Staying up late to rant about a favorite tool is a special kind of frustration. You love the product, you pay full price, and suddenly you can’t use what you bought. That’s the mood in a new breakdown from Matthew Berman, the creator who spent an evening writing an essay on what’s going wrong at Anthropic and then walked through it live.
The expert still loves Opus 4.6 and 4.7. He uses them daily. But he argues Anthropic’s communications and quota games are pushing power users toward Codex, and it all traces back to one call Dario made roughly 18 months ago.
🧠 The one miscalculation
Anthropic built what the creator calls the greatest flywheel in modern business: a top coding model, enterprise revenue, coding data, better next-gen models, repeat. The catch? Dario chose not to YOLO on CapEx. On the Dwarkesh podcast, he said buying a trillion dollars of compute would risk bankruptcy if growth slipped. Safe bet, except demand actually did keep going 10x. Now Anthropic can’t serve it, and OpenAI is scooping up the overflow.
⚠️ The three moves that sparked the backlash
- Claude Code quietly removed from Pro and Free tiers as a “2% pricing experiment,” even though users across the board saw it. No announcement, just a pricing page edit.
- OpenClaw restrictions dropped at 4pm on Easter Friday, giving users under 24 hours to move off Opus. Boris promised clarity in the docs. Three weeks later, still none.
- Peak-hour quota throttling that the post’s author says targets the 7% of power users running agentic harnesses like OpenClaw and Hermes.
📉 Uptime and token math working against them
The LinkedIn creator pulled the status pages live. Claude.ai sitting at 98.8%. platform.claude.com at 99%. Meanwhile Codex runs 99.98%. On top of that, Opus 4.7 ships with a new tokenizer that burns 1 to 1.35x more tokens on the same input, plus extra thinking tokens at higher effort. More compute crunch, same quota.
🎯 What the savvy professional wants Anthropic to do
His ask is refreshingly simple:
- Be clear about what’s in policy and what isn’t. No more employee tweets that contradict the docs.
- Raise prices or lower quotas if the math doesn’t work. Just don’t restrict how paying users spend the tokens they already bought.
- Give real notice before breaking workflows. A Friday-afternoon email isn’t “plenty of notice.”
Meanwhile Tibo and Rohan from the Codex team are throwing resets like candy, Sam Altman is drunk-tweeting “okay boomer,” and Cursor just teamed up with xAI’s Colossus supercomputer. The creator sketched a compute-vs-demand spectrum: xAI compute-rich and demand-light, Anthropic demand-rich and compute-starved, OpenAI healthier, Google comfortably selling TPUs to everyone including its competitors.
💡 The takeaway for builders
If you’re running agents in parallel on a Max plan, treat this as a signal. Shift async jobs to off-peak hours while the 2x weekend bonus lasts. Keep a Codex subscription warm as a backup. And watch the AWS Trainium deal, 5 GW over 10 years, because nothing ships overnight.
The original poster isn’t writing Anthropic off. He calls Opus a fantastic model and notes Rune’s point that their main problem is everyone wants the product. But the half-measures, opaque policies, and comms flip-flops are doing real damage.
Watch the full video for the drawn-out flywheel diagram, the Dario clip, and the drunk Sam Altman tour.