▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=y5TEATUDDfc
• Original title :: Anthropic can't stop winning

Anthropic renamed its $200/month "Max 20x" plan to Ultra and removed the 5-hour session limit for verified Ultra accounts. Theo explains why the 5-hour limit exists at all: it spreads GPU demand across the day so subscription users don't eat the peak-hour compute Anthropic sells to enterprises. Verification is vague by design, because it targets distillation farms that burn subscriptions. The catch for T3 Code: Claude subscriptions used through the Agent SDK burn the 5-hour window about 1.8x faster than the interactive CLI, so $20 and $100 users of third-party apps are hit harder. Mid-recording Anthropic informally okayed a T3 Code change to its entry-point identifier that removes the penalty.

The new plan lineup

• Four tiers: Free, Pro ($20), Max ($100), Ultra ($200). Ultra is the old 20x plan at the same price.
• The rename fixes misleading naming: "20x" only ever applied to the 5-hour limit. Weekly limits, added later, were set at 10x Pro. The head of product for Claude subscriptions publicly apologised for the confusion; the plan pages now state the multipliers explicitly.
• Theo, with hill-climbing jobs turned off, struggles to use half of one Ultra account. For most realistic dev work, Ultra is more than enough.
• Ultra can deliver up to roughly $12,000/month of API-priced inference. The labs make little or no money on it; the point is to get developers used to that volume so employers later pay API prices.

Why the 5-hour limit exists

• Two limits: a weekly cap, and a 5-hour window that stops one burst from consuming the whole week.
• The issue is compute availability, not cost. GPUs already exist; what matters is whether subscription use takes away capacity that could be sold to enterprises during work hours.
• T3 Code analytics: prompting ramps from about 4 a.m. Pacific, peaks 7-9 a.m., and drops sharply from 6 p.m. to midnight. Peak and trough differ about 2x on a normal weekday.
• With the 5-hour window, a user can spend only about 10-15% (by Theo's later estimate, about 30%) of the weekly budget in any one window. Anthropic once went further and cut limits during peak hours.
• Off-peak GPUs sit idle, so labs don't mind night and weekend usage. Codex limit resets tend to land at those times.
• Minor upside for users: the window limits the damage of a runaway prompt that burns far more tokens than expected.
• The May SpaceX compute deal (Anthropic took over a mostly idle SpaceX data centre) let Anthropic double Claude Code's 5-hour limits and drop the peak-hour cuts. The wording said "Claude Code" specifically, and that turns out to matter.

"Verified" accounts

• The limit is removed for most existing subscribers. Newer accounts keep it until verified, "to prevent fraud and abuse". Phone verification at checkout does not count. No current-session bar under Settings → Usage means the account is verified.
• Theo dislikes the lack of transparency but understands it. The target is distillation: labs, often in sanctioned countries, buy subscriptions ($200 in, $12k out) and burn them to collect request/response pairs. Removing the window makes that easier, so the checks have to be opaque.

The Agent SDK penalty

• Users reported, and Theo confirmed, that T3 Code burns the 5-hour window about 1.8x faster per token than the interactive Claude Code CLI. claude -p has the same problem. The doubled 5-hour limit applied only to the interactive Claude Code entry point.
• Ultra users are no longer affected. $20 and $100 users of third-party apps get both the smaller window and the SDK penalty.
• OpenCode's new Claude subscription support also runs on the Agent SDK, so it has the same penalty. Any app using Claude subscriptions must either drive the CLI visibly or use the SDK.
• Workaround: the Claude Code entry-point field is set to sdk-ts. Any other value removes the penalty. Theo argues reporting t3-code is simply more accurate. While filming, Anthropic contacts gave semi-official permission to merge that PR, so T3 Code users get the same 5-hour and weekly usage as Claude Code itself.
• Theo's own multi-account setup routes through a CLI proxy that already did this. He hints that the config is in a gist linked from the FAQ, and disclaims responsibility for any bans.

Coming next

• A video comparing the new Codex plans with Claude's. Before this change Codex was poor value next to Claude; Theo says the math is now very different.