How One Bank Cut Its AI Coding Bill to a Tenth — and What It Means for You
When British fintech Revolut rolled out AI coding assistants to roughly 2,000 engineers, the bill did what AI bills always do: it exploded. Some engineers were spending $1,000 — even $10,000 — a month on AI tokens with little measurable return. So Revolut did something almost no company of its size has tried: it ditched the frontier APIs as its default and moved its engineers onto a self-hosted, open-weight AI model. The result, according to Revolut's own numbers, was a roughly 90% cut in AI coding inference costs — plus coding that got about 60% faster.
This is the Revolut story explained in plain English: what they did, the numbers they reported, the honest caveats, and what regular businesses and developers can actually learn from it.
What Revolut actually did
Revolut — the UK-based digital bank with around 80 million customers — laid out the move in a talk by David Abitbol, who runs AI operations and infrastructure at the company, at the AI Engineer conference in New York. The talk, "Shipping like a startup, regulated like a bank," was posted on 29 October 2026, and reported widely on 3 November.
The timeline, per Abitbol's talk:
- Early 2025: Revolut rolled out Google's Gemini company-wide — roughly 20% faster coding time, 10% better lead time.
- Early 2026: Claude Code arrived, but with strict per-engineer budgets, because spending was getting out of hand. Measured impact was modest.
- April 2026: Revolut began self-hosting open-weight models, starting with GLM-5.1, an open-weight model from Chinese AI lab Z.ai — with no token limits at all.
- Later: Upgraded to GLM-5.3 (roughly 754 billion total parameters, about 40 billion active per token), served from about 56 Nvidia B200 GPUs — seven eight-GPU DGX B200 systems' worth of compute.
The default coding model for Revolut's roughly 2,000 engineers is now that self-hosted GLM. Anthropic's Claude Code and OpenAI's Codex haven't vanished — they've been pushed to an "exception" list, reached for only when the self-hosted setup genuinely can't do the job.
The numbers that matter
All of these figures come from Abitbol's talk — they are Revolut's own measurements, not independent audits, but they are strikingly concrete:
- ~90% cost cut: Revolut's blended serving cost landed at about $0.30 per million tokens, and its AI coding invoice is roughly one tenth of what it paid Anthropic and OpenAI. (By FourWeekMBA's arithmetic, the previous bill was near $3 per million tokens — Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens.)
- ~60% improvement in coding time and ~70% improvement in lead time (task pickup to merge), measured across about 2,000 engineers.
- 76% more requests per second with no extra GPUs after tuning, and an 88% reduction in time-to-first-token — from about 14.3 seconds to 1.7 seconds.
- GPU count held steady at 56 B200s from April through August even as usage grew, thanks to optimisation work.
The killer detail: before the switch, engineers were rationing AI because every API call carried a metered bill. Abitbol said some engineers burned through $1,000–$10,000 a month with little return on investment. Unlimited, flat-cost tokens changed the behaviour — and the behaviour, not a smarter model, drove most of the speed gain.
Why self-hosting worked: the end of token rationing
The most important sentence in this whole story might be the one Startup Fortune pulled from the talk: "cheap unlimited tokens, not smarter AI, is what speeds up coding." Revolut's engineers didn't get a better model than Claude Opus — GLM-5.3 is a step down from the frontier on many benchmarks. They got a model they could use without thinking about the meter running.
That unlocked three things:
- No rationing. Engineers stopped doing mental accounting before each prompt. Heavy agentic loops — the kind that burn thousands of tokens re-reading files and iterating — became free at the margin.
- Flat, predictable billing. "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," Abitbol said. CFOs love that sentence.
- Latency control. Self-hosting let Revolut tune inference for its own engineers — cutting time-to-first-token by 88% — instead of waiting in a shared API queue.
What Revolut didn't throw away
This is not a story about open models "beating" Claude and GPT. Revolut is explicit that frontier models remain in the stack for the jobs where they're clearly better. GLM is text-only, so anything involving UI work or multimodal reasoning still goes to the frontier labs. Abitbol said Revolut still uses Codex and Claude Code and called the latest models "fantastic" — the frontier is the exception, not the rule.
There's also a compliance angle worth noting: Revolut runs the open weights on US-based infrastructure rather than through Z.ai's hosted endpoint in China. For a regulated bank, self-hosting isn't just cheaper — it's how you keep data inside your own legal and regulatory perimeter. That argument applies to any company with data-residency requirements, not just banks.
