Search the phrase and you get a dozen pages that open with a table and close with a winner. Most of them blend two kinds of number: figures a vendor published, and figures the review site produced in its own test.
This page is narrower. Every Claude figure below traces back to a page Anthropic publishes, and the ChatGPT column is empty, because OpenAI’s pricing and model pages returned HTTP 403 from the network this was written on. The prompt that produced the table sits further down.
Why most comparison tables cannot survive a source check
Comparison pages rank by answering fast, and a fast answer wants a verdict. So the writer runs a test and puts the score in the table.
That work is honest and still not transferable. The top-ranking page for this phrase says it ran a ten-question benchmark plus a long-document pass; its eight out of ten is that test’s result, not a property of the product. Two sites can run different tests and rank the same pair in opposite orders.
Only one side of this table has a publisher behind it
Anthropic publishes the terms in more detail than most buyers expect. Free sits at $0. Pro is $17 per month on annual terms with $200 billed up front, or $20 billed monthly. Max starts at $100 per month with five times or twenty times the usage of Pro.
One wrinkle before you quote that. Anthropic’s help centre presents the same tiers a second way: Pro at $20 per month or $200 per year, Max 5x at $100, Max 20x at $200. Two official pages, two presentations, so quote both and check the checkout page. Both note that prices exclude applicable tax and that plans can change at Anthropic’s discretion. The full plan breakdown takes the tiers one at a time.
The model line-up is a matrix rather than a sentence, and one row does the work. Claude Fable 5.1 reads “No” on Free, “Usage credits” on Pro, and “50% of weekly limits” on both Max tiers. Claude Opus 5 is “No” on Free and “Yes” above it. That is a billing difference, not a scoring one.
Side by side, with the unsourced cells left empty
| What is compared | Claude, from Anthropic’s published pages | ChatGPT |
|---|---|---|
| Entry paid tier | Pro $17 monthly on annual terms, $200 up front, $20 monthly | the vendor has not published this |
| Higher tiers | Max from $100 per month, 5x or 20x Pro usage | the vendor has not published this |
| Current models | Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5 | the vendor has not published this |
| Context window | Newest paid models up to 1M tokens, others 500K or 200K | the vendor has not published this |
| Usage metering | Rolling five-hour sessions, weekly limits, no fixed message count | the vendor has not published this |
| Claude Code | Included in all paid plans, same usage pool | the vendor has not published this |
| Image and video | Not listed among Anthropic’s published plan capabilities | the vendor has not published this |
One column readable, one deliberately blank. Nothing on the right claims ChatGPT lacks a feature; every cell says no figure appeared in a page this article could open, so none is quoted. A comparison with one honest column beats a symmetrical one resting on a test nobody can inspect.
The terminal changes where coding work happens
Anthropic’s sentence is “Claude Code is included in all paid plans”, and the same page adds that it “shares the same usage limits as the rest of your plan, so your work in the terminal and your chats draw from one pool”. Read that as a budget rule: a long refactor and a morning of questions spend one allowance. Moving the work to the API changes the billing model completely, which the API getting-started guide covers.
The context window has an official sentence, and throughput does not
The familiar claim that Claude holds a 200k window is out of date. Anthropic writes that “the context window size depends on which model you’re using”, that the newest models on paid plans “support up to a 1M token context window”, and that others support 500K or 200K. It also separates usage limits, which cap how much you use across conversations, from length limits, which cap one conversation’s depth. How fast any of it answers stays unpublished.
Worked example: the prompt that produced this table
The prompt, run in a fresh conversation with the table above pasted underneath it.
Audit this comparison table row by row.
For each row do three things:
1. Name the page that supports the figure on the Claude side, with its URL.
2. Name the page that supports the figure on the ChatGPT side, with its URL.
3. If you cannot name one, write "unsourced" in that cell.
Do not estimate. Do not count a review site, a leaderboard or a blog
post as a source. Only the two vendors' own domains count.
Finish with a list of every row that came back unsourced.What it produced was not a verdict. Seven rows survived on the Claude side, each pointing at a plan card, a help-centre article or an API price table, and every ChatGPT row came back marked unsourced. That closing list became the section below almost word for word. The prompt never picks an assistant for you; it shows where your table has no floor under it.
What neither publisher will hand you
Anthropic publishes no fixed message count for any tier, so a ranking by messages per day sorts a number the vendor does not publish. The weekly limits exist without a figure. No conversion rate connects usage credits to tokens, so the cost of reaching the top model from Pro cannot be worked out in advance, and the annual discount has no percentage. The free plan is described the same way: session based, resetting every five hours, no number.
The OpenAI side is empty for a blunter reason. Both openai.com and platform.openai.com answered 403 to every request, twice. That is a fact about a network, not about a product.
Frequently asked questions
So which one is better? The question hides a task. The table above settles published terms, not output quality, and quality needs one test run on both.
Why leave a column empty instead of filling it? An empty cell with a stated reason can be checked in a minute; a filled cell from an unnamed test cannot be checked at all. If your network reaches the OpenAI pages, run the prompt above.
Where do I start with Claude itself? The topic hub collects the guides, and the plan breakdown linked near the top is the natural next read.
A comparison you cannot audit is a review you have to trust.