Understand AI
AI models and tools: how to choose for the job
Know what you are comparing across OpenAI, Claude, Gemini, Grok, Llama and Perplexity, and pick a tool for one business job with a trial plan, not a ranking.
For: Owners and working professionals choosing an AI tool for a real task, without a technical background.
- Format
- Reference + selection worksheet
- Published
- September 9, 2026
- Last checked
- September 9, 2026
- Written by
- Dryvn AI editorial
- Edited by
- Dryvn AI editorial
- Reviewed by
- Not yet reviewed by a named person
- Read as
- Vendor-reported (from official documentation)
What you are actually comparing
Someone asks "should we use ChatGPT or Claude or Gemini?" and the honest answer is: for what, and which part of them? Each name covers several different products. Comparing the app from one company against the API from another is like comparing a truck to an engine. Sort the pieces first and the choice gets easier.
| Layer | What it is | Example of the kind of thing | Who usually chooses it |
|---|---|---|---|
| Provider | The company that trains models and runs the service. | OpenAI, Anthropic, Google, xAI, Meta. | Whoever owns your data and vendor decisions. |
| Model | The underlying system that reads input and produces output. Providers ship families with several sizes. | A GPT or reasoning model, a Claude model, a Gemini model, a Grok model, a Llama model [S20][S21][S22][S23][S24]. | Developers, or the app on your behalf. |
| Assistant app | The product you type into. It wraps a model with memory, files, search and connections. | ChatGPT, the Claude app, the Gemini app, Grok, Perplexity. | You, per person or per team. |
| API / platform | The developer service for building your own software on a model. | OpenAI API, Claude Developer Platform, Google Cloud and Google AI offerings, xAI API [S20][S21][S22][S23]. | Your developer or a vendor like Dryvn. |
| Tool | A capability the model can use: web search, file search, code execution, connecting to your calendar or drive. | OpenAI's hosted tools and connectors; Claude's Connectors and web search; Gemini's connected Google services; xAI's web search and function calling [S20][S21][S22][S23]. | Set by the app or the developer, limited by your permissions. |
The layer most owners should care about is the assistant app, because it decides what the model can reach. A model on its own knows nothing about your business. An app with your files connected and web search turned on is a different animal, and the same model behaves differently inside two different apps.
The major ecosystems, as the vendors describe them
Every statement in this section comes from an official page we fetched on 9 September 2026. These are vendor claims. We have not run side-by-side tests, and plan names, limits and prices change without notice. If a detail matters to your decision, open the source and check the date.
| Ecosystem | Assistant app | Developer access | Connected capabilities named on the official page | What we could not verify |
|---|---|---|---|---|
| OpenAI | ChatGPT (consumer page not fetched in this check). | OpenAI API with GPT and reasoning model families; Responses API [S20]. | Web search, file search, code interpreter, computer use, image and speech, MCP servers and connectors [S20]. | ChatGPT plan names and limits. |
| Anthropic (Claude) | Claude app: Free, Pro, Max, Team, Enterprise [S21]. | Claude Developer Platform, also called the Claude API [S21]. | Web search, Projects, Connectors, Claude Code, Claude in Chrome, Claude for Microsoft 365, Skills [S21]. | Usage limits per plan. |
| Google (Gemini) | Gemini app: Free, Google AI Plus, Google AI Pro, Google AI Ultra; Gemini for Google Workspace for business [S22]. | Enterprise and developer access through Google Cloud, per the same page [S22]. | Deep Research, Gems, Canvas, Gemini Live, connections to Gmail, Docs, Drive, Maps and YouTube [S22]. | Developer platform details beyond the mention of Google Cloud. |
| xAI (Grok) | Grok app (consumer site not fetched in this check). | xAI API with the Grok model family; Grok Build for coding [S23]. | Function calling, web search, structured outputs, batch processing [S23]. | Consumer plans and the X-integrated features. |
| Meta (Llama) | No assistant app named on the developer page. | Llama 4 and Llama 3 model families available through GitHub and Hugging Face for developers to run or host [S24]. | None named; the point of open-weight models is that you or a partner run them. | Licence terms. Read the licence before any commercial use. |
| Perplexity | Perplexity, an answer engine that responds with linked sources [S04]. | Not covered on the page we checked. | Web search with citations; a crawler used for search results, not for training models [S04]. | Plan names, research features, model choice. |
Choose by the job, not by the brand
The useful question is what the job needs. Five kinds of work cover most of what a small business asks an assistant to do. For each, the table says what to look for and what usually goes wrong.
