Use the smallest, fastest model that does your task well, and step up only when it falls short. Quick questions and simple edits need a fast model. Everyday writing, analysis and coding suit a balanced one. Save the deepest, slowest tier for hard, multi-step problems where a wrong answer costs more than a slow one.
The only reliable way to know is to try a few models on your own work. This guide gives you the questions that matter, a decision table and a test you can run in an hour.
- Match the model to the job: fast for simple, high-volume work, balanced for everyday tasks, deep for hard problems.
- Every major lab sells the same three tiers under different names.
- Seven questions decide it: task, quality bar, speed, cost, context size, privacy and tool use.
- Benchmarks narrow the field, but only a test on your own tasks picks the winner.
- Re-check your choice every few months, because the lineup changes fast.
The seven questions that decide it
Before you compare names, answer these about the job itself. Most choices are obvious once you do.
| Question | Why it matters | Rule of thumb |
|---|---|---|
| What is the task? | Some tasks need several steps of thought; most do not | Lookups and rewrites are easy; planning and debugging are hard |
| How good must it be? | A draft you will edit forgives more than a contract clause | The higher the cost of a mistake, the bigger the model |
| How fast must it answer? | Big models make you wait, sometimes for minutes | Live chat needs speed; an overnight job does not |
| What can you spend? | You pay per token, many times over at volume | Price the whole job, not one request |
| How much must it read at once? | Long documents and codebases need a big context window | Check the window before you check the price |
| Where can the data go? | Some data cannot leave your control | Read the provider’s data terms first |
| Does it need tools? | Agents that browse, run code or call APIs need reliable tool use | Test tool-heavy work separately |
A token is a small chunk of text, about three quarters of an English word. If that is new to you, read tokens and context windows, explained without math first.
Fast, balanced and deep: the three tiers
Nearly every AI lab sells its models in sizes. Smaller models answer faster and cost less. Larger ones handle harder problems, but you wait longer and pay more. Our guide to small language models covers just how small a capable model can be.
Think of it like staffing a task at work. A quick errand goes to whoever is free. A tricky design problem goes to your most experienced person, even though their time costs more.
Here is how the current lineups line up. The model names are the vendors’; the grouping is our reading of how each vendor describes them.
| As of September 2026 | Fast | Balanced | Deep |
|---|---|---|---|
| Best for | Quick, high-volume tasks with a clear goal | Everyday writing, analysis and coding | The hardest problems and long agent runs |
| Anthropic | Claude Haiku 4.5 | Claude Sonnet 5 | Claude Opus 5.5, Claude Fable 5.1 |
| OpenAI | GPT-6 Luna | GPT-6 Sol | GPT-6 Astra |
| Gemini 3.5 Flash-Lite | Gemini 3.8 Flash | Gemini 3.1 Pro (preview) |
Our own product, Cyborb, follows the same pattern with its built-in agent. Air is instant and free, “the model you leave running”. Liquid is the balanced everyday worker. Plasma is the deepest reasoning, for the hardest problems. The models docs cover the details.
Many models also have an effort or thinking setting that slides one model up or down this scale. Anthropic’s own guide says tuning effort is often a better lever than switching models. Our guide to reasoning models explains when extra thinking pays off.
Which AI model should I use? A decision table
Find your task, start with the suggested tier, and move up only if your test shows you need to.
| If your task is... | Start with | Why |
|---|---|---|
| Quick questions, rewording, short replies | Fast | Answers in a moment, and the quality gap is small |
| Sorting, tagging or extracting data from thousands of items | Fast | Costs multiply with volume; check accuracy on a sample |
| Drafting emails, summarizing reports, everyday coding | Balanced | Good quality at a fair price and speed |
| Reading a long contract, book or codebase in one go | Any tier with a large context window | Everything has to fit at once |
| Fixing a stubborn bug, planning a system, tricky analysis | Deep | A wrong answer costs more than a slow one |
| Agent work that runs on its own for an hour or more | Deep | Needs reliable planning and tool use over many steps |
| Private data that cannot leave your control | A model you can run yourself, or a provider with suitable data terms | Privacy decides before quality does |
How much does an AI model cost?
