How to choose an AI model
Match the model to the job: small for volume, mid for daily work, top or reasoning for hard problems and local for private data, with a decision tree and a test.
There is no best model. There is a best model for a given job, budget and level of privacy. This guide shows how to pick one in five questions, and how to check your pick with a small test.
Model names change within months. The ones below are examples from each provider's own pages as of 26 October 2026. For current prices, see AI API prices.
Three classes of model
Most providers sell the same three tiers. Names differ; the idea does not.
| Class | Good for | Trade-off | Examples |
|---|---|---|---|
| Small and fast | Sorting, extraction, short summaries, routing, high volume | Weaker on hard or subtle tasks | Claude Haiku 5.5, GPT-6 Luna, Gemini 3.5 Flash-Lite |
| Mid-range | Daily writing, analysis, code help, chat | A balance of speed, cost and skill | Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash |
| Top model | Hard coding, long agent work, deep analysis | Slower and several times pricier | Claude Opus 5.5 and Fable 5.1, GPT-6 Astra, Gemini 4 Argon |
Anthropic describes Haiku 5.5 as built for high-volume, latency-sensitive tasks such as classification, extraction and routing, Sonnet 5.5 as its best mix of speed and intelligence, and Fable 5.1 for demanding reasoning and long-horizon agent work. OpenAI lists GPT-6 Astra as its most capable model and GPT-6 Luna as its most efficient. Google calls Gemini 4 Argon its flagship reasoning model and Gemini 3.5 Flash-Lite its fastest, most cost-effective 3.5 model.
The decision tree
Five questions in detail
1. How private is the data?
If text must not leave your control, think about this first. Options: a consumer app with training turned off, a business plan or API with a data agreement, or a local model on your own machine. Read what not to paste. Open models you can run yourself are called open-weight models. Examples on the Ollama library: gpt-oss (20b and 120b sizes), Gemma 4 (sizes from E2B to 31B, with image input) and Qwen 3.5 (0.8B to 122B). Small sizes run on a laptop; large ones need strong hardware. Quality is lower than the top cloud models, but often enough for summaries and drafts.
2. How hard is the task?
A reasoning model spends extra steps thinking before it answers. That helps with maths, tricky bugs and planning, and it costs more tokens and time. OpenAI lists reasoning levels from low to max for its GPT-6 line. Anthropic's current models use adaptive thinking steered by an effort setting. For plain writing or chat, a lower effort or a mid-range model is usually enough.
3. How much volume?
Thousands of small jobs favor a small model. Price gaps are large: on the providers' price pages, Claude Haiku 5.5 and GPT-6 Luna list at $0.10 input and $0.50 output per million tokens (Haiku up to 100,000-token prompts), against $2 and $10 for the mid class and $4 and $20 for Claude Opus 5.5 (checked 2026-10-26). Add prompt caching and the Batch API when they apply. For a full table, see AI API prices.
4. How long is your input?
The context window is how much text a model can take in one request. Current flagship lines list about one million tokens: Claude models 1M, GPT-6 models 1.05M, Gemini 4 Argon 1M. That fits a long report or a small codebase. But more input costs more, and long prompts can cost extra: Haiku 5.5 moves to higher rates above 100,000 tokens. Send what the task needs.
5. Does it need images, audio or files?
A multimodal model takes more than text. All current Claude models and the GPT-6 flagship line accept text and image input. Google lists separate models for audio, speech and image or video generation. Check the model page for the exact inputs before you build around one.
Quick picks by task
- Summarize a long report: a mid-range model with a large context window, or split the text into parts.
- Sort 10,000 support tickets: a small model, run through the Batch API if the result can wait.
- Fix a hard bug in a big codebase: a top or reasoning model, often inside a coding agent.
- Draft emails and posts: a mid-range model; a small one for short, plain mails.
- Read a photo, a scan or a chart: a multimodal model.
- Chat about private notes: a local model, or a plan with strong data terms.
- Build a helper that calls other tools: check that the model supports tool use before you build an AI agent.
Speed matters too. Small models answer faster, which counts for live chat and voice. Top models and high effort settings think longer, so a reply can take much more time. Pick the speed your user can accept.
Check your pick with a test
Benchmarks are a hint, not proof. They test general tasks, not yours.
- Collect ten real tasks from your own work.
- Run each on a small model, a mid model and, for hard tasks, a top or reasoning model.
- Score the answers: correct, usable, wrong.
- Pick the cheapest model that passes most of them.
- Test again when a new model arrives. Models are retired too: Anthropic, for example, lists retirement dates for each model.
Also check the knowledge cutoff if your task needs recent facts. Anthropic lists June 2026 for its current models. Models can still hallucinate, so verify anything that matters.
Cost and plan
Pay-per-token and flat plans are different deals. See subscription or API for the math. In a plan, the app decides which models you can pick; in the API, you choose per call.
Tip: Start with a mid-range model. Move down if answers are good enough, up only if they are not.
Warning: A model name is not a quality guarantee. Prices, limits and names change; confirm on the provider page before you commit.
Questions and answers
Is the biggest model always the best choice?
No. A top model costs several times more and answers more slowly. For simple tasks such as sorting, short summaries or extraction, a small model is often good enough.
When do I need a reasoning model?
When the task has many steps that must fit together: maths, hard code bugs, planning with constraints, careful analysis. For plain writing and chat it adds cost without much gain.
Does a bigger context window mean better answers?
No. It lets you paste more text. Models can still miss details in very long input, and you pay for every token you send.
How do I know which model works for me?
Test it. Take ten real tasks, run them on a small and a larger model, and compare the results. Keep the cheapest model that passes.
Read next
- AI API prices compared: cost per 1 million tokens
A table of AI API list prices per 1 million tokens, with source and check date for every row.
- Context window
The context window is the maximum amount of text, counted in tokens, that a model can read and write in one request.
- Reasoning model
A reasoning model spends extra tokens on hidden step-by-step thinking before it writes its answer, which helps on hard problems but costs more.
- Multimodal
A multimodal AI model can take in, and sometimes produce, more than one kind of data, such as text, images, documents, audio or video.
- Local model
A local model is an AI model that runs on your own computer or server instead of a provider's cloud, so prompts need not leave your device.
- Open-weight model
An open-weight model is an AI model whose trained weights are published for download, so you can run and often adapt it yourself.
- Benchmark
An AI benchmark is a fixed set of tasks with a scoring rule, used to compare models on the same test.
- Knowledge cutoff
The knowledge cutoff is the date after which a model's training data stops, so it knows nothing about later events unless you give it fresh information.
- What not to paste into AI chatbots
Passwords, client data, IDs and secrets do not belong in a chatbot: a traffic-light checklist, provider data rules, Swiss and EU basics, and safer options.
- Subscription or API: what is cheaper for you?
A flat $20 plan or pay per token: one formula and three examples show where each wins, plus the limits and privacy settings that count beyond cost.