Choose a model for your job

This page helps you decide which model to put behind your agent for the work you actually do. It applies to every agent in this guide, and to cloud and local models alike. Pick your model and budget already rates each option from 1 to 5, and this page shows how to match those rows to a job. It also explains the scores on model pages, and how to run a short test of your own.

Everything here was checked on 15 November 2026. Benchmark versions and rankings change every few weeks, so treat each number as a snapshot rather than a promise.

Start from the job, not the leaderboard

A model at the top of a leaderboard can still be the wrong pick, because each job needs a few specific things from the model. Find your job in the first column, check what it needs, and then look up the matching rows in the full table.

If two jobs matter to you, choose for the harder one, or give each job its own model as Mix a cheap model with a strong one shows. For local models side by side, see Local models compared.

Check what a model can do before you test it

Every model page lists what the model can do, and a quick look there saves a wasted test. On Ollama's library, each model shows badges such as tools, vision, and thinking, and the search page can filter by Tools. For a cloud model, open the provider's own model page, since Google's page for gemini-4-hydrogen marks Function calling and Computer use as supported. Then check the pricing page as well, because a free tier can switch off a feature that the model itself supports.

What the numbers on model pages mean

A model card usually ends with a table of benchmark scores. Each benchmark tests one kind of work, so a high score predicts success only at that kind of work. The list below covers every benchmark this guide quotes, described from each benchmark's own site.

• Terminal-Bench 2.1 gives an agent real tasks in a terminal, with names such as fix-git and train-fasttext. A test script then checks whether each job is done. It predicts how well a model handles agents that run commands. Version 2.1 fixed 28 of the 89 tasks in version 2.0.
• SWE-bench Verified hands the model a codebase and a real GitHub issue, and asks for a change that fixes the problem. Verified is a set of 500 problems that engineers confirmed can be solved. It predicts fixing bugs in existing code.
• SWE-bench Pro is a harder set of 1,865 coding tasks from Scale AI, including some private company code that models are unlikely to have seen. At launch, Scale said models scoring above 70% on Verified scored about 23% on Pro. It predicts larger coding changes across several files.
• LiveCodeBench v6 uses problems from coding contests on LeetCode, AtCoder, and Codeforces, published from May 2023 to April 2025. It predicts writing a short, self-contained program, not running a long chain of tool calls.
• TAU2-Bench has an agent help a simulated customer while following a written policy and calling tools. In its newer tasks, the customer has to take some of the steps too. It predicts following rules over a long conversation with tools.
• BFCL-V4, the Berkeley Function Calling Leaderboard, checks whether a model calls tools correctly, and version 4 adds tests of agent work. It predicts whether a model sends each tool request in the right shape. Some cards still quote the older BFCL v3.
• DeepSWE 1.1 is a set of 113 long engineering tasks from Datacurve, written from scratch so models cannot have seen them. It predicts long coding jobs that take many steps.

Publisher claims and independent results

A score on a model card is the publisher's claim, because the company that made the model also ran the test. That does not make it false, but it was not checked by anyone else. This guide labels those numbers as claims, and an independent result is one that another group ran under its own rules.

Scores from different test setups are not comparable, even on the same benchmark. The program that drives the model during a test changes the result. The Terminal-Bench team reported that the 2.1 fixes raised one model's score by 12.1% when Claude Code drove it, but by only 0.9% under Terminus 2, a different test program. Epoch AI, which tracks benchmarks, warns that small changes in prompts or settings can shift results a lot. So a card that says "Terminus" beside a score cannot be lined up against another card's score from a different setup.

Versions matter too. The Terminal-Bench site now shows version 4.1 by default, while the cards in this guide quote version 2.1. The Terminal-Bench team calls its 4.0 changes breaking changes that need every test run again, so a 2.1 score and a 4.1 score cannot be compared.

Some benchmarks have known faults. Epoch AI rated both SWE-bench Pro and DeepSWE 1.1 as flawed in September 2026. It cited audits that found 30% or more of SWE-bench Pro tasks broken, and problems in at least 23 of DeepSWE's 113 tasks. A gap of a few points on either one tells you little.

Independent places to compare models

None of these sites knows your job, so use them to build a short list and then run your own test.

• The benchmarks' own leaderboards. Terminal-Bench, SWE-bench, SWE-bench Pro, TAU2-Bench, BFCL, and DeepSWE each keep a public leaderboard. DeepSWE runs every model through the same program, which makes its rows easier to compare. Several of these boards do not say on the page whether each row was checked or sent in by the model's maker.
• Arena, formerly LMArena and before that Chatbot Arena. Visitors compare two anonymous answers and vote, and the names appear only after the vote. That measures which answer people prefer, which is close to everyday chat and writing. Its separate Agent leaderboard ranks models from real agent sessions instead of votes. Arena says it may share conversations in public research datasets, so never paste private material there.
• Artificial Analysis. It says it runs its own tests on every model under identical conditions, and it also measures price, speed, and how soon the first token of a reply arrives. Its Intelligence Index, now version 5.0.1, blends coding, science, knowledge work, and general questions into one number. One number cannot tell you which model suits your own job.
• Epoch AI's Benchmarking Hub. It mixes tests Epoch ran itself with results from official leaderboards, and its reviews flag benchmarks with broken tasks. Read a review there before trusting a close race on any benchmark.
• Local Agent Bench. It tests local models through Ollama on tool-calling and multi-step tasks, using mostly 24 GB graphics cards. Its results page says testing was done in October 2026, and it does not name who runs it. It is not a paper. It ranks qwen3.8:27b second of 32 models that fit 24 GB, and speeds there compare only within one machine.

