Model theory
Separate the model from the tools around it.
A practical explanation of AI models, tokens, context windows, tool access and reasoning effort for choosing a useful ChatGPT or Claude workflow.
Updated
The model is not the entire product
A text model receives inputs and generates outputs. The surrounding application determines which files, search tools, editors and connected sources it can use. This distinction matters: a capable model without the needed document cannot perform a grounded document review.
Before upgrading the model, identify what the task requires. For an advertising asset, distinguish copy, image generation and exact layout. For a code change, distinguish suggesting code from editing and verifying a repository.
For [task], list the required inputs and capabilities. Separate language or reasoning work from search, calculation, visual generation, file editing and external actions. State what is available in this environment and what remains missing.
Tokens are units, not a quality score
Models process text in tokens: chunks that do not map exactly to words. A context window limits the material considered within a request. Input and output share that budget, with reasoning also counting for relevant models. Exact limits depend on the model and product.
A larger capacity can accommodate more material, but it does not establish that every relevant detail was used correctly. Ask for evidence references and inspect the difficult passages. Do not evaluate a report workflow by file-size capacity alone.
Before analysing these files, identify the sections relevant to [questions], unreadable material and any missing source. Return a source map. Do not claim complete coverage unless the required material was actually accessible.
Model choice and effort are separate decisions
Reasoning effort adjusts how much work a supported model devotes to a response; changing the model changes the underlying capability choice. Higher settings can be useful for demanding work, but missing context is still missing context. Product controls and availability differ.
Use a task-specific starting policy. A short rewrite needs preservation of meaning. A conflicting-source analysis needs evidence reconciliation. A coding change needs observed behaviour and verification. Compare the editing and checking required, not simply the length of the answer.
| Problem | First thing to check | Possible next step |
|---|---|---|
| Wrong facts | Were appropriate sources supplied and used? | Fix grounding before changing model |
| Shallow analysis | Was the decision and uncertainty clear? | Try a stronger model or effort setting |
| No usable file | Does the environment support file creation? | Use the appropriate work tools |
| Wrong visual asset | Was an image generator actually available? | Choose a visual production workflow |
Use a small task before making a broad commitment
Choose a representative input and define what successful output contains. Include a difficult case, such as an unreadable table or conflicting dates. Record model, settings and the revision work needed. That gives you local evidence for your workflow rather than a universal intelligence ranking.
Design a comparison for this task using the same input and requirements. Define observable failures, the review method and what metadata to record. Explain what this small test cannot establish about the models more generally.
Common questions
Does a bigger context window mean a better answer?
It means greater input/output capacity, not guaranteed source use or correctness. Check the relevant evidence and omissions.
Is model choice the same as choosing Chat or Work?
No. Model choice concerns capability; a work environment concerns the sources, tools and workflow available for producing the result.
Keep exploring
Sources and editorial note
Guidance checked 9 October 2026. These AI-assisted guides and original templates are reviewed against the linked sources. Results depend on your inputs and available tools. Model recommendations are editorial starting points, not comparative test results.