| Kind of work | What the tool needs | What usually goes wrong | Who should check the output |
|---|---|---|---|
| Research | Web search with sources shown, dates on the results, an easy way to open the original. | Confident answers with no source, or a source that says something different. | Whoever will act on it opens two of the sources. |
| Writing | A way to give it your examples and rules (a project or a saved instruction), and simple editing controls. | Generic tone, invented facts about your business, wrong terms. | The person whose name goes on it. |
| Coding and formulas | Code execution or a way to run and test what it writes, and version control if it touches real software. | Code that looks right and fails on real data. | Someone who can run it on a copy first. |
| Document work | File upload or a connection to where your documents live, with clear limits on what it can read. | Answers drawn from the wrong document version, or from nothing at all. | Whoever owns the document checks the quoted passage. See how AI uses your documents. |
| Connected actions | Connectors to your calendar, email, files or systems, plus explicit approval before anything is sent or changed. | An action nobody authorized: an email sent, a record overwritten. | The owner of the record. See AI approval boundaries. |
Notice that the right-hand column never says "nobody". Every tool on this page can produce a plausible wrong answer. The job of the checker is part of the job, and it should be decided before the tool is chosen. If no one can check the output, that is a reason not to use AI for that task yet, which is the same rule as in what to automate first.
App or API?
If a person will read every output, an assistant app is almost always enough. Apps come with the file handling, search and memory already built, and the plan price covers usage for one person or a team. The API matters when you want software to call the model without a person in the middle: a form that gets summarized automatically, a document that gets classified as it arrives. That is a development project, with its own testing, monitoring and cost per call. Many small businesses never need it, and the ones that do usually get it through a product rather than building it.
Illustrative example · fictional
A bookkeeping practice picks tools for three jobs
A fictional two-person bookkeeping practice wants help with three things: answering client questions about CRA filing deadlines, drafting monthly summary emails, and pulling totals out of scanned receipts. They start by listing the jobs, not the brands.
For the deadline questions, the job is research with sources, and the risk is a confidently wrong date. They look for an assistant with web search that shows the source, and they make the rule that the answer is only used after someone opens the CRA page it points to. For the monthly emails, the job is writing in their own tone, so they want a tool that keeps a project with their past emails and a note of what never to promise. The partner whose name goes on the email checks every one. For the receipts, the job is document work on client financial data. Before choosing anything they ask where uploaded files are stored, whether the plan they are on excludes their data from training, and whether they can delete uploads. Two of the three apps they consider answer those questions on a public page; the third does not, and it is out.
They run a one-week trial on real work with the checks in place. One app is better at the receipts, another at the emails. They keep both on the lowest plan that fits, write down which job each one is for, and put a date in the calendar to look again in three months. No ranking was needed. This example is fictional and illustrates the method, not any vendor's performance.
A trial plan that takes one week
- 01Name the job and the checkerOne task, one person who will check the output, and what "good" looks like in one sentence.
- 02List the data the tool will touchCustomer names, financials, contracts, nothing sensitive. Read the vendor's data-use page for the plan you are on before you upload anything.
- 03Pick two candidate apps, not fiveChoose from the table above by capability, not by reputation. Use the free or lowest plan for the trial.
- 04Run the same ten real items through bothReal inputs from last month. Keep the outputs and the checker's notes side by side.
- 05Count what the checker had to fixNot a score, a list. Wrong facts, wrong tone, missing steps. This is your evidence.
- 06Decide, write it down, set a review dateWhich app, for which job, with which check, until when. Plans and models change; your note is what lets you re-run the trial fairly.
Worksheet
Choosing for the job
One sheet per job. Fill it in before you open a pricing page.