Test it yourself

Your own jobs are the best test, and a short trial takes an afternoon. Do it before you pay for a year of anything.

1. Pick three to five real jobs you do every week, such as summarizing a report or comparing three prices. Write each one as a request with the four parts from Prompting strong and weak models.
2. Use files that are not private. A cloud model reads everything you send, so start with public pages or made-up sample files. Move to private files only once you have picked a model whose privacy terms you accept.
3. Choose two candidate models from the same rows of the full table, such as the Google Gemini free key and the light hybrid. Set a spend limit at each paid provider first.
4. Run each job on both models in a fresh chat, with the same words each time.
5. Score each run on four questions. Did it finish the job? Did it use the right tools, and did the file it claims to have saved really exist? Did it stay on the job without wandering off? What did it cost, according to your provider's usage page?

If you are testing a local model, run this in a second terminal window while the agent works:

ollama ps

The PROCESSOR column shows whether the model fits in graphics memory, and the CONTEXT column shows how much of the conversation it can hold.

What you should see a short note for each job and each model that records finished or not, any tool mistakes, and the cost. For a local model, ollama ps should show a context of at least 64,000 tokens for agent work, as Run a local model with Ollama explains.

Pick the cheaper model if it finished every job you care about. Otherwise pick the stronger one, or use the cheaper model for routine steps and the stronger one for the hard ones.

Warning signs a model is too weak for agent work

These signs mean the model, its settings, or both are the problem, rather than your wording.

• It loops, repeating the same step or the same answer. On a local model, check the context length first, since Ollama starts at 4,096 tokens under 24 GiB of graphics memory, and the start of the job falls out of view.
• The tool call shows up as text in the chat instead of running. Hermes Agent's docs list this as a sign the model or its server does not support tool calling.
• It describes the work instead of doing it. Agent Zero's docs describe small local models that explain a command rather than calling the tool.
• It asks for a tool your agent does not have, often with a made-up name. Arena's Agent leaderboard counts this as tool hallucination, one of the signals it scores.
• It claims a step it never took, such as saying a file was saved when nothing is in the workspace. Hermes Agent's docs say small local models often confirm a save without making it.
• It forgets earlier instructions or ignores the safety rules halfway through a long job.
• It stops early with no error, or reports success on a job that is clearly unfinished.

One sign on one run can be bad luck. The same sign on most runs means the model is too weak for that job. Try the one-step style in Prompting strong and weak models, and if that still fails, move the job to a stronger model. When a weak model has already done damage, After it does something wrong walks you through recovery.

Trade-offs: safety versus capability

A cheaper or weaker model, or a stronger one. This is the model you choose in your agent's model settings, for the whole agent or for one kind of step.

• You gain: with a cheaper model, a smaller bill or none, and with a local model your prompts stay on your computer. With a stronger model, more jobs finish on the first try and you spend less time checking.
• You risk: a weaker model may loop, misuse tools, or claim work it never did, and you must check every result. A stronger cloud model costs more per step, and a stuck agent can turn into a surprise bill without a spend limit.
• Makes sense when: a cheaper or local model fits routine, single-step jobs that you check afterward, especially on private files. A stronger model fits long jobs, coding, and anything that runs with nobody watching.

For the three questions behind these choices, see Safety versus capability.

Sources

Checked 15 November 2026. Each page and what it said is recorded in the model evaluation research notes.

• Terminal-Bench site and leaderboard, for the current 4.1 default board and its columns.
• Terminal-Bench 2.1 announcement, for the 28 fixed tasks and the Claude Code and Terminus 2 results.
• Terminal-Bench 4.0 announcement, for breaking changes that need every test run again.
• Terminal-Bench README, for tasks in real terminal environments, each with a test script.
• SWE-bench documentation and SWE-bench leaderboards, for the codebase, the issue, the fix, and the 500 Verified problems.
• Scale AI, SWE-Bench Pro, 19 September 2025, for the 1,865 tasks, private code, and the drop from Verified scores.
• LiveCodeBench and its GitHub page, for contest problems and the v6 date range.
• τ-bench leaderboard, the τ²-bench paper, and its GitHub page, for simulated customers, policies, and shared steps.
• Berkeley Function Calling Leaderboard, for tool-calling accuracy and version 4.
• DeepSWE, for the 113 long tasks and one test program for every model.
• Epoch AI Benchmarking Hub, its SWE-bench Pro review, and its DeepSWE review, for setting changes that shift results and the two flawed ratings.
• Arena FAQ, Arena about page, and Agent Arena methodology, for anonymous votes, shared conversations, and the Agent leaderboard.
• Artificial Analysis methodology and its Intelligence Index page, for identical test conditions, speed and price, and version 5.0.1.
• Local Agent Bench results, for local tool-calling tests, the 24 GB ranking, and its caveats.
• Ollama library page for Qwen3.8 and Ollama search by tools, for capability badges and filters.
• Google's model page for gemini-4-hydrogen, for supported features and the 1,048,576-token window. The free-tier limits are in the free and cheap APIs chapter.
• Ollama context length, for the defaults by graphics memory, 64,000 tokens for agents, and the ollama ps columns.
• Hermes Agent, Providers, Hermes Agent, Memory, and Agent Zero, Local model tool use, for the tool-calling warning signs, as cited on Prompting strong and weak models.

Next

• Setting up a local model to test is on Run a local model with Ollama.
• How to word requests for strong and weak models is on Prompting strong and weak models.
• Back to the start: the home page.