Choosing an AI assistant without reading another benchmark

Benchmark tables change every few weeks and almost none of them measure the thing you do all day. A more useful approach is to work out what constrains you, then pick accordingly.

How long are your inputs?

If you routinely paste in long documents, context window is the constraint and everything else is secondary. If your prompts are a paragraph, it is irrelevant and you should ignore the marketing entirely.

Does it need to run unattended?

An assistant you supervise can be wrong occasionally, because you will catch it. An assistant running inside an automation cannot, because nobody is reading the output before it acts. Unattended work needs a more capable model and a validation step after it.

Where does the data go?

If you handle client data under contract, the question is not which model writes best but which provider your agreement permits. That often decides it before capability enters the conversation.

What does it cost at your volume?

• Occasional use: free tiers are genuinely sufficient for most people.
• Daily drafting: a single paid subscription, roughly the cost of lunch.
• Inside automations: API pricing per token, which behaves completely differently and can surprise you on long documents.

The honest summary

For most individual work, the frontier models are close enough that the one with the interface you prefer is the right answer. The differences that matter show up at the edges: very long inputs, unattended operation, and data handling obligations. Pick for your edge case and the general case takes care of itself.