Through an API, you pay per token, with separate prices for what you send and what the model writes back. As of September 2026, Anthropic’s lineup runs from $1 in and $5 out per million tokens for Claude Haiku 4.5 to $10 and $50 for Claude Fable 5.1. Google’s runs from $0.30 and $2.50 for Gemini 3.5 Flash-Lite to $2 and $12 for Gemini 3.1 Pro.
The price list understates the gap. Deep models often think before they answer, and that hidden thinking is billed as output. A hard question can cost many times more on a deep model than the same question on a fast one. Our full pricing comparison, LLM API pricing compared, lines up input, output, cached and batch prices across five labs.
Chat apps and agents usually charge a flat monthly fee with usage limits instead, which hides tokens but not the trade-off. Deeper tiers use your allowance faster. Our guide to cutting your AI bill covers the savings that do not hurt quality.
How to test models on your own tasks
Public leaderboards are a useful first filter, and our guide to AI benchmarks explains what they measure. But they test someone else’s tasks. Anthropic’s docs call a good evaluation set “the most important step in the process.” Our step-by-step guide to LLM evaluation for beginners shows how to build and grade one. Yours can take an hour.
Collect five to ten real tasks
Pull them from your actual work, including one or two that went badly before. Real inputs beat invented ones.
Write down what good looks like
For each task, note the facts that must be right, the format and the length. This is your answer key.
Run each task on two or three models
Use the same prompt and files, in a fresh chat each time. Note how long each one took and, if you pay per token, what it cost.
Compare the answers blind
Remove the model names, label the answers A, B and C, and score them against your answer key. A second model can help grade, but check its verdicts yourself.
Pick the cheapest model that passes
If the fast model passes, use it. Move up a tier only for the tasks where it failed.
You are grading answers to the same task from different AI models. The task: [paste the task] A good answer must: [paste your answer key] For each answer (A, B, C), score accuracy, completeness and format from 1 to 5, and list every factual error. Then say which answer you would use and why. Do not guess which model wrote which.
Keep your task set. When a new model ships, rerunning it takes minutes, and you will know whether the upgrade is real for your work.
Mix models instead of marrying one
You do not have to choose one model for everything. Anthropic’s docs describe two common patterns. In one, a cheaper model does the work and escalates hard decisions to a stronger one. In the other, a strong model plans and hands the bulk work to cheaper helpers.
Most people can use a simpler version: a fast or balanced model by default, and a deep one when a task gets hard. Switching takes seconds, and it keeps the expensive model for the moments that need it.
FAQ
What is the best AI model right now?
There is no single best model for every task, and the leaders change every few months. Shortlist two or three current models from the major labs, then test them on your own work.
Is a more expensive model always better?
No. For simple tasks, a fast model often gives an equally good answer, sooner and for a fraction of the price. Bigger models earn their cost on hard, multi-step problems.
Do I need a reasoning model?
Only for tasks that need several steps of thought, such as math, planning or debugging. For quick questions and rewrites, extra thinking adds delay and cost without a better answer.
Should I use one model for everything?
It is simpler, but it usually means overpaying for easy work or underpowering hard work. A fast default plus a deep model for hard tasks covers most needs.
How often should I revisit my choice?
Whenever a major model ships, or at least every few months. Keep your test tasks, so checking takes an hour instead of a week.
- Use the smallest model that passes your test, and step up only where it fails.
- Fast, balanced and deep tiers exist at every major lab under different names.
- Price the whole job: volume, output length and hidden thinking all add up.
- Five to ten real tasks with an answer key beat any public leaderboard.
- Mix models: a cheap default for routine work, a deep one for hard problems.
Next, learn when thinking longer pays off, or read what AI benchmarks actually measure.
- Choosing the right model, Anthropic, September 2026
- Models overview, Anthropic, September 2026
- Gemini models, Google AI for Developers, September 2026
- Gemini Developer API pricing, Google AI for Developers, September 2026
- OpenAI’s GPT-6 Sol and GPT-6 Luna now available, GitHub, September 2026
- OpenAI launches GPT-6 Sol and Luna, TechCrunch, September 2026
