Hitting your Claude limit mid-task feels like a power cut in the last over of an India vs Pakistan match.

You know something good was about to happen, and now you're sitting in the dark waiting for the inverter to kick in.

Revolut's engineers lived in that power cut, except the meter was still running.

Some were spending up to $10,000 a month each on Claude Code, about ₹9.5 lakh a month per person, which is basically a salary paid to a chatbot. And the company says the gains were small.

So Revolut moved around 2,000 engineers to GLM, an open model from China's Z.ai, running on its own setup with no per-engineer budgets. Claude Code and Codex are now on the exception list.

The bill dropped to about a tenth, and work now goes from ticket to merge around 70% faster. That 70% came from the most boring place possible.

Some housekeeping first, then I'll show you.

Presented by Tiger Data

When analytics on Postgres slows down, most teams add a second database. Then come the pipelines, the sync jobs, and a copy of your data that's always a little behind.

TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Hypertables partition your data automatically as volume grows. Hypercore compression cuts storage up to 95%. Continuous aggregates keep dashboards live without re-querying everything.

CERN runs Postgres this way for sensor data from the Large Hadron Collider.

No split architecture, no pipeline lag, no new query language to learn. Same SQL, same drivers, same tools.

Start on Tiger Cloud and get $1000 in credits.

Get $1000 Credit

Presented by Superhuman AI

The next wave of millionaires will be people who figured out how to make AI work for them.

The window to get ahead is still open. But not for long.

Here are 200+ proven ways to make money with AI in 2026.

Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.

Claim your free list

Friday Work

When Claude hits its limit mid-task, it hands the work to another coding agent you've set up, like Codex or Cursor, so the task keeps going. It's open source and built for coding agents, so forward it to your developer friends.

Teamily AI

A group chat where your teammates and AI agents sit together. Tag an agent, give it a task, and it remembers what was said across chats, which already makes it better than most colleagues. Free plan to start.

PasteGuard for Mac

A free menu bar app that masks emails, card numbers and API keys the moment you copy them, so they don't land in an AI chat by accident. Everything stays on your Mac, and one shortcut brings the originals back.

Gemini went company-wide at Revolut in January 2025 and coding got about 20% faster.

Then came Claude Code with strict budgets per engineer, so people saved their tokens for the "important" tickets and the gains stayed small.

In April, Revolut started self-hosting GLM (5.1 first, now 5.3) at about 30 cents per million tokens, and people stopped saving. 

Engineers ran tasks in parallel and let agents ping them when done, the way you put rice in the cooker and go reply to emails.

Now the boring bit I promised.

Coding time improved 60%, but lead time (from picking up a task to merging it) improved 70%. 

That extra jump came from the work around the code, like digging through Jira tickets and inboxes.

Nobody burns a tight budget on a status update, so that stuff only went to AI once tokens were cheap.

All of this runs on 56 Nvidia B200 chips serving 2,000 engineers. David Abitbol, Revolut's AI operations lead, says once waiting crosses five seconds, his Slack starts blowing up.

After 5pm, when load drops, engineers get about four times the speed, which might be the first good reason anyone's had to work late.

And because it's a bank, the model runs on US servers instead of Z.ai's China endpoint, with zero data retention.

Agents work in a sandbox, a human approves the work, and every step is traced.

Big engineering org with compliance teams: Revolut just did your homework, though Abitbol called the out-of-the-box performance "terrible," and making it fast took serious tuning.

50-person startup: Please don't, you'd be paying engineers to babysit GPUs. Rent open models from providers like Together AI or Fireworks and let someone else deal with the hardware.

Indian IT services firm or GCC paying AI bills in dollars: This one should make you nervous. Thousands of engineers on per-head licences is the same rationing Revolut walked away from.

One catch before anyone copies it: these are Revolut's own numbers, nobody has audited them, and the 90% saving leaves out GPU, staff and monitoring costs. Some developers say GLM struggles in long, complex sessions. And since the model switch and the end of limits happened together, nobody can say which one did the heavy lifting.

You don't need 56 GPUs to stop rationing.

Claude's still at Revolut, just as the expensive specialist you call when things get hard.

In June I said most of you should keep renting AI instead of running your own, and I still stand by that. Revolut isn't most of you.

My guess is that for seasoned engineers the gap between GLM and Claude is small, while juniors might still need the smarter model.

Cheap AI works best for people who can smell bad output in seconds, and if that's not you yet, that's the skill to build before you worry about which subscription to pay for.

For two years we've argued about which model is smartest. Revolut got its biggest gain by letting people use AI on everything, including the boring work nobody puts in a demo.

My bet is the teams that ship fastest next year will be the ones whose people stopped asking "is this worth a prompt?"

Reply and tell me what you're still rationing AI for.

Until next time,
Vaibhav 🤝