Choosing an AI Model: OpenAI vs Anthropic vs Local LLMs

The biggest practical advantage of a model-agnostic assistant like OpenClaw is that this decision is reversible. You bring your own key, you can change your mind in a minute, and you can use different models for different jobs. This guide is about how to make the choice well rather than which brand to pick, because the brands move faster than any page can track.

Start with the criterion that actually breaks things

Before capability, price or context length, check one thing: does the model support tool calling? An assistant acts by having the model emit a structured call to a tool, and not every model can do that. Point OpenClaw at one that cannot and you get a fluent conversational partner that describes what it is about to do and then does nothing at all. Nothing errors, which is why this is the single most misdiagnosed problem in self-hosted assistants and why it has its own section in why an assistant stops responding.

A softer version matters too. Some models support tool calling but do it unreliably, producing malformed calls that fail and get retried. A model that retries is not cheaper than a stronger one that gets it right first time, whatever the price per token says.

The four dimensions worth comparing

• Reasoning quality. How well it handles a task with several steps and some ambiguity. This is what you notice on hard work and never notice on easy work.
• Context window. How much it can hold at once. Important for long documents, and more expensive than it looks, because a long conversation re-sends its own history on every turn.
• Latency. How long you wait. On a chat assistant this matters more than people admit: an assistant you wait thirty seconds for is one you stop using.
• Privacy posture. Where the content goes and what the provider's retention and training policy says about it. Read the policy for the tier you are actually on, since it frequently differs between consumer and API access.

Notice that price is not on that list. It matters, but it is downstream: a model that needs fewer retries and shorter prompts to do your job well can be cheaper at a higher headline rate.

The hosted providers

OpenAI

A strong general default with reliable tool use, broad ecosystem support and a wide range from small fast models to large capable ones. If you have no specific reason to choose otherwise, starting here and changing later costs you nothing. You pay OpenAI directly with your own key.

Anthropic

Claude models are known for careful, well-structured reasoning, long context and close instruction-following. They suit an assistant that does a lot of reading and writing, and workflows where you want the model to do exactly what was asked rather than something adjacent to it.

Google

Gemini's headline strength is very large context, which suits work dominated by long documents, transcripts or codebases that need to be reasoned over in a single pass rather than chunked.

Aggregators

Services such as OpenRouter put many providers behind one key and one API, which is genuinely useful when you want to compare models or route different tasks to different ones without managing several accounts. The trade is another party in the path and, on free tiers, real limits: free model variants are capped per minute and per day, and those caps bite an assistant harder than a chatbot because an agent makes several requests per task.

Local models, realistically

Running a model on your own instance through Ollama is the strongest privacy guarantee available, because nothing is sent anywhere. It is worth being clear about the trade rather than romantic about it.

The arithmetic first. A 7B model quantised to 4 bits is roughly 4 GB of weights before any conversation context and before OpenClaw itself, so the entry tier cannot hold one and 8 GB is the realistic floor. Larger models scale accordingly, and without a GPU, inference runs on CPU and is slow enough that most people who try it for everyday use go back to a hosted model. See the server requirements for how this changes sizing.

The honest framing is that local inference is right for specific tasks rather than as a default: content you are not willing to send anywhere, work that has to keep functioning without internet, or a high-volume workload where per-token pricing is the constraint. For everything else a hosted model with a self-hosted assistant already keeps your conversation history, credentials and integrations on your own server, which is most of what people mean by privacy.

Quick comparison

How to actually decide

Comparisons written by other people are a starting point and a poor substitute for two hours of your own testing. The method that works:

1. Pick the task you most want the assistant to do well. Not a benchmark, a real one.
2. Run it against two candidate models with the same prompt.
3. Compare on the three things you will live with: was the answer right, how long did you wait, and what did the provider dashboard say it consumed.
4. Keep the winner as your default and leave the other configured, because the next task may swing the other way.

This takes an afternoon and settles the question far better than any table, including the one above.

You do not have to choose only one

Because OpenClaw is model-agnostic, the real answer is usually a split: a capable hosted model as the default, something cheap and fast for routine reformatting, and a local model for the handful of things you would rather not send anywhere. Switching costs nothing structural, because the model is configuration rather than architecture. Your integrations, history and setup are untouched by the change.

Ready to try it? Follow the self-host OpenClaw guide, check the server requirements if a local model is on your list, or deploy from the homepage.