Claude review: The one we open when the work gets hard

The first time we leaned on Claude for a problem that actually mattered, it asked us a question before it touched the keyboard. The job was a database migration scattered across half a dozen files, with subtle timezone quirks nobody on the team remembered. "Before I propose changes," it said, "is the `created_at` column storing UTC or local time? It affects how I'd write the backfill." That single sentence caught a bug that would have shipped silently. The migration went out clean the next morning.

That was a year ago. We've gone back and forth with ChatGPT since the second generation of Claude landed, and Claude is the one we open now when we know the work is going to be hard. The question worth answering is whether it's worth the same $20 a month as everything else, and what you give up by picking it as your daily driver.

The long context window is the feature

Most AI assistants will let you paste a few thousand words and call it context. Claude lets you paste a full repository, a 300-page PDF, or six months of design docs and still reason coherently across all of it. The million-token window on Opus is the headline. The practical version is that we almost never bother building RAG plumbing for one-off analysis anymore. Drop everything in, ask the question, get an answer that cites the right page.

It shows up in places you wouldn't expect. Code reviews on an unfamiliar repo are more useful when the model has read the whole thing. Long contract analysis catches the clause on page 47 that contradicts the clause on page 12. There is no summarize step, so there is no summarize step where things go wrong.

It says "I don't know"

This sounds like a small thing. It isn't. ChatGPT, even today, will confidently invent an answer for a question it has no business answering. Claude is more likely to stop and say "I don't have reliable information about this specific version" or "I'm guessing here because the docs you shared don't cover this case." We triple-check obscure factual claims less often as a result.

It still hallucinates library APIs and confuses versions on niche packages. But the rate is lower, and it flags its own uncertainty more often when it happens.

Claude Code is the reason to subscribe if you write software

The CLI tool runs agentically in your terminal. It can edit files, run tests, navigate a repo, and recover from its own mistakes mid-task. We've handed it a real feature - schema migration, server route, frontend wiring, tests - and watched it ship the thing over an afternoon with light supervision.

Cursor is still the better moment-to-moment editor when you're writing code yourself. Claude Code is a different shape entirely. You hand it a real task, walk away for twenty minutes, and there's a working PR when you get back. Both live on our laptops.

Where it loses to ChatGPT

The consumer surface is small. No native image generation. No voice mode worth using. No GPT store, no equivalent to Sora, no real-time browsing experience that keeps up with what OpenAI ships every few months. If you want one AI assistant to do everything in your day, generate an image for a deck, talk to you while you cook, answer trivia by voice, ChatGPT is the right pick.

Anthropic seems okay with this. The product is for people who care about answer quality more than feature breadth, which is a sharp pitch for a small audience.

The free tier is too thin to evaluate

You won't get a feel for Claude on the free plan. The message limits are tight enough that one real debugging session burns through the day. The honest recommendation: subscribe to Pro for a month if you want to try it properly. Most people who do that stay on Pro, but the on-ramp is steeper than ChatGPT's, where the free tier is genuinely usable.

Two features that get no marketing attention

Projects keep a set of files and instructions attached to a conversation thread. Drop in your style guide, the repo conventions doc, a draft you've been iterating on. Every chat in that project inherits the context. We use it for client work especially, where the same constraints come up over and over.

Artifacts opens a side pane for any code or document Claude generates. You can edit it in place, run it for some languages, and have it stay out of chat scrollback. After a few weeks of using it, going back to a chat UI without it feels primitive.

A note on the cautious dial

Claude is sometimes too careful. It will occasionally refuse a benign ask, or wrap a clearly fine request in disclaimers nobody asked for. This is much better than it was in the Claude 2 era, when the model would lecture you about a coffee recipe. But it still surfaces in places where it really shouldn't, and if you've been burned by overcautious AI before, you should know it isn't fully gone.

So should you switch?

If your work is research, writing, coding, or analysis, try Claude as your default for a month. The depth of the answers is the differentiator, and you won't feel it until you've lived with it for a while.

If your work is quick questions, image generation, or a stack of custom GPTs you actually rely on, stay on ChatGPT and keep a Claude tab open for the hard ones.

We do both, and on a forced choice it would be Claude.