Revolut slashes AI coding costs 10x by cutting out Claude and going

British fintech Revolut slashed its AI coding inference costs by roughly 90 percent in less than a year, a 10x cut, by making self-hosted models the default instead of relying primarily on commercial services like Claude. David Abitbol, ops and infrastructure lead, described that rollout at the AI Engineer conference in New York last month in theregister.com's report by systems editor Tobias Mann, published Tuesday 3 November 2026 at 20:07 UTC. Revolut adopted GLM-5.1, then moved to GLM-5.3, which it serves on 56 Nvidia B200 GPUs.

Platform leads and engineering managers comparing self-hosted coding models with Claude Code or OpenAI Codex are the audience. This piece covers the token caps before GLM, how the B200 stack is tuned, why Jira, Confluence, and debugging soaked up usage, who still uses Gemini or GPT Luna, and the latency and privacy limits Abitbol set.

Revolut slashes AI coding costs 10x: what actually changed

Revolut was experimenting with commercial AI before GLM. It deployed Gemini in early 2025 and Claude Code in early 2026. "We were highly skeptical about the cost of AI. We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI, and so when we launched it we launched it with strict budget limits," Abitbol said. Those limits gave engineers a limited number of tokens. The shortage is what led Revolut to investigate self-hosting its models.

In less than a year, self-hosted models became the default and AI coding inference costs fell by roughly 90 percent. Revolut adopted GLM-5.1 first, then moved to GLM-5.3, which it serves on 56 Nvidia B200 GPUs, equivalent in GPU count to seven eight-GPU DGX B200 systems. At $0.30 per million input/output tokens, self-hosting GLM 5.1 cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code. Once tokens were that cheap, engineers stopped worrying about how many they used or what they used them for.

Revolut slashes AI coding costs 10x: how it works

GLM 5.1 comes from Z.ai, a Chinese open-weights developer. It has roughly 754 billion total parameters, with 40 billion activated for each token generated. The weights need roughly 380 GB of GPU memory, plus another 6 to 26 GB per user to maintain context, depending on precision. Revolut serves GLM-5.3 on those 56 B200s. Abitbol said the greatest performance gains came from tuning hyperparameters, switching to NVFP4 quantization, implementing disaggregated prefill and decode, and tuning context lengths, rather than from adding GPUs.

A one-million-token context window can be served and still fail to be useful or efficient. "Long context is a scam. You really need to be mindful of that," Abitbol said. "Even though you can serve one million tokens, 256 thousand is really what you need." With those optimizations, Revolut reached a sub-3-second time to first token on context windows of around 256,000 tokens and generation speeds of around 200 tokens a second. He called both numbers the sweet spot for the company's software engineers.

Revolut slashes AI coding costs 10x: why it matters

Revolut did see improvements to coding speed: a 60 percent improvement in coding time and a 70 percent improvement in lead time across roughly 2,000 engineers. Abitbol said the biggest bottleneck was not the code. The biggest gains came from letting the models deal with debugging and with Jira and Confluence tickets. "When you have unlimited tokens, you're free to do that," he said. "You no longer have to save your tokens for the end of the quarter or the most difficult task."

"What you want to aim for is a sub-3 second time to first token. This is empirical evidence. As soon as you exceed 5 seconds, my Slack inbox starts to blow up," Abitbol said. Throughput matters less. Revolut serves 200 tokens per second interactively. He said people are fine with that and do not need to go higher. They run tasks in parallel and check in when the task is done. After 5 pm, engineers get closer to 800 tokens a second.

Revolut slashes AI coding costs 10x: who is affected

The rollout sits with roughly 2,000 Revolut engineers, the group that recorded the 60 percent coding-time gain and the 70 percent lead-time gain. They are the same engineers who had been spending $1,000 a month and $10,000 a month on commercial AI with not so much ROI, then working inside strict token budgets. Self-hosting removed the reason to ration. Debugging, Jira tickets, and Confluence tickets became normal model jobs, because nobody had to save tokens for the end of the quarter or the most difficult task.

GLM has not replaced GPT or Claude entirely. The text-only model is a poor fit for UI work, so a router sends vision queries to Gemini or GPT Luna and keeps code generation on GLM. "Open source is now a priority at Revolut, and our economics are favorable. The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," Abitbol said. Privacy and security concerns, he said, center on APIs running out of China. On-prem or on regional cloud, there is nothing to worry about.

Revolut slashes AI coding costs 10x: what to watch

Abitbol's target is a sub-3-second time to first token on windows of around 256,000 tokens. At 5 seconds his Slack inbox starts to blow up. Interactive generation sits near 200 tokens a second. He said people are fine with that. They run tasks in parallel and check in when the task is done, and they no longer look at the stream of thinking tokens. After 5 pm, engineers get closer to 800 tokens a second. One million tokens can be served. He said 256 thousand is what you need.

Self-hosting GLM 5.1 cost $0.30 per million input/output tokens, about a tenth of OpenAI's Codex or Anthropic's Claude Code, and the inference bill fell by roughly 90 percent in less than a year. GLM-5.3 on 56 Nvidia B200 GPUs stays the code default, and vision still routes to Gemini or GPT Luna. Open source stays the priority, with flat billing, no end-of-month bill shock, and the frontier as the exception. APIs running out of China remain the privacy concern. On-prem or regional cloud, he said, leaves nothing to worry about.