Self-hosting an open coding model pays off when usage is heavy, not just when the team is big. GPUs bill around the clock and rent in 8-GPU nodes, about $39,070 a month each. At heavy usage ($600 per developer a month on the API), self-hosting costs less from about 113 developers. At Anthropic's average usage it barely breaks even, even at Revolut's 2,000 engineers. For a 25-person team it costs $55,045 a month against $4,875 on the API.
What Revolut reported
At AI Engineer NYC in October 2026, David Abitbol, who leads ops and infrastructure at Revolut, described moving about 2,000 engineers to a self-hosted GLM model from Z.ai, running on 56 NVIDIA B200 GPUs. The reported figures:
- About $0.30 per million tokens, roughly a tenth of what Revolut paid Anthropic and OpenAI.
- Some engineers had been spending up to $10,000 a month on AI tools.
- About 60% faster coding time and 70% faster lead time for changes, as reported by the company.
- Claude Code, Codex and Gemini are still used for some tasks.
These are company-reported numbers from one talk, covered by Dealroom and Runtime Wire, among others. Coverage differs on the exact GLM version, and reports mention months of tuning before the switch.
What $0.30 per million tokens leaves out
A cost per token from self-hosting is the GPU bill divided by the tokens served. That is the right number to compare with an API, as long as you remember what sits outside it:
- 01Idle hours
Rented GPUs bill 24 hours a day. Coding happens in working hours. A cluster sized for the busiest hour sits mostly idle at night and at weekends. How utilization moves the cost per token.
- 02Minimum size
B200s rent as 8-GPU nodes. A small team pays for a whole node however few tokens it uses.
- 03People
Serving, upgrades, monitoring and on-call. Revolut's savings were large enough to cover a team; a small company's may not cover one engineer.
- 04Quality
A model that needs more attempts or more review per task saves less than its token price suggests. Measure cost per finished task, not per token.
The math for your team
Assumptions: B200s at Lambda's on-demand $6.69 per GPU-hour, whole 8-GPU nodes running 730 hours a month, 36 developers per GPU (Revolut's ratio), $15,000 a month of staff, and 80% of usage moved, with the rest staying on the API. API spend is Anthropic's average of $13 per active day over 15 days, or $195 per developer a month.
| Developers | GPUs | All on the API | Self-hosted | Difference |
|---|---|---|---|---|
| 25 | 8 | $4,875 | $55,045 | $50,170 more |
| 100 | 8 | $19,500 | $57,970 | $38,470 more |
| 250 | 8 | $48,750 | $63,820 | $15,070 more |
| 500 | 16 | $97,500 | $112,639 | $15,139 more |
| 1,000 | 32 | $195,000 | $210,278 | $15,278 more |
| 2,000 | 56 | $390,000 | $366,487 | $23,513 less |
At Revolut's size the same assumptions give 56 GPUs, the number Revolut reported. At average usage, though, self-hosting saves only $23,513 a month there, about 6%. A 90% saving implies far heavier use than the average developer, which fits engineers spending up to $10,000 a month, and probably GPU prices well below on-demand rates.
Notice that bigger isn't always cheaper. At average usage self-hosting first costs less at about 848 developers, then costs more again at 1,000, because the next node has to be added before it is full. Without the staff cost, the first crossing comes at about 251 developers.
Usage is the real lever. At heavy usage, $600 per developer a month ($30 a day, the level Anthropic says 90% of users stay under), self-hosting costs less from about 113 developers, and the saving grows with every developer after that.
| Developers, heavy usage | GPUs | All on the API | Self-hosted | Difference |
|---|---|---|---|---|
| 100 | 8 | $60,000 | $66,070 | $6,070 more |
| 250 | 8 | $150,000 | $84,070 | $65,930 less (44%) |
| 500 | 16 | $300,000 | $153,139 | $146,861 less (49%) |
| 2,000 | 56 | $1,200,000 | $528,487 | $671,513 less (56%) |
Try your own numbers in the self-hosting calculator. Usage figures: Anthropic; GPU price: Lambda.
When self-hosting is worth it
- Heavy, steady usage, such as agents running for hours, where per-token bills are large.
- Enough developers to fill whole nodes, at least a hundred or so heavy users, kept busy through the working day.
- Code that can't leave your infrastructure, where cost is not the main reason anyway.
- A team that already runs GPU infrastructure, so the people cost is shared.
Below that, the cheaper move is usually a cheaper model, not your own hardware.
What to do before you buy GPUs
- 01Measure spend per developer
Averages hide the few people who drive most of the bill. Track Claude Code cost per developer with OpenTelemetry first.
- 02Re-price your real usage
Ganivra's Coding agents view shows what your recorded tokens would cost on Sonnet, Haiku, GLM-5.3 or your own GPUs at a rate you enter. Once GPUs are running, split the GPU bill into a cost per call.
- 03Try the open model through an API
GLM-5.3 runs on hosted per-token APIs. Test quality on real tasks before committing to hardware.
- 04Move the simple work first
Revolut kept frontier models for some tasks. Routing easy work to a cheap model captures much of the saving with none of the hardware.
Self-hosting AI coding models: frequently asked questions
Did Revolut really cut AI coding costs by 90%?
That is Revolut's own figure, presented by David Abitbol, who leads its ops and infrastructure, at AI Engineer NYC in October 2026 and reported by several outlets. It hasn't been independently verified, and reports differ on which GLM version is in use.
How many GPUs does a self-hosted coding model need?
At least one 8-GPU node: B200s are rented that way, and NVIDIA's own recipes run a GLM-5 model on 4 B200s per copy. Revolut reported 56 B200s for about 2,000 engineers, roughly 36 engineers per GPU.
What does a B200 cost to rent?
Lambda lists $6.69 per GPU-hour on demand for 8-GPU instances, which is $39,070 a month for one node running around the clock. Other providers range from about $6 to $8 on demand, and reserved capacity costs less.
Is a self-hosted open model as good as Claude Code or Codex?
For some tasks it may be. Revolut kept Claude Code, Codex and Gemini for some work and reported months of tuning before the switch. Measure it on your own tasks: a cheaper model that needs more attempts per task saves less than its token price suggests.
What is the cheapest way to try an open coding model?
Use one through a hosted per-token API first, such as GLM-5.3 on Together AI or Fireworks. You pay only for tokens, with no GPUs to run, and can measure quality on real work before committing to hardware.
Before you decide
See what your own usage would cost on another model.
Send Claude Code's usage to Ganivra and re-price it on cheaper models or your own GPUs. 14-day free trial, no credit card required.