Home › Agent Harness › Which model

AI · Choosing

Which AI model for which job

There is no best model. There is a best model for each kind of work, and it changes every few months. So this page gives you a pick per job and a way to test, not a ranking. For data and agent work the rule is short: pay for the strongest model when you build, and use the smallest model that passes your own tests when an agent runs.

Written 20 November 2026. No benchmark scores are quoted, because leaderboards disagree and age within weeks. This page was written with Claude, a model from one of the vendors discussed, which is a reason to trust the method (test it yourself) and not the picks. The last section lists what to verify.

The short answer

1 · Build with the bestCoding, specifications and long edits: the strongest frontier model. A weaker one costs you rework.
2 · Run with the smallest that passesAn agent in production: a mid-tier or open model, chosen by your own scenarios, not by a leaderboard.
3 · Keep the choice cheap to changeThe model is a setting behind one interface, so swapping is a config change and a re-test, not a rewrite.

A pick for each kind of work

Read down for the job you have. The examples are what we saw on offer in November 2026 and will date.

JobWhat it needsClass of modelExamples on offer nowVerdict
Building: coding, long edits, writing specifications and pagesThe best reasoning and the longest reliable work on a task. A person watches.The strongest frontier model you can pay forClaude's top tier (Opus or above), OpenAI's flagship, Gemini's top tierPay for the frontier here. The cost is small next to your time, and a weaker model costs you rework.
An agent in production: tool calls in a loop, structured answersReliable tool calling, valid JSON every time, follows instructions, steady cost per task.A mid-tier model, tested on your scenariosClaude's mid tier (Sonnet class), OpenAI's mid models, Gemini Flash class, or an open model such as gpt-oss-120bPick the smallest model that passes your own scenarios, then keep re-testing. Do not start at the top.
Bulk and cheap: classify, extract, route, tidy textSpeed and low cost; the task is narrow and checkable.A small modelClaude's small tier (Haiku class), Gemini Flash-lite class, small open modelsUse the smallest that meets your accuracy bar on a labelled sample. Check the errors, not the average.
Judging other models' answersBetter than what it grades, consistent, and ideally from a different family.The strongest model you can afford, checked against human labelsA frontier model from a different vendor than the one being gradedCalibrate it on cases people have labelled. An uncalibrated judge is an opinion.
Reading long documents and specificationsLong context that is actually used, not just accepted.A long-context model, tested on your documentsGemini's long-context models, Claude's long-context optionsTest recall on your own files: plant a fact in the middle and ask for it.
Regulated or private data (health, pharma, finance)The data terms, the region, and who may see prompts.A model reached through your own cloud account or hosted by youModels offered by your data platform or cloud (for example Databricks Foundation Model APIs, Azure, Bedrock, Vertex) or open-weight models you hostWhere the model runs decides first, which model second. Read the retention and training terms.
Personal and learning projects, no budgetFree, good enough, no card needed.A free hosted open-weight modelgpt-oss-120b and Llama 3.3 70B on a free data-platform workspace (we ran both)Start free. Move up only when a test shows the free model fails.
Search by meaning (embeddings)A good embedding model and a stable one.An embedding model, not a chat modelbge or gte class models on your platformChanging it means re-embedding everything, so pick once and record why.

How to choose, in five steps

What to measure

MeasureThe question
Tool-call reliabilityDoes it call the right tool with valid arguments, and stop when it should?
Valid structureDoes it return the JSON shape you asked for every time, or only usually?
GroundingDoes every number in its answer match what a tool returned? Count the exceptions.
Refusing correctlyDoes it decline what it must (advice, filing, claims of cause) and still help with what it may?
Cost per successful taskNot price per token: total cost of runs that finished correctly, including retries.
SpeedTime to a finished task, and how it behaves under load.
Data termsRegion, retention, training use, and whether your cloud already approves the provider.
AvailabilityIs it offered in your region and on your platform today, and for how long?
One real example. A drug-safety agent we built has a rule that every number in a draft must equal a tool result. We test candidate models on twenty drug and event pairs and count how many drafts break that rule. That one deterministic check told us more than any published score, and it works for any model you try.

Where the model runs matters as much as which model

Check before you rely on this

Where to go next