Explore AI
How to reach a model
The model matters less than how you buy access to it. Eight routes, honestly compared.
Sign up, type, get answers. Every major lab runs one, and most have a usable free tier.
- Setup
- None
- Privacy
- Your text goes to the provider
- Good for
- Most people, most of the time.
Watch out You are renting access, not owning it. Terms, models and limits can change without notice, and your conversations are on their servers.
You get a key and call the model directly. Billing is per token — typically quoted per million, and output costs several times more than input.
- Setup
- An account, a card, a key
- Privacy
- Your text goes to the provider
- Good for
- Building something, or using a harness.
Watch out Every provider is a separate account, a separate bill, and a separate rate limit. Trying four models means four sign-ups.
One account and one key that reaches hundreds of models from many providers. You can switch model by changing a single string, and compare real costs side by side.
- Setup
- One account, one key
- Privacy
- Your text passes through a third party
- Good for
- Trying many models, or building something model-agnostic.
Watch out An extra party in the path, and pricing is set by the aggregator rather than the lab. Availability of a given model can change.
A flat monthly fee for generous or unlimited use of one provider's models, usually including their best ones and their apps.
- Setup
- None
- Privacy
- Your text goes to the provider
- Good for
- Heavy individual use of one lab's models.
Watch out Flat pricing rewards heavy use and punishes light use. If you use it twice a week, per-token is cheaper.
A program that drives a model in a loop — reading files, running commands, editing code, checking its own work. The model is the engine; the harness is the car.
- Setup
- Install and configure
- Privacy
- Depends entirely on the model you point it at
- Good for
- Coding, and any multi-step task.
Watch out This is where most of the practical value now sits, and it is the least understood part. The same model in a good harness substantially outperforms the same model in a chat box.
Download open weights and run them yourself with a runtime. No account, no per-token cost, no internet required.
- Setup
- Real
- Privacy
- Nothing leaves the machine
- Good for
- Private work, bulk work, and learning.
Watch out You are the infrastructure. Capability is capped by your hardware, and speed by your memory bandwidth.
Rent a machine with the memory you need, by the hour. Run any open model at full size, then switch it off.
- Setup
- Moderate
- Privacy
- Your data is on rented hardware
- Good for
- Models too large for your machine, occasionally.
Watch out You pay for idle time if you forget to stop it. Setup is real work the first time.
Several providers offer free access to capable models, sometimes including open-weight ones, with rate limits.
- Setup
- An account
- Privacy
- Your text goes to the provider, and is often used for training
- Good for
- Learning, and light use.
Watch out Free tiers usually mean your data may be used to train future models. That is the price. Never put anything confidential through one.
What most guides get wrong
They compare models. The more useful comparison is between routes — because two people using the same model through different routes have entirely different costs, privacy, and results.
The single biggest gap is the last one on the list: a harness. The same model inside a good harness will outperform the same model in a chat box by a wide margin, because the harness gives it tools, memory of the task, and the ability to check its own work. If you only change one thing about how you use AI, change that.
The combinations that make sense
- Curious, no budget → a free hosted tier, and a small local model for anything private
- Heavy individual use → one subscription, plus a local model for private work
- Building something → an aggregator while you experiment, then direct API keys once you have settled on a model
- Coding → a harness, pointed at whichever model is best that month
- Confidential work → local, or a rented GPU. Nothing else is honest here.