PARTLY

Partly true: Claude Code usage limits on Max plans are real, documented and have moved around in 2026, but 'Opus 5 was built to throttle you' is the poster's own theory with no evidence, and the follow-up model got faster and cheaper

Facebook · Oct 2026

The claim'In 2027 there's going to be a massive shift from cloud-based LLMs to open-weight and full open-source AI.' 'We can't trust these guys... they're continuing to do these price hikes. We don't know what they're doing with these tokens... they can charge us whatever they want.' 'With Claude Code on the Max plan, they are regulating. They're throttling the usage to essentially make it less performant.' 'Part of my theory with Opus 5 is that Opus 5 is an upgrade to allow them to throttle the AI, that it has something in it that makes it intentionally slower so it doesn't have as much compute overhead.' On-screen slides list 'Throttle what you already paid for' and 'Bill you on a count only they can see.'

There is a real grievance under this. Claude Code on Pro and Max runs on five-hour and weekly usage limits that Anthropic sets and changes: it doubled five-hour limits in May 2026, then on September 14 cut weekly Claude Code capacity by about 17% versus a temporary promotion (Anthropic says that is still 25% above the pre-May level). A proposed class action filed in June 2026 alleges Max delivers less usage than advertised; that is an allegation, not a ruling. Anthropic's own April 2026 postmortem admitted Claude Code had lowered default reasoning effort from high to medium for about a month, plus a caching bug and a prompt change, all reverted, with usage limits reset; it said model weights and the API were untouched. So 'limits are tight and shift' is fair, and 'quality dipped from a product change' happened once. The Opus 5 part is different. The speaker labels it 'my theory' and offers no evidence. Opus 5 launched July 24, 2026 with selectable effort levels (default high in Claude Code), so slowness is mostly a setting. Anthropic has repeatedly stated it does not intentionally degrade models because of demand. And Opus 5.5 shipped September 22 at 40% lower price with reportedly faster output. Token counts are not invisible either: every API response returns usage. The 'Throttle what you already paid for' slide cites the speaker's own stream as its source. The case for open-weight models is a reasonable opinion; the caption sells a members stream.

What holds up

Full transcript and 16-frame contact sheet reviewed; the throttling slide's source line credits the speaker's own livestream.
InfoQ coverage of Anthropic's April 2026 Claude Code postmortem: effort default lowered March 4 and reverted April 7, a caching bug fixed April 10, a verbosity prompt reverted April 20; 'The API and underlying model weights were not affected.'
Implicator: Anthropic announced on Aug 29, 2026 that weekly Claude Code limits would fall about 17% from promo levels on Sept 14, 25% above pre-May; five-hour limits unchanged.
Release coverage: Opus 5 on July 24, 2026 with five effort levels; Opus 5.5 on Sept 22, 2026 at $4/$20 per million tokens versus $5/$25, about 40% cheaper.
June 2026 proposed class action over Max usage limits reported by several outlets; no ruling found.

What doesn’t

'Opus 5 is built to be intentionally slower' is self-described speculation with no evidence offered.
A slide cites the speaker's own stream as proof of throttling.
Price-hike examples on screen are from GitHub Copilot and Google, not Anthropic, which cut Opus pricing.
Caption funnels to a paid members stream.

The catch

The limits are real and Anthropic has admitted one stretch where Claude Code quietly got worse, so the frustration is earned. But the jump from 'limits tightened' to 'the new model was engineered to be slow' has nothing behind it, and the next release went the other way on price and speed.

How to actually do it

Check your usage meter in Claude Code (the /usage command) or in Settings, Usage on claude.ai to see where you sit against the five-hour and weekly limits before assuming throttling.
If it feels slow, check the effort setting; high or max effort thinks longer by design, and medium is much faster for routine edits.
Check Anthropic's status page for incidents and its engineering posts for postmortems before drawing conclusions.
If you want insurance against vendor changes, keep a local open-weight model (via Ollama or LM Studio) for routine tasks and pay per token on the API, where usage is itemized per call.

Usage caps and their changes are documented, and one quality dip was confirmed and reversed by Anthropic. The Opus 5 throttling theory is unsupported, and the vendor-independence goal is achievable today with local models plus metered API use.

Confidence: Medium
Posted by: a tech livestream podcaster (about 15,400 plays) promoting a paid YouTube members stream; reel sent to Buddy Wed Oct 28, 2026, late night

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.