How the picks are made
The finder is a set of fixed rules applied to the same daily model data as the pricing explorer. No language model is involved, so the same answers always give the same picks, and you can see why.
- Hard filters. Models are dropped if they lack a must-have (tool calling, structured output, vision, reasoning), have too small a context window, are scheduled for retirement, are more than 18 months old, or do not match where you want to run them. For local inference, a model has to fit your GPU memory, estimated at about 4.5 bits per weight plus 2 GB of headroom. Models whose parameter count is not published cannot be checked and are left out.
- Quality. The Artificial Analysis index that fits the job: the coding index for coding agents, the agentic index for support and automation, the intelligence index otherwise. A model must score within a floor of the best candidate to be recommended. The floor is 55% of the best when cost comes first and 95% when quality does.
- Cost. Monthly cost at your request volume and token sizes, compared on a log scale so a 10× difference counts the same at any price level. Batch latency uses batch-API rates where they exist.
- Your slider sets the weight between quality and cost. Real-time latency penalizes models that always reason. Latency is inferred from the model type, not measured.
You get three picks. Recommended is the top score within your budget. Budget is the cheapest model scoring at least 85% of the recommendation. Higher capability is the best-scoring model regardless of budget. Open Top 10 and how they scored to see the quality and cost points behind each.
What it cannot tell you
Benchmarks measure benchmarks. A model that ranks third on an aggregate index can be first on your task, and the reverse. Treat the picks as a shortlist: run the top two or three on fifty real examples from your workload, score the outputs, and let that decide. Backtesting Lies to You covers how to keep that eval honest.
Running models locally
Choose My own hardware and set your GPU memory to see open-weight models that fit. Mixture-of-experts models need memory for every expert even though only a few are active per token. The finder shows the active and total parameter counts so you can judge speed. If a model almost fits, offloading experts to system RAM is possible but slow; I tested this in MoE expert offloading: 125B on a 12GB GPU.
Related
- Buy, Build, or Wait: evaluating AI capability as a CTO
- How to Evaluate an AI Tool for Your Business
- AI Agent Cost Simulator: price the pick across a whole agent loop.
FAQ
What is the best LLM for AI agents?
There is no single answer, which is the point of the tool. For agents, tool calling and reliable structured output are non-negotiable, and the agentic and coding indexes predict multi-step performance better than general chat benchmarks. Within that set, the right pick depends on how many calls you make: a model that is 20% better and 10× the price is worth it for a few high-stakes calls a day and ruinous for a million.
What is the best LLM for coding?
Select Coding agent and the finder ranks by the Artificial Analysis coding index. Move the slider to quality first to see the strongest model regardless of price, or toward cost first for high-volume work such as code review on every commit.
What is the best local LLM for my GPU?
Choose My own hardware and your GPU memory. The finder lists open-weight models that fit at 4-bit quantization and ranks them by benchmark score. Larger context windows need more memory for the KV cache than the 2 GB headroom assumed here, so leave margin if you need long prompts.
Where do the benchmark scores come from?
From the Artificial Analysis intelligence, coding and agentic indexes as published in OpenRouter’s public model list, refreshed daily. Models without a published score are flagged and only recommended when nothing scored qualifies.
Does the finder use AI to make recommendations?
No. Prices, capabilities and scores come from the published data, and the ranking is a transparent formula. That keeps it fast, free to run, and consistent between visits.