Codex vs Claude vs Gemini Usage Limits: Whose Quota Rules Can You Trust?

The question I’ve gotten most these past two weeks: which AI subscription gives you the most for your money, Codex, Gemini or Claude? Honest answer, I can’t tell you. Not because I’m lazy. None of the three publish what their top tiers actually include, so there’s nothing to divide the price by.

OpenAI even spells it out on its Codex pricing page: credit prices alone don’t determine included subscription usage. Okay then. So I asked a different question, one I can actually answer. Whose quota rules are more transparent, and whose are more stable?

My take, up front

On how clearly the rules are written, Claude and OpenAI edge out Google. On how much things moved around in the last two weeks, OpenAI wins by a mile. That’s not a compliment.

Nobody publishes absolute token numbers for Pro, Max or Ultra. OpenAI gives Plus and Business Standard a rough “messages per 5 hours” range. Plus on GPT-6.1 Sol is 15 to 160, a range you could park a truck in. Google and Anthropic only give multipliers like 5x and 20x relative to their Pro plan. Google is the only one that says the multiplier is in tokens: Ultra 5x is “5X worth of tokens compared to the $20/month Pro plan.”

OpenAI: trimming quota, handing out resets

This is the part I find genuinely interesting. OpenAI’s Pro tiers help article says new Pro 200 subscribers who don’t qualify as existing customers get a “lower usage allowance.” Price is still $200. If you had an active subscription between September 22 and September 29, you keep the old allowance until October 29. Same money, less quota if you showed up late.

Meanwhile Tibo, who runs Codex, launched a 28-day challenge: every day, either ship a noticeable improvement or do a full reset. I counted seven global resets between October 3 and October 18 (Oct 3, 7, 8, 11, 13, 15 and 18). The latest, Day 13, just says “Le limit reset.” We’re doing French puns now.

Some of the challenge days were real improvements, to be fair. On Day 2, auto-review stopped counting against usage, which OpenAI says can be 2 to 10% of a plan. Day 12 was about better cache hits through subagent context reuse and smarter routing, sold as more “effective usage.” Nice phrase. The bucket isn’t bigger, each scoop just costs a bit less, and you can’t verify either number yourself.

Now, pure speculation on my part. I have no evidence the two are connected and OpenAI has never said so. But if you reset everyone every two days or so, heavy users basically never hit the weekly cap, so they can’t feel how much the nominal allowance dropped. The challenge started October 5 and should wrap around November 2. Legacy allowances end October 29. Those dates sit awfully close, so I’m waiting until early November to see what the quota really looks like with no resets propping it up.

“My quota evaporated” is often the multipliers

There are plenty of posts complaining quota shrank. One user, @dovlan1990, says Codex is burning 3.8x faster and his weekly quota dropped from 33% to 0%. I can’t verify that, and he had Codex calculate the 3.8x itself, so the method is a mystery. His own numbers do show model requests going from 48,867 to 81,421 week over week. That’s 67% more. So, you know.

The official rules explain a lot of this. Fast mode costs 2.5x of included usage, Ultrafast on Pro 500 costs 8x, image generation runs about 3 to 5x. GPT-6 Astra’s credit rate is 5x that of GPT-6.1 Sol (250 vs 50 for input). And the one everybody forgets: ChatGPT Work and Codex share the same pool. Another user, @HermesWatcher, posted a support reply claiming ChatGPT automatically switches chats into Work mode, which then eats Codex quota. The shared pool is official. The auto-switching I couldn’t find documented anywhere, and the screenshot isn’t verified.

Add the fact that GPT-5.5 was pulled from ChatGPT-sign-in Codex on October 15, pushing people onto other (sometimes pricier) models. Multipliers don’t explain everything, though. A sudden 33%-to-zero drop doesn’t fit the rules, and I can’t rule out real backend changes. I’d just check whether Fast mode is on before rage-posting.

Google: “depends on capacity,” in writing

Google’s issue isn’t hiding things. It tells you upfront there’s no guarantee. The Antigravity plans doc says baseline rate limits are “primarily determined by available capacity” and usage limits are subject to modification. Pro refreshes every five hours until you hit a weekly limit. Ultra comes in 5x ($100) and 20x ($200, down from $250).

Gemini 4 Argon has its own rules on top. Varun Mohan says paid Pro gets a separate weekly Argon quota that doesn’t touch Flash. Trial or gifted Pro gets no Argon and no Claude Sonnet/Opus 5.5 at all. So when someone says the $20 Google AI Pro is the best deal around, ask whether theirs came free with a phone.

On October 17 Varun reset Argon and Gemini quotas for everyone. In the replies someone said Argon runs out too fast, and he said they’re working on getting more out of the model within quota. Same day, community posts claimed a wave of 403s on proxy accounts and bundled Pro plans flagged as trials. Unverified, scale unknown. And Theo posted that Gemini’s cache misses 27% of the time, making it 50% pricier than it should be; Varun replied that’s not what Google sees. Neither side showed reproducible numbers, so I’m just enjoying the show.

Claude: boring, in a good way

Anthropic did nothing to quotas in the last two weeks. As far as I could find, no official changes and no global resets, just a couple of Max users grumbling about limits to an audience of single-digit views. Boring. I like boring.

The Claude pricing page is fairly straight with you: every plan resets on a rolling five-hour window, paid plans add weekly limits on top, Claude and Claude Code draw from the same pool, there’s no fixed message count, and they may add weekly or monthly caps. Pro is $20, Max 5x is $100, Max 20x is $200, multipliers relative to Pro. “No fixed number” sounds vague, but at least they’re not pretending there is one.

So how should you pick?

Decide whether you want predictable or cheap. If you want predictable, Claude is the calmest option right now. OpenAI is raining resets, which feels great short term, but new Pro 200 buyers should know they’re on the reduced allowance and nobody knows what November looks like. Google suits people who can live with fluctuation. The docs literally say capacity decides.

Screenshot the plan page the day you subscribe. These pages change without notice, and if you ever end up arguing with support you’ll want receipts.

To stop burning quota: leave Fast off unless you need it, and definitely don’t flip on Ultrafast casually. Use the cheaper model for routine work and save Astra for genuinely hard problems. Chatting in ChatGPT spends your Codex quota too. Image generation is expensive, so maybe don’t use it as a meme machine. And Argon on Antigravity is weekly, so going all-in on Monday means watching from the sidelines by Thursday.

Last thing. Before you cancel over a “my quota vanished in a day” post, check what mode, what model and how many extra requests that person ran. I counted Tibo’s reset posts three times this week, and my conclusion is: see you in November.