment with prompts and model answers" is slightly passive.
Suggestion: "Open the Workbench. Try a prompt, see the result, tweak it." (Active, concise, no jargon).
Recording conditions and limits
Local LM Studio CLI. Reasoning off; remaining sampling defaults not pinned. The Improvement and Suggestion lines are excerpted; Markdown emphasis and section labels were removed for display. Other parts of the review are omitted.
Claude · a fuller rewrite
Recorded review excerpt · Claude Opus 5.5 · Medium
Hey! I’m Patch. Tiny computer, big curiosity. Welcome to Tinker and Tokens. Curious what AI can actually help with? Start with Everyday AI. We’ll try one small, useful thing together. Like taking things apart? Head to the Workbench to build prompts and compare real model answers we’ve saved. Need a page? Just ask me. One reminder: AI can sound sure and still be wrong. Stay curious, check what matters, and let’s tinker.
Recording conditions and limits
Captured from the Claude app in an existing design conversation. The revised-script paragraph is quoted verbatim. Conversation context differs from the other reviewers; no API parameters or exact backend snapshot recorded.
Gemini · a shorter route
Recorded review excerpt · Gemini Flash · High (app labels)
Hey! I’m Patch—tiny computer, big curiosity. Welcome to Tinker and Tokens. Wondering what AI can actually do? Start in Everyday AI with one quick, practical project. Want to see how it works under the hood? The Workbench lets you experiment. Ask me anytime you need directions. Remember: models can sound confident and still be wrong. Ready? Let’s tinker.
Recording conditions and limits
Captured from the Gemini app in an existing design conversation. The proposed-script paragraph is quoted verbatim. App labels are shown as observed; exact backend model snapshot not recorded.
How this case was recorded
All three received the same longer review brief on October 13, 2026. The shared task excerpt above is its welcome-video section; the omitted sections covered site architecture, download safety, and revenue. The outputs shown are selected excerpts, not full replies. Existing app histories and settings differed. This is a record of a real collaboration, not an isolated or controlled benchmark.
What these records cannot establish
OpenAI makes ChatGPT, but these OpenAI responses were recorded through Codex. App instructions and available tools can change the experience. One response per model, or two runs on one small task, cannot establish general reliability, the effect of model size, or a speed ranking. A recorded model label is the identity observed at that time, not a current availability promise.
Try a careful reading exercise first: choose one claim, find its support in the supplied prompt, and write down what remains unknown. Work through the three guided cases →
