For most of three weeks this spring, my work kept stalling against Claude's usage limits. This is the part nobody warns you about — the field doesn't wait for you. A first-person account of what happened in AI while I couldn't touch it, and why it's the reason llmdeal.me exists.
· llmdeal.me
I am not a casual user. In a normal stretch I'll put 40-odd million tokens through a model in 17 of 21 days. That isn't abuse — it's what building production software with an agentic toolchain actually looks like in 2026. For a long time a flat subscription absorbed that fine. Then, through late April and into May, it stopped.
The limits moved underneath me. A session that used to run five hours started tapping out far sooner under sustained load. On April 21 Claude Code briefly vanished from the $20 Pro plan entirely before being put back about a day later. I started planning my day around a rolling usage window instead of around the work. And on the bad days I did the thing that still feels absurd to write down: I closed the laptop and waited. Not because the work was done — because I had run out of allowance and the meter would not reset for hours, sometimes longer.
Here is what that taught me, and it is the whole reason this company exists: when your stack depends on one vendor's capacity, their bad week becomes your lost week. And the field does not slow down to match your allowance. While I was rationed, the rest of AI sprinted. This is the catch-up I had to do when I got back.
Every item below is dated and sourced. Three and a half weeks; I've kept it to the things that actually mattered.
April 23 — Anthropic published an unusually candid engineering postmortem, admitting three missteps that had quietly degraded Claude for weeks: reasoning effort silently lowered, a context-caching bug, and a response-length cap that measurably cost quality. Usage limits were reset for all subscribers as a goodwill gesture. May 6 — Anthropic doubled Claude Code's 5-hour limits, removed peak-hour throttling, and raised Opus API rate limits 2–16×; the same day it announced a SpaceX/Colossus compute deal (220,000+ GPUs) and Dario Amodei disclosed roughly 80× year-on-year growth. May 13 — weekly Claude Code limits were raised another 50% through mid-July, and a $200/month Agent SDK credit was announced for Max 20x subscribers (effective June 15). Read that sequence again: the relief came after the cancellations, in three waves, each one chasing the backlash.
April 23 — OpenAI launched GPT-5.5. May 5 — GPT-5.5 Instant became the default model in ChatGPT for everyone. May 7 — OpenAI's Codex Chrome extension crossed 4 million weekly users. May 14 — OpenAI shipped a Codex mobile app and ran an explicit "switch to Codex" promotion aimed straight at Claude defectors. The same day, xAI launched Grok Build, a Claude Code-style CLI, on its $300/month SuperGrok Heavy tier. Google wasn't idle either: a broader Gemini 3.1 Pro rollout on May 4 and gemini-3.1-flash-lite reaching general availability on May 7. The coding-agent market re-shuffled in the exact window I couldn't participate in it.
This is the part that actually changed my mind. April 24 — DeepSeek released V4-Pro and V4-Flash: open weights, MIT-licensed, 1M-token context, and pricing low enough to make a frontier subscription look like a luxury. April 29 — Mistral shipped Medium 3.5. Days before the window opened, Moonshot's Kimi K2.6 had landed too. While I was being told to wait for capacity I was already paying for, models I could run or rent for a fraction of the price were getting good enough to do most of my work. The "use a cheaper model" advice stopped being a compromise.
April 24 — Google committed up to $40 billion to Anthropic. May 12 — Bloomberg reported Anthropic in talks for a $30 billion round at a $900 billion valuation. The same week, Bloomberg also reported that somewhere between a third and half of planned 2026 US data centres were being delayed or cancelled. The capacity crunch that rationed me is not a glitch that ends next quarter — it's the structural condition of the whole industry right now, and the valuations are being written as if it isn't.
Line all of it up and one fact stands out. Not one of the limits I hit was about my workload. They were vendor decisions — capacity allocation, plan economics, an experiment I happened to be inside of. Anthropic's own postmortem confirmed the quality I'd been getting in April was not the quality I'd calibrated against in January, and I found that out from a usage meter, not an email.
I don't think Anthropic is the villain here. Claude is still one of the best models in the world, the postmortem was honest, and the fixes were real. The villain, if there is one, is the architecture of the dependency: a single vendor, a single subscription, a single capacity pool, holding your ability to ship. When that vendor has a bad three weeks, so do you. That is not a model problem you can fix by switching models. It is a structural problem you fix by not having a single point of failure.
I didn't set out to start an AI company. I set out to never lose another three weeks to one provider's capacity planning.
So that's what llmdeal.me is. One OpenAI-compatible key. Smart routing across cheap, open-weight, and frontier models. When any single provider has a bad week — throttles you, degrades quietly, prices you out — it becomes a routing decision instead of a dead stop. The easy 60–80% of calls go to a cheap or self-hosted model that, as of this spring, is genuinely good enough; the hard ones escalate. It's EU-resident, paid in crypto, with no KYC, because by the time you've been rationed by a vendor you also start noticing that your prompts and your billing trail are sitting in someone else's jurisdiction.
That's the founder's-honest version of the pitch. I built the thing I needed during the worst working month I'd had in a year. If you've had that month too, you already understand the product. If you want the rest of the story — the numbers, the tier design, the build — that's what the rest of this site is for: see the pricing or how it works.
Dates and events checked against primary reporting, May 2026. Article published 2026-05-16. This is a first-person founder note; the AI-industry events it describes are independently sourced above.