Which AI Model Should I Use?

An LLM selector is a tool that recommends AI models for a specific job. BenchLM’s selector, also called an LLM picker or AI model finder, reads the task, budget, input size, and where the model has to run, then shortlists models from public benchmark evidence and API prices, with every assumed answer marked.

It depends on the job and on whether you mean the chat app or the model behind it. The selector compares models on benchmark evidence and API prices. Subscription plans, message limits, and app features such as voice or file uploads differ by plan, so check those with the provider before you pay.

No single model leads every job. The overall leaderboard names the current leader across categories, and it changes with each release. For a specific job, the selector answers a narrower question: which models have evidence for that work and still pass your budget, input size, and deployment limits.

Describe the job, or pick it from the list. The page shows what it read, fills every remaining question with a stated default, and computes a shortlist at once from the public dataset: a relevant category estimate or observed task result, filtered for the stated budget, context, and deployment needs. Each answer you change reruns the shortlist and states what moved. There is no hidden weighted fit score.

Start with the job, not the leaderboard. Name the kind of work, then say what would rule a model out: budget, input size, or data that cannot leave your machine. The selector filters the models with evidence by those limits and orders the rest on evidence for that job. Then test the top two on three of your own examples.

A decision model fills in the same multiple-choice form you see on the page, and nothing else. It recognizes requirements and models you mention; it never sees a score or chooses the shortlist. Every answer it fills is marked as coming from your description, and you can change any of them. The shortlist is computed on this page from the dataset, exactly as it would be from clicks. BenchLM keeps the description, including drafts read as you type, for 180 days without your IP address or account, to see which jobs people need.

A shortlist is provisional while any answer under it is an assumed default rather than something you said or chose, or while the job was read with low confidence. The assumed answers are marked. Confirm or change them and the label goes away.

Choose Build or fix software, then the kind of work you do. Coding-agent results depend on the model, editor or agent harness, and effort setting. Compare the suggested models on representative repository tasks before switching.

Choose that it must stay on your own hardware, and the shortlist is limited to open-weight models. The selector does not check your hardware. The VRAM calculator takes your GPU memory, card count, and quantization and shows which local models fit; the local LLM ranking compares the open-weight candidates.

A free assistant plan, an open-weight model, and free hosting are different things. For local deployment, the selector limits candidates to open weights. For assistant plans, model access and message limits must be checked with the provider; API prices are not subscription prices.

A task may lack suitable benchmark evidence, or no measured candidate may satisfy your constraints. Missing prices cannot pass a price limit. The selector keeps these gaps visible instead of silently relaxing your requirements.

Name it in the description. The page then states where that model stands for the job: on the shortlist, filtered out by a constraint you stated, below the cut with its score, or without evidence for this job.

Yes. The selector is free and needs no account; anonymous use has a daily limit on description readings. The shortlist is computed on the page from the public dataset, and a shared link carries your answers but not your description, so it reopens the same answers against current data.

Every answer is written into the page address, so the link reopens the same shortlist against the current dataset. The same answers work as query parameters on the /api/tools/recommend endpoint for scripts and agents.