Picked from the 11 tools on this site by what they're actually used for — not by benchmark scores.
| Tool | What it is | Free tier | Context |
|---|---|---|---|
| Claude | Anthropic's assistant built for long documents and code edits | Yes | Up to 1M tokens, varying with the selected model (checked 2026-10-29) |
| Kimi | Moonshot's assistant built around 1M-token reading | Yes | 1,048,576 tokens (K3 series); 262,144 tokens on the K2 series (checked 2026-10-29) |
| Perplexity | Answer engine that cites sources on every reply | Yes | — |
Research is the use case where citations either exist or the answer is worthless. A tool that gives you a confident summary with no links is not saving you work — it's moving the work to verification later, usually after you've already repeated the claim. Prefer tools that show the source next to the sentence.
Then check what happens when the answer isn't on the web: ask something that requires reading a document you supply. The good ones quote the paragraph and the page; the rest quietly fall back to their training data, which is exactly the failure mode you can't see.
Start with the one that already has a free tier covering your volume, then check the context column: research work usually breaks on input size before it breaks on model quality. If two of them look equal, pick the one whose paid tier you'd actually renew, because that's the one you'll keep using after the novelty wears off.
Everything else — ChatGPT, Gemini, Grok, DeepSeek — is on the main comparison table; they show up here only if their own documentation frames them that way.