Why Claude Max felt way less restrictive than Pro, even at 4x the usage

Between May 20 and June 1, I hit Claude's five-hour session limit 12 times on Pro. After moving to Max 5x, I hit it 3 times in the next 29 days.

I wanted to know what was actually happening, so I mined 10 weeks of my own logs (the JSONL files Claude Code keeps under ~/.claude/projects). I found 18 session-limit hits across 10 weeks and 148 sessions: 12 on Pro, 3 around the upgrade, and 3 on Max 5x.

A quick note: I was using subscriptions (like most people I assume), so these are not actual money I paid. I price the token fields in the logs at the API rates from the day they ran. That gives me one consistent unit for comparing very different models and token types, but it is still only a proxy for Anthropic's unpublished limit accounting.

On Pro, 11 of the 12 hits happened after roughly $7.60-$16.80 of visible API-equivalent usage in that five-hour window. One outlier reached $24.40. The median was about $11. The reset countdowns added up to about 32 hours. That is time remaining until reset when each hit occurred, not a claim that I spent all 32 hours staring at the timer.

In June 10, I started using Fable like every one else! I hit the limit three more times over the next 12 hours, and I upgraded. I treat those three hits as an ambiguous transition period rather than assigning them to either plan (I don't remember whether my limits reset or not).

On Max 5x, the three hits landed at $58, $76, and $93 of visible API-equivalent usage. Anthropic says Max 5x provides five times Pro's per-session capacity. My API-value medians are closer to 7x.

The part I did not expect was what happened to the distribution. My visible API-equivalent usage went from about $13 a day on Pro to $52 a day on Max, roughly 4x as much usage. With a plan offering 5x the per-session capacity, the naive expectation is that I should still hit the limit almost as often. I did not.

The reason is that my usage is not spread evenly. The observed Pro hit level sat around the 57th percentile of my reconstructed Pro windows, so an ordinary busy window could cross it. The Max hit level sat around the 94th percentile of my Max windows. I generally needed one of my rare, bursty five-hour stretches to get there. That is the distinction that stuck with me: the limit itself is a quota, but whether I keep feeling it is a percentile question. The bigger plan did not just raise the ceiling. It moved the ceiling from the middle of my personal usage distribution into its tail.

It also gives me a practical way to think about plan choice: decide how often I am willing to collide with the cap, then look at where the observed hit level falls in my own five-hour distribution. For example, if my tolerance is roughly 1 in 10 windows, I want that threshold at p90 or higher. Pro at p57 was too tight for me; Max at p94 is on the right side of that line.

The clearest example was one brainstorm that turned into a 19.6-hour build session: $205.87 of API-equivalent usage, 1,207 turns, and 41 subagents. Its busiest five-hour stretch was $84, which sits inside the $58-93 range I observed at the three Max hits.

Btw beyond hit limits, the expensive part was carrying a large conversation for so long. After an overnight break, two calls wrote about 473k tokens back into the prompt cache and produced 395 output tokens. At API rates, those two calls alone were about $19. Across the whole corpus I found only 9 compaction-sized context drops; 55 of the 63 sessions that grew past 150k tokens had no such drop at all. The logs do not tell me whether those drops were automatic or manual. My guess is roughly half and half, but the logs cannot prove it.

What I changed: I now start a fresh session when a follow-up does not need the old conversation, use `/compact` at real phase boundaries (it actually matter when it is done and from time to time I regret doing a `/compact` when I see how the session diverges after it), and send bulky exploratory reads to subagents so their full payload does not stay in the main thread.

One important note: this is not a clean A/B comparison. The Pro hits were concentrated in a crunch, while the Max month includes quieter days (and the Fable suspension). The hit counts are also a lower bound because I can only see hits recorded by Claude Code, while the limit is shared with other Claude products, mainly chat at claude.ai. I use chat infrequently, and usually only for quick questions, so I expect the missing usage to be marginal.

I would be interested in another set of usage-at-hit numbers, especially from Max 20x.

(Here is the tool I used (and built) if any one wants to try it )

3 comments

Why Claude Max felt way less restrictive than Pro, even at 4x the usage