Which model should you use?

Start with the job, not the model. Name the task, weigh the cost of being wrong, then pick the cheapest model that clears the bar. Below is that decision, broken down by the work you are actually doing, with live picks for each.

Updated Jul 22, 2026

Picks pulled live from the model catalogue

The decision

One question decides most of it

What is the cost of being wrong on this step, and what is the cheapest model that clears that bar?

That single question sorts almost every choice. When a mistake is expensive and you are not running the step millions of times, buy the top of the board and move on. When the step is high-volume and a good-enough answer is fine, the cheapest capable model wins on economics every time. The frontier now spans roughly a fifty-fold range in output price, so the gap between a careful pick and a lazy one is real money.

Everything below is that question applied to specific jobs. Find the row that looks like your work, then use the live picks as a shortlist to test against your own task.

Building a coding agent or writing code

When it fits: The model sits behind an IDE assistant, a code-review bot, or an agent that edits real files across a repo.

Optimize for: Coding Index first, then context window so it can hold enough of the codebase.

Full coding board →

Live pick	Provider	Coding
GPT-5.6 Sol	OpenAI	77
GPT-5.6 Terra	OpenAI	77
Claude Fable 5	Anthropic	77

Running an autonomous agent

When it fits: The model plans, calls tools, and stays on task across a long chain of steps without a human at each turn.

Optimize for: Agentic Index, which is separate from raw intelligence for a reason. A clever model that loses the thread mid-task is worse than a steady one.

Full agentic board →

Live pick	Provider	Agentic
GPT-5.6 Sol	OpenAI	54
Claude Fable 5	Anthropic	53
Kimi K3	Moonshotai	50

High-stakes analysis and decisions

When it fits: An architecture call, a contract review, a financial model, or an eval judge, where a wrong answer is expensive and the volume is low.

Optimize for: Intelligence Index. Buy the top of the board and do not think twice. The token cost is trivial next to the cost of being wrong.

Full intelligence board →

Live pick	Provider	Intel
Claude Fable 5	Anthropic	60
GPT-5.6 Sol	OpenAI	59
Kimi K3	Moonshotai	57

High-volume, lower-stakes work

When it fits: Classification, extraction, tagging, first-draft generation, and anything you run millions of times where good enough is the goal.

Optimize for: The cheapest model that still clears the quality bar. This is where the money is won or lost at scale.

Full cheapest board →

Live pick	Provider	Price
Mistral Small 3.2 24B	Mistral	$0.100 in
Llama 4 Scout	Meta	$0.100 in
Gemini 2.5 Flash Lite	Google	$0.100 in

Long documents and whole-repo work

When it fits: Whole-codebase reviews, long contracts, research over large corpora, or agents that carry heavy state between steps.

Optimize for: Context window first. If it does not fit in the prompt, nothing else matters.

Full context board →

Live pick	Provider	Context
Llama 4 Scout	Meta	10M
Qwen3 Coder 480B A35B	Qwen	1.0M
Gemini 2.5 Flash	Google	1.0M

A note on writing and voice

Not every job sorts on a benchmark. For drafting that has to sound human, tone and instruction following matter more than any index, and the Claude family is the common favourite for it. Read the model families guide for the temperament differences, then trust a quick side-by-side on your own copy over any score.

The expensive mistake

Do not pay frontier rates for everything

The most common and most costly habit is picking one top model and pointing every task at it. It feels safe and it quietly burns money, because most steps in any real workflow do not need frontier reasoning. A support summary, a data extraction, a first draft: these run fine on a model that costs a fraction as much.

The better pattern is to route. Match each step of a workflow to the cheapest model that is good enough for that step, and reserve the expensive models for the few steps that carry the risk. Done across a busy workflow, that is often the difference between a bill you shrug at and one you have to explain.

How model routing works →

Common questions

How do I choose which AI model to use?

Start with the job, not the model. Name the task, decide the cost of being wrong, then pick the cheapest model that clears the quality bar for that specific step. Sort on the metric that matches the work: Coding Index for code, Agentic Index for agents, Intelligence Index for high-stakes reasoning, price for high-volume work, and context window for long inputs.

Which AI model is best for coding?

The top of the Coding Index board, which is refreshed live from independent benchmarks. It changes as new models ship, so the picks on this page are pulled from the current catalogue rather than named once and left to go stale.

Do I need the most expensive model?

Only for the steps where a wrong answer is expensive. For high-volume, lower-stakes work, a mid or budget model that clears the quality bar wins on economics. Paying frontier rates for everything is the most common and most costly mistake.

Can I use more than one model?

Yes, and most serious workflows do. Different steps have different needs, so matching each step to the right model is cheaper and better than forcing one model to do everything. That practice is called model routing.