AI models, apps, APIs and agents: a practical guide

You ask an assistant to prepare a sponsor report. It writes a convincing summary, but the figures come from last month's spreadsheet and no file appears in your project folder. Is that a weak model, the wrong app or a missing connection? Those are different problems, and buying a more expensive model may leave all of them in place.

The distinction becomes easier when you follow one job through the system. This guide uses an illustrative sponsor-report task: read approved campaign figures, compare them with agreed targets and produce a draft for review. No report is sent automatically.

Four layers, one deliverable

These are practical definitions; vendors do not use every term identically. An app can contain an agent, and that agent can use an API to call a model. You do not have to choose one of the four and discard the others.

Watch the sponsor report move through the layers

Start with the simplest version: upload a permitted spreadsheet in an app and ask for a draft. The model can reason only from the information it receives, while the app determines how the file reaches it and where the result appears. An uploaded file name alone does not prove every sheet or formula has been interpreted correctly.

Before drafting, ask for a small intermediate table: campaign, reporting period, actual result, target and missing information. Compare those rows with the original file. If September's actual result is compared with August's target, stop at that stage. A beautifully written conclusion would only conceal the wrong comparison.

Now imagine wanting the same report every Monday from a shared folder. That adds work around the model: choosing the current file, obtaining access, handling missing data, saving the report and recording failures. An app may already provide the needed workflow. If it does, building an API integration solely to use the same model adds maintenance without necessarily improving the report.

A custom API integration makes sense when a specific requirement cannot be met in the available app—for example, fitting the report into an existing internal system with its own access rules. The API supplies a connection to model capabilities. Your software still needs to decide which records to send and what to do with the response.

An agent becomes useful when the path varies: it may need to find a missing target in another approved source or inspect a second file before completing the table. If every step is fixed, ordinary automation may be easier to inspect. Either approach needs the same acceptance criterion: the correct period, correct figures, labelled gaps and a draft in the intended location.

Example: GPT-6 and ChatGPT

OpenAI's current model catalogue lists GPT-6 Astra, Sol and Luna. Those are model identifiers and families. ChatGPT is a product experience; an account's available controls are a separate question from the API catalogue. Official model catalogue.

When asking for help, name the app, the model shown in it and the task. “Use GPT-6” is less reproducible than a description of the actual environment and expected output. Do not expose credentials when sharing that context.

Example: Claude Code and Claude 5.5

Claude Code is a coding application. Opus 5.5 and Sonnet 5.5 are models described in Anthropic's release announcements. A problem may depend on the application version, selected model, repository instructions or tool permissions. Claude Code overview; Sonnet release.

Keep those details separate when comparing results. If you change the model, prompt, repository state and available tools at once, you cannot tell which change helped.

Example: voice interface and reasoning backend

GPT-Live illustrates how one application can use separate components: a voice model manages the conversation while a backend handles delegated reasoning and tools. Gemini also distinguishes its Live app feature from a developer Live API. GPT-Live guide; Gemini API overview.

A demonstration of natural speech is not proof of reliable business actions. Check whether the correct record was read or changed, and whether the system respected the user's instruction when they corrected it.

Diagnose the earliest broken step

Return to the sponsor report before changing models. Inspect the evidence in order: selected file, extracted rows, calculation, written explanation and saved result. Fixing the earliest failure prevents unnecessary rewrites later.

The last row is the point where model choice becomes an isolated experiment. Keep the same source packet and acceptance criteria so an improvement can be attributed more sensibly. Changing the file, prompt, model and tools together makes diagnosis harder.

Write the brief before buying the stack

For this illustrative job, a useful brief is: “Every Monday, prepare a draft sponsor report from the approved weekly file. Include actuals and targets for the same period. Flag missing targets. Save the report for review, and do not send it.” Add who owns the source, who reviews the result and where the output belongs.

Then attach one small example with a known result. Suppose a campaign has 12,000 actual views against a 10,000-view target. The difference is 2,000 and the result is 20% above target. If the target is absent, the report should show it as missing rather than calculate a percentage. These invented records let you test both success and an exception without sharing client data.

Approve the workflow when the report, its calculations and the saved artifact all agree. A statement that the report was saved should point to something you can open. If the task is occasional, the manual app workflow may be sufficient; recurrence alone does not require a custom agent.

Avoid three common mistakes

Do not treat an API price as a consumer subscription quote. Do not infer a feature from a model name without checking the app. Do not call a workflow autonomous without deciding how it handles ambiguous or irreversible actions.

Start with the smallest useful process and expand after testing. For practical next steps, read prompting for real work, Codex tips or the live voice comparison.