The honest caveats
Before anyone reads this as "self-host AI and save 90%," the fine print matters:
- These are Revolut's numbers. The company hasn't isolated the model change from other causes — removing token limits alone could explain much of the productivity gain, as Revolut itself acknowledges.
- The 30-cent figure is incomplete. StartupHub notes the talk left cloud spend, staffing, and observability costs out of the $0.30-per-million-token math. Those are real costs; Revolut just absorbed them into existing infrastructure teams.
- Fifty-six B200 GPUs is serious money. That's the compute budget of a mid-size AI lab. A 20-person startup can't replicate this on a spare server — though, as we covered in our best AI coding assistants guide, smaller open-weight models like DeepSeek V4.1 Flash or Kimi K3 now fit on far less hardware.
- It took real tuning. The 76% throughput gain came from engineering work, not from downloading weights and pressing go. Self-hosting is an operations project, not a purchase.
What this means for you
Most readers of this blog won't self-host a 754-billion-parameter model. But the story still changes the calculus:
- For startups and small teams: the open-weight ecosystem keeps getting cheaper. Models like DeepSeek V4.1 Flash are MIT-licensed and downloadable today, and inference APIs for open models now undercut frontier pricing dramatically. Revolut's story is the extreme version of a trend that already reaches small businesses — we track it in our AI tools for small business coverage.
- For developers: the skill that matters is knowing which tasks need a frontier model and which don't. Routine coding help, refactors, and test generation run fine on open models; novel architecture and hard debugging still justify the frontier bill. Use frontier models like Revolut does now: as the exception.
- For enterprises: predictable billing and data residency are now first-class arguments for self-hosting, alongside raw cost. Revolut — a regulated bank — just validated the pattern at scale.
- For the AI labs: this is the threat frontier vendors quietly fear. If the "good enough" tier of work migrates to self-hosted open weights, metered API revenue concentrates in the hardest 10% of tasks.
The open-model landscape moves fast — our DeepSeek Super-Agent post covers another open-weight coding story from this month, and Moonshot's Kimi K3.1 Pro is due to release its full weights any day now.
Verdict
Revolut's experiment is the strongest real-world evidence yet that unlimited, cheap tokens beat smarter models for everyday coding work. The headline — 90% cost cut, 60% faster coding — comes with genuine caveats (their own figures, incomplete cost math, serious GPU budget), but the direction of travel is unmistakable: open-weight models have crossed the "good enough for production" line for the bulk of coding work.
If you run an engineering team and your AI bill makes your CFO flinch, the Revolut playbook — self-host an open model for the everyday work, keep the frontier for the exceptions — is no longer a theory. It's a documented playbook with numbers attached.
Frequently Asked Questions
Which AI model did Revolut switch to?
Revolut's default coding model is now GLM-5.3, an open-weight model from Chinese AI lab Z.ai, self-hosted on about 56 Nvidia B200 GPUs. It started the self-hosting experiment in April 2026 with the earlier GLM-5.1.
Did Revolut stop using Claude and ChatGPT entirely?
No. Revolut still uses Claude Code and OpenAI's Codex for tasks where the self-hosted model isn't good enough — GLM is text-only, so UI work and the hardest problems still go to frontier models. The frontier tools are now the exception rather than the default.
How much did Revolut actually save?
Revolut reports its AI coding inference invoice fell to roughly one tenth of what it paid Anthropic and OpenAI — a ~90% cut — with a blended serving cost of about $0.30 per million tokens. Note this excludes cloud, staffing, and observability costs, which the talk left out of the math.
Can a small company copy Revolut's approach?
Not at this scale — 56 B200 GPUs is serious infrastructure. But smaller open-weight models (DeepSeek V4.1 Flash, Kimi K3, GLM variants) are downloadable under open or permissive licenses and run on far less hardware, and hosted inference for open models undercuts frontier API pricing. The principle — open models for routine work, frontier for exceptions — scales down even if the GPU count doesn't.
Sources: figures from David Abitbol's talk at the AI Engineer conference in New York (posted 29 Oct 2026), as reported by The Register, Dealroom, FourWeekMBA and Startup Fortune (3 Nov 2026). All cost and performance claims are Revolut's own and have not been independently verified.
No comments:
Post a Comment