▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=Q7n0PGbMW_U
• Original title :: They lied...

Anthropic announced that from September 14th it will "permanently raise standard weekly Claude Code limits by 25%" for Pro, Max, Teams and seat-based enterprise plans. Theo unpacks the arithmetic: the 25% is measured against the original base, not against today's limits, which have been running 50% above base since May. The net effect for current users is a ~17% reduction. The video is less about the number than about the communication pattern — Anthropic posted, deleted, and re-posted a message that leads with good news to describe a downgrade — and about why Anthropic keeps being squeezed on compute.

What actually changed

Base weekly limits go up 25% permanently on September 14th. The temporary 50% boost, in place since May and extended repeatedly, ends the same day. Going from base×1.5 to base×1.25 is a 17% cut for anyone who has been using Claude Code during the promotion window. The first version of the announcement was misleading enough that Anthropic deleted it; the replacement opens almost identically.

The history that got here

• May :: A SpaceX compute partnership let Anthropic double the 5-hour rate limits for Pro/Max/Teams, remove peak-hour reductions, and raise Opus API limits. Theo's complaint at the time: the binding constraint was the weekly limit, not the hourly one — you can burn a week's allowance in two sessions.
• May 13th :: Weekly limits raised 50%, nominally until July 13th.
• Fable removal and return :: When Fable was pulled from the plans, subscribers lost the thing many had subscribed for, so the increase was extended again. When Fable came back, it came with a separate sub-limit: only about half of a weekly allowance can be spent on Fable before you are forced onto Opus or Sonnet.
• August/September :: One further extension, then the September 14th reset.

The Fable sub-limit is the change that hurts Theo most. He maintains five Claude accounts and uses them essentially only for Fable, so the split means he effectively loses half of each plan. He notes, as a correction of his own earlier take, that Opus is not good right now and he was wrong to say it was.

Why the comms matter

Most Claude Code customers are not on Twitter. They learn what they are buying from email and from the plan description at signup — and the majority of heavy users likely signed up or upgraded after the 50% boost landed, so the only limit they have ever experienced is the one about to disappear. Announcing that as an increase is, in Theo's reading, close to misinforming customers, and cancelling in response is a reasonable reaction.

He rewrites the announcement to show the alternative: lead with the bad news, state the promotion is ending, give the honest net figure (25% above base, 17% below today), acknowledge the impact, and say what you are doing about capacity. He is explicit that the goodwill gesture in his version — dropping the separate Fable weekly limit — is invented for the example and not something Anthropic has said.

Anthropic can communicate well when it suits them

The contrast case is Anthropic's chief compute officer commenting on the OpenAI/Cursor breakup: warm about Cursor, pointed about OpenAI, and forward-looking about SpaceX. OpenAI's own note on cutting Cursor off gave maximum contractual notice and centred the affected developers. Anthropic writes with that kind of empathy when there is a competitor to dunk on or a compute partner to flatter, and not when its own users are the ones absorbing the hit.

The compute economics underneath

Theo's read: subscriptions have become a loss leader that competes for the same GPUs as per-token enterprise API sales, which are far less subsidised. During business hours, when enterprise demand peaks, subscription traffic is what has to give. He expects peak-hour restrictions to come back as a pattern across the industry, since it is the cleanest way to protect margin on flat-rate plans.

Takeaways

• The headline "25% increase" is measured from a base nobody has been on since May; the real change is a 17% cut.
• Anthropic's pattern is to hide, deflect, or reframe its own bad decisions rather than admit them — deleting and re-posting the same spin is an unusually clean example.
• The Fable weekly sub-limit compounds the reduction for anyone who subscribes specifically for Fable.
• Restrictive, anti-user policy (Fable sub-limits, bans for OpenClaw mentions in commit history) contrasts with OpenAI letting Codex tokens be used wherever the customer wants, and is a competitive liability if OpenAI's models keep closing the gap.
• Theo's position is not model criticism: Fable 5 is, in his view, unbelievably good and worth the money. The objection is entirely to how the business treats the people paying for it.