Blog ·
OpenClaw, Hermes Agent, Claude Code, Codex and OpenCode don't think for themselves. Each is a harness: the program that gives a model tools, memory and a loop to work in. The thinking is done by the model you connect, from Anthropic, OpenAI, Google or another provider.
That makes the model the biggest single factor in how well your agent works, and in what it costs. This guide shows how to choose one without relying on benchmark charts.
Harness, model, computer: three separate choices
- The harness decides how the agent works: chat apps, memory, skills, schedules. See OpenClaw vs Hermes Agent or Claude Code vs Codex vs OpenCode.
- The model decides how well it reasons, follows instructions and uses tools, and most of the monthly bill.
- The computer decides whether it stays on, what it can reach and whether you can watch it. See where should your AI agent run?.
You can change each one without the others. Most harnesses let you switch models in a minute, so treat your first choice as a starting point, not a commitment.
What makes a model good for agents
A model that writes great essays can still be a poor agent. For agent work, look at:
- Tool calling. Every agent works by calling tools: open this page, run this command, save this file. A model that formats calls badly, or invents results instead of waiting for them, fails no matter how clever it is. This matters most.
- Following instructions over a long task. Can it still remember on step 40 that you said “don't send anything”? Agents run for many steps, and some models drift.
- Knowing when to stop. Good agent models finish, or stop to ask, instead of looping.
- Context length. How much it can read at once. Long context helps with big documents and codebases, but filling it costs money on every step.
- Vision. Agents that look at a screen or screenshots need a model that reads images well.
- Speed. A fast model makes a chat assistant feel responsive. For overnight jobs it matters less.
- Price. You pay per token, for everything the model reads and writes. Agents read far more than they write.
- Privacy. What the provider keeps, for how long, and whether it trains on your data.
The providers at a glance
These are the providers most agent harnesses support, and the ones you can pick in the Cubicles New agent form:
| Provider | Known for | Worth a look if you… |
|---|---|---|
| Anthropic (Claude) | Strong tool use, coding and long multi-step tasks | Want a dependable default, especially for coding agents |
| OpenAI (GPT) | Broad capability, wide tool support | Already use OpenAI, or run Codex |
| Google (Gemini) | Very long context and strong with images and video | Work with large documents or lots of media |
| OpenRouter | One key for hundreds of models from many providers | Want to compare or switch models without new accounts |
| DeepSeek | Low prices and open-weight models | Run high volumes on a tight budget |
| GLM (Z.ai) | Open-weight models that are good at coding and agent work, at low prices | Want a capable budget option |
| Mistral | A European provider with open-weight options | Prefer a provider based in Europe |
| Venice | Privacy first: says it doesn't store prompts or responses | Care most about keeping conversations private |
| Custom endpoint | Any OpenAI-compatible server, including your own | Host your own model, or use a provider not listed |
Model line-ups change every few months, so this compares providers rather than individual models. It reflects them as of November 2026. Check each provider's own pages for current models, prices and data policies.
Big model, small model: use both
The most expensive model isn't the best choice for every job. Frontier models are worth it for hard, open-ended work. Routine jobs, like checking an inbox, summarising a page or confirming a site hasn't changed, often run just as well on a smaller model at a fraction of the price.
- Split by agent. Run your everyday assistant on a mid-sized model and a separate agent with a top model for hard research or coding. See how to run multiple AI agents.
- Split by task. Some harnesses, OpenCode for example, let you pick a model per session, so you can plan with a strong model and do routine edits with a cheaper one.
- Use a router. OpenRouter lets you change models by changing one line, which makes this kind of experiment cheap.
How to test a model on your own work
Public benchmarks measure general ability. You care about your tasks. An afternoon of testing beats a week of reading leaderboards:
- Pick five real tasks you'd give the agent, from easy to hard, with a clear finished state.
- Run each with two or three models in the same harness, with the same instructions.
- Score each run. Did it finish? Was the result right? Did it stop to ask at sensible moments? Did it claim to have done something it didn't?
- Check the cost in each provider's usage dashboard.
- Pick the cheapest model that passes, and keep a stronger one for when it doesn't.
Keeping the model bill under control
- Set a monthly spending limit with your provider before you start. A stuck loop then stops at a number you chose.
- Use a separate API key per agent, so you can see which one spends what and revoke one without breaking the others.
- Trim what goes in on every step. Long standing instructions and big memory files are re-read every call.
- Make scheduled checks cheap. A check every few minutes adds up fast. Run it less often, or on a small model.
For a full breakdown, read how much it costs to run an AI agent.
Privacy: where your data goes
Everything your agent reads, from emails to web pages to files, is sent to the model provider. Before you connect an agent to anything sensitive:
- Read the provider's API data policy. API use is often treated differently from consumer chat apps, with shorter retention and no training on your data by default, but check.
- Look for business options such as zero data retention if you handle client or customer data.
- Consider a privacy-focused provider like Venice, or your own endpoint, for the most sensitive work.
- Send less. An agent that only reads one mailbox folder shares far less than one with your whole inbox.
Open-weight and local models
Open-weight models, from DeepSeek, Z.ai, Mistral and others, can be run by any host or on your own hardware. They're often much cheaper through hosted providers, and running them yourself keeps data in-house.
The catch with running locally is hardware: a capable model needs a strong GPU or a lot of fast memory, which is a different machine from the one your agent works on. A common setup is to run the model on one machine and point the agent at it through a custom, OpenAI-compatible endpoint.
Setting the model on Cubicles
On Cubicles, the New agent form asks for three things: a provider, a model id (for example claude-sonnet-5-5) and your API key. Pick Custom to use any OpenAI-compatible endpoint by giving its URL. Venice's models are listed for you to search as you type.
Your key is stored encrypted and sent only to your agent's computer. You can change the provider, model or key later from the agent's page, and Cubicles sets the agent up again for you. Your first computer is free, so it's a cheap place to run the five-task test above. Sign up to try it.
Frequently asked questions
What is the best model for OpenClaw?
There's no single best. Pick a model that is reliable at tool calls and follows instructions over long tasks. Many people start with a recent Claude or GPT model, then move routine jobs to a cheaper one once they know what works.
Can I switch my agent's model later?
Yes. OpenClaw, Hermes Agent, OpenCode and most other harnesses let you change the provider and model at any time. On Cubicles you change it in the agent's settings.
Is a cheaper model good enough for an AI agent?
Often, for routine work like summaries, inbox sorting and simple checks. Test it on your own tasks. Keep a stronger model for open-ended research, coding and anything where mistakes are costly.
Do I need a GPU to run an AI agent?
No. If the model runs at a provider like Anthropic, OpenAI or Google, the agent's computer only runs the browser, terminal and tools. You only need a GPU to run the model itself on your own hardware.
What is OpenRouter and should I use it?
OpenRouter is a service that gives you one API key and one bill for hundreds of models from many providers. It's handy for comparing and switching models. If you settle on one provider, going direct can be simpler.
Can I use my ChatGPT or Claude subscription instead of an API key?
Some coding agents can: Codex signs in with a ChatGPT plan and Claude Code with a Claude plan. Personal agents like OpenClaw and Hermes usually use an API key, and the Cubicles New agent form asks for one.