- The job
- One task, in one sentence
- Inputs
- What the tool will be given: text, files, web, your records
- Output needed
- Draft, answer with sources, table, action taken
- Who checks it
- A named person and what they check
- Data sensitivity
- Customer, financial, contractual, none
- Connected actions needed
- None / read only / draft only / act with approval
- Candidate tools (two)
- Which app and plan, and why by capability
- Trial plan
- Ten real items, one week, what counts as a fix
Downloads · no email required
- Download .md
AI tool selection worksheet
The 'Choosing for the job' sheet as a fill-in Markdown file, with a separate fictional example.
Questions people ask
- What is the difference between an AI model and an AI app?
- The model produces text or other output from what it is given. The app wraps the model with memory, file access, search and connections, and decides what the model can see and do. Two apps on the same model can behave very differently. For most owners, the app is the real choice.
- How do I choose an AI model for a business task?
- Start from the task, the data it touches and who checks the result. Pick two apps that have the capability the task needs, run ten real items through each for a week, and count what the checker had to fix. Choose from that evidence and write a review date.
- Is an open-weight model like Llama cheaper?
- It can be, if you have someone to run it and keep it running. Meta distributes Llama families for developers to host themselves or through partners [S24]. The licence and the hosting cost are the real questions, and neither is on this page.
- Which one is best?
- For which job, checked by whom? Every vendor here publishes strong claims. We have not tested them against your work, so we will not rank them. The worksheet gets you an answer that is true for your business.
Related resources
- AI assistant, automation, agent or business software: which do you need?Tell the four categories apart by the job each one does, and pick the right one for a specific task in your business, including the case where what you already own is enough.
- How multi-model AI systems work, and when one model is enoughSee what happens when several AI models share a job, why disagreement is useful, what it costs, and how to write down roles and checks before you build.
- How AI uses your documents: search, context and retrievalUnderstand what actually happens when an assistant answers from your company documents, why that is not the same as training, and what to check before you trust the answer.
- Where business AI may go next: facts, forecasts and open questionsSort what AI can demonstrably do for a business today from what vendors have announced and what is still a scenario, and decide what evidence would change your mind.
- BlogMore AI models do not automatically make a better systemAdding a second or third model to a business process only pays when one of them has a distinct job, including the job of disagreeing. Otherwise you have bought delay and cost.
Sources
Dates are when each source was last checked by the editor. Sources support specific claims; they are not endorsements.
- S20OpenAI API platform overview · OpenAI · checked September 9, 2026Supports: OpenAI offers an API platform with GPT and reasoning model families, and hosted tools including web search, file search, code interpreter, computer use, and MCP connectors. The consumer ChatGPT pages could not be fetched in this check, so ChatGPT plan names are not verified here.
- S21Claude product overview · Anthropic · checked September 9, 2026Supports: Claude app plan names (Free, Pro, Max, Team, Enterprise), the Claude Developer Platform (Claude API), model family names, and features such as web search, Projects, Connectors, Claude Code and Claude for Microsoft 365.
- S22About the Gemini app · Google · checked September 9, 2026Supports: Gemini app, plan names (Free, Google AI Plus, Google AI Pro, Google AI Ultra), Deep Research, Gems, connected Google services, and Gemini for Google Workspace. Prices shown on that page change and are not repeated here.
- S23xAI developer documentation overview · xAI · checked September 9, 2026Supports: the xAI API, the Grok model family, and developer tools including function calling, web search and structured outputs. The consumer x.ai site could not be fetched, so Grok app plan names are not verified here.
- S24Meta AI developer platform (Llama) · Meta · checked September 9, 2026Supports: Llama 4 and Llama 3 model families distributed through GitHub and Hugging Face for developers to run themselves or through partners. Licence terms were not verified on this page and must be read before use.
- S04Perplexity crawlers · Perplexity · checked September 9, 2026Supports: Perplexity answers questions with linked sources and its crawler is for search results, not for training foundation models. Consumer plan names could not be verified in this check.
R13 · Published September 9, 2026 · Next scheduled review November 9, 2026 · Teaches process management; not legal, warranty, safety or engineering advice. Examples are fictional unless stated. Part of the Dryvn resource library (17 resources).