Why session limits exist and how compute changes them

Two limits control different things

The weekly limit caps total included usage over seven days. The five-hour limit restricts how much can be consumed in one burst; Max still has it. 4:38
Theo estimates that $200/month can buy inference equivalent to as much as $12,000 at API prices. This is his valuation of usage, not a claim that Anthropic incurs $12,000 in direct costs. 5:01
His explanation centers on available GPUs: subscription traffic competes with enterprise inference for capacity already provisioned by the lab. Concentrated usage can crowd out capacity Anthropic could sell. 5:25 9:09

T3 Code weekday traffic begins rising around 4 a.m. Pacific, peaks at 7–9 a.m., and falls into a lull from about 6 p.m. to midnight. Theo reports roughly a 2x difference between daily lows and highs. These are T3 Code observations, not Anthropic-wide measurements. 5:25 5:50
Five-hour caps prevent users from spending an entire weekly allowance during the busiest window. Theo initially illustrates this with a rough 10–15% allowance per window, rather than a measured universal quota. 6:21
Later, his calculation for the limits under discussion is roughly 30% of a weekly allowance per five-hour window, or about 3.5 full session allowances per week. Keep this estimate separate from his earlier illustration. 11:29
Theo connects past peak-hour reductions and off-peak generosity to compute scarcity. He also interprets late-night and weekend Codex resets as a way to use spare capacity. 7:01
He argues that low-priced subscriptions encourage habits that later produce employer spending at API prices. This is his business interpretation, not demonstrated profitability data. 7:37
Session limits offer users one protection: an unexpectedly long-running agent cannot exhaust the whole weekly allowance in one burst. 7:37 8:03

Theo recalls a May compute partnership with SpaceX as a multibillion-dollar deal providing access to GPUs in an older data center. He speculates that Anthropic may now operate the whole facility. 8:20 8:44
He reports that Anthropic then doubled Claude Code's five-hour limits and removed Claude Code's peak-hour reductions. The product-specific wording matters for SDK apps discussed later. 8:44 9:34