I’ve paid Anthropic $60 over the last three months, and I kept a running note the whole time. Every session where Claude Pro saved me an hour went in one column. Every session where I hit a wall, waited out a timer, or rewrote a confident-but-wrong answer went in the other.
That note is the basis for this Claude Pro performance review. Not a spec sheet read back to you, not a launch-day hot take. Three months of real work: client copy, Python scripts, a 92-page PDF I had to summarise, and one very long argument with a spreadsheet.
Here’s what held up and what didn’t.
What $20 a month actually buys
The plan sits above the free tier and below the $100 and $200 Max tiers. The headline features are straightforward:
- Roughly five times the free tier’s usage, metered in a rolling five-hour window, with a weekly cap on top
- Access to the newest models, including limited Opus usage alongside the current Sonnet
- Projects, which hold persistent context for a workspace so you’re not re-explaining your brand voice or codebase every morning
- Artifacts, the side panel that renders code, documents and diagrams live as you iterate
- Priority access when traffic spikes, plus the desktop app and browser extension
Nothing there is exotic. The value lives entirely in how much of it you can use before the meter stops you.
Speed and reliability across three months
For most of the period, responses started streaming in under two seconds and a mid-length answer landed in eight to fifteen. Long document questions took noticeably longer, sometimes 40 seconds on a 60-page PDF, which is fair enough when you consider what’s happening under the hood.
I logged four outages that lasted more than ten minutes. Two were brief, one ate an afternoon, and one hit during a client deadline, which is exactly when you notice how much of your workflow has quietly become dependent on one company’s uptime.
Degradation is the subtler issue. On busy weekdays the model sometimes felt thinner. Shorter answers, less willingness to push back. Hard to prove, easy to feel across hundreds of prompts.
Where Claude Pro genuinely earns its keep
Long documents
This is the clearest gap between Claude and everything else I’ve tested. I fed it a 92-page regulatory PDF and asked for the five clauses that would affect a small software company. It came back with five, correctly ranked, each quoted with a page reference I could check in under a minute. Two other assistants I ran the same test on either skimmed the first twenty pages or invented a clause number.
Writing that doesn’t read like a chatbot wrote it
Ask for a first draft and you get something serviceable. Ask it to cut 200 words from your own draft without flattening the voice, and that’s where the model separates itself. Over three months I stopped using it to generate copy and started using it to compress mine. The editing passes are sharper than the drafting.
Code review
I use it mostly as a second pair of eyes rather than an author. Point it at a 300-line function and it will flag the off-by-one, the unhandled null, and the retry loop that never backs off. It’s also good at explaining why a test is flaky. What it won’t do is guarantee the fix is correct, and the industry is now building entire companies around that gap. Qodo’s $70M raise for code verification exists precisely because AI-written code needs checking by something other than the thing that wrote it.
The friction points nobody puts in the demo
The limit wall arrives without warning
You don’t get a fuel gauge. You get a message telling you that you’ve reached your limit and when it resets. The five-hour window means a heavy morning can leave you locked out for the afternoon, and the weekly cap introduced in 2025 makes Friday a genuine gamble if you’ve been leaning on it since Monday. I now batch my heavy PDF work into the morning and keep a fallback tab open.
Confidence and accuracy are not the same thing
It still fabricates. Fewer times than a year ago, and usually in ways you can catch, but a made-up citation delivered in fluent prose is more dangerous than an obvious error. My rule is blunt: anything factual gets verified before it goes near a client.
Memory is better, not solved
Projects help enormously, but cross-conversation memory is patchy. It will forget a preference you set last Tuesday and then apply it faithfully a week later, which is somehow more maddening than consistent forgetting.
How it stacks up against the rest of a $20 budget
Most people comparing plans are really asking whether to move the $20 from one subscription to another. The honest answer is that ChatGPT’s free vs paid split is worth checking before you decide, because the two plans fail in different places. ChatGPT is broader: image generation, voice, connectors, a bigger ecosystem. Claude is deeper on long text and code reasoning.
I’ve also spent six months paying for eleven tools at once, and the pattern that emerged was consistent: roughly three carried 90% of the actual work. The rest were subscriptions I kept out of habit. You can read the full breakdown in the honest performance review of 11 AI tools, but the lesson applies here. Claude Pro earns a slot if long documents or code review sit at the centre of your week. If you mostly want help drafting emails and summarising articles, the free tier plus a second opinion will cover you.
Who should pay, and who shouldn’t
Pay for it if you handle documents over 30 pages, review code daily, or do editing work where voice matters. Pay for it if you bill your time and the $20 disappears inside fifteen minutes of recovered work.
Skip it if your usage is bursty and light, if you need image generation in the same tab, or if you’re still working out whether AI fits your process at all. There’s no shame in the free tier. The quality gap between free and paid on a short question is close to zero.
The routine that made it worth keeping
What changed my results wasn’t a clever prompt. It was a structure. Each morning I open a Project with three files attached: a style guide, a folder of past work, and a running list of things the model got wrong last week. Then I work in short bursts rather than one marathon conversation, because long threads degrade and a fresh context with the right files attached beats a 60-message history every time.
I also stopped asking it to be creative first. I draft badly myself, hand over the messy version, and ask for compression plus a structural read. That single change removed most of the “this sounds like AI” problem I was fighting in month one, and it costs nothing to copy.
Three months and $60 later, the note has far more entries in the saved-time column than the wasted-time one. It’s not flawless, it hits a ceiling sooner than I’d like, and I’ve learned to keep a backup open in another tab. But when I tally the hours back, the subscription has paid for itself before the first Tuesday of every month so far.