Cheapest LLM for AI agents

The cheapest model is only a bargain if it can do the job. Below we rank models by what a typical agent workload would cost per month, not just by the price per token, and explain how to test whether a cheap model is good enough.

Ranking for a typical agent

Workload: 500 tasks/day, 6 steps per task, 3,000 input tokens (half of them cached where the model supports caching) and 400 output tokens per step. Only models with at least a 128K-token context window and no open price warnings are included.

The very cheapest entries are small models (1B–14B parameters). They are great for simple steps like sorting or extracting, but usually too weak to run a whole agent on their own.

Note on DeepSeek: its official prices change by time of day. We list the peak price; off-peak hours are half price (DeepSeek pricing).

Cheap is not always cheaper

• More steps: a weaker model may need extra steps or retries, which can cancel out the lower price.
• Output-heavy work: if your agent writes long answers, compare the output price first.
• Tool use: check the model reliably calls tools (function calling). Small models fail here more often.

A simple way to choose

1. Pick 20 real tasks your agent handles.
2. Run them on your current model and on 2–3 cheaper ones.
3. Count how many each gets right and how many steps it needs.
4. Put those step counts into the calculator to see the true monthly cost.

Many teams end up mixing models: a cheap model for easy steps and a stronger one only where needed. See How to reduce LLM token costs and side-by-side model comparisons.

Prices are list prices per million tokens from our pricing table (last refresh 2026-11-01); check providers before you commit. This guide was written with help from AI tools and checked by a human.