Revolut made GLM the default for 2,000 engineers. The gains came from no longer rationing tokens
Revolut's engineers used to save tokens for the end of the quarter. David Abitbol, the fintech's ops and infrastructure lead, told the AI Engineer conference in New York last month how that ended, and The Register's write-up on Tuesday plus a tweet at 738,000 views brought it back around. About 2,000 engineers, GLM-5.3 on 56 Nvidia B200s, about $0.30 per million tokens, roughly a tenth of Codex or Claude Code. Revolut reports inference costs down about 90%, coding time improved 60% and lead time 70%.
The cost number is the headline, and the behavior change is the story. Revolut launched AI coding with strict budget limits after seeing engineers spend $1,000 to $10,000 a month with little return. Once inference was flat and cheap, Abitbol says, the biggest gains came from debugging and Jira and Confluence tickets, the jobs nobody would spend rationed tokens on. "You no longer have to save your tokens for the end of the quarter." Metered pricing shapes which work gets done, and the highest return may sit in the work it discourages.
There are operating details worth stealing. "Long context is a scam," Abitbol said: you can serve a million tokens, but 256K is what you need. Aim for under 3 seconds to first token, since past 5 the Slack complaints start. 200 tokens a second is plenty, because people run tasks in parallel and stop watching the stream. The big gains came from tuning and quantization, not more GPUs.
Now the caveats. These are self-reported figures from a talk, which we read through The Register and not the video, and Revolut did not dump Claude. Vision work routes to Gemini or GPT Luna, and "the frontier is the exception". HN commenters also disputed whether $0.30 is impressive next to Haiku 5.5 pricing, and the coverage gives no staffing cost for running 56 GPUs at this level.
Gartner offered a useful sanity check the same day via The Register, from a note titled "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps". Nine in ten steps are routine. Open models cost as little as a twentieth of a frontier model on every step, but they are more verbose and retry more, so the analysts put the real per-task saving at five or six times. That second number is the honest one. Per-token price is marketing, and cost per finished task is the bill.
The Register on Revolut: https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345 Talk: https://www.youtube.com/watch?v=Ok6QdZHCcI8 Gartner via The Register: https://www.theregister.com/ai-and-ml/2026/11/03/open-models-can-do-90-of-agentic-work-at-a-twentieth-of-the-cost-of-a-frontier-llm-gartner/5307132
← Back to all articles
The cost number is the headline, and the behavior change is the story. Revolut launched AI coding with strict budget limits after seeing engineers spend $1,000 to $10,000 a month with little return. Once inference was flat and cheap, Abitbol says, the biggest gains came from debugging and Jira and Confluence tickets, the jobs nobody would spend rationed tokens on. "You no longer have to save your tokens for the end of the quarter." Metered pricing shapes which work gets done, and the highest return may sit in the work it discourages.
There are operating details worth stealing. "Long context is a scam," Abitbol said: you can serve a million tokens, but 256K is what you need. Aim for under 3 seconds to first token, since past 5 the Slack complaints start. 200 tokens a second is plenty, because people run tasks in parallel and stop watching the stream. The big gains came from tuning and quantization, not more GPUs.
Now the caveats. These are self-reported figures from a talk, which we read through The Register and not the video, and Revolut did not dump Claude. Vision work routes to Gemini or GPT Luna, and "the frontier is the exception". HN commenters also disputed whether $0.30 is impressive next to Haiku 5.5 pricing, and the coverage gives no staffing cost for running 56 GPUs at this level.
Gartner offered a useful sanity check the same day via The Register, from a note titled "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps". Nine in ten steps are routine. Open models cost as little as a twentieth of a frontier model on every step, but they are more verbose and retry more, so the analysts put the real per-task saving at five or six times. That second number is the honest one. Per-token price is marketing, and cost per finished task is the bill.
The Register on Revolut: https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345 Talk: https://www.youtube.com/watch?v=Ok6QdZHCcI8 Gartner via The Register: https://www.theregister.com/ai-and-ml/2026/11/03/open-models-can-do-90-of-agentic-work-at-a-twentieth-of-the-cost-of-a-frontier-llm-gartner/5307132
Comments