First Deploy AI

For engineering leaders

Should our team self-host an open model instead of paying for Claude Code?

Short answer: Probably not yet. Revolut’s headline is real but it compares inference prices for about 2,000 engineers, not total cost for a 30–200 person team. Most teams get the first savings without GPUs: route simple work to cheaper hosted models, cap runaway sessions, and measure cost per finished task. Start with the $150 burn-audit hour.

What Revolut reported

Per The Register’s November 3, 2026 report, Revolut’s ops and infrastructure lead told the AI Engineer conference that the company cut AI coding inference costs by roughly 90% in under a year by making self-hosted open models its default. It serves GLM-5.3 to roughly 2,000 engineers on 56 Nvidia B200 GPUs, and self-hosting GLM-5.1 at about $0.30 per million input and output tokens cost about a tenth of Codex or Claude Code. Computing’s coverage adds that some engineers had been spending $1,000 to $10,000 a month on commercial tools and that the company first tried strict spending limits.

Two details matter for everyone else. GLM is text-only, so Revolut still routes vision work to other models through a router. And the reports give no total for the GPUs, hosting or the people who tuned and run the cluster, so the 90% is an inference-price comparison, not total cost of ownership.

Why a 30–200 person team is different

  • Scale. A GPU cluster is a fixed cost. It pays off when thousands of people keep it busy. A team of 40 developers rarely does.
  • People. Someone has to serve, tune, patch and monitor the model. That is a job, not a weekend project.
  • Quality. A cheaper model that needs more retries can cost more per finished task than an expensive one that gets it right.
  • The real lesson is routing. Revolut didn’t move everything. It kept specialised work on commercial models and sent the bulk to the cheap one.

Start with the $150 burn-audit hour. One hour with an operator. You leave with a written map of the jobs, tokens and flows to cut or rebuild. Pay the hour, then we schedule the session. See a sample audit (example data) or try the free burn check.

Book the hour — $150 Operator pack — $625 depositAI Concierge — $2,000/mo

The pack is $1,250 for 10 hours; checkout takes the $625 deposit. Questions first? Email daniel@agenthiveinc.com or call our AI front desk demo at (320) 335-6186.

What to try first

  1. Find your heaviest users and sessions. Anthropic’s Claude Code cost guide explains how spend varies by person and how teams can see it.
  2. Pick three routine task types, such as search, small edits and test fixes, and run them on a cheaper hosted model for a week.
  3. Compare cost per merged change, not cost per token.
  4. Set spend limits so a runaway session can’t eat the month while you test.

If routing on hosted models already gets most of the saving, you’ve avoided a hardware project. If it doesn’t, you now have the numbers to price self-hosting honestly.

The three ways to work with us

  1. $150 burn-audit hour. One working hour. We map the agents, routines and flows your team runs and hand back a written list of what to cut, fix or rebuild.
  2. $1,250 operator pack ($625 deposit). Ten hours of build and fix work on the flows from the audit. The balance is due as the work is scheduled.
  3. AI Concierge, $2,000/mo. A senior AI operator on retainer: two 45-minute working sessions a month, unlimited async help between them, and a shared asset inventory. Month to month; cancel any month after the first 30 days.

Common questions

Is the 90% figure audited?

Not that we could find. It comes from Revolut’s own talk as reported by The Register and Computing. Treat it as one company’s inference-price comparison.

Can you set up a self-hosted model for us?

That is outside these offers. We help you measure what you spend now and test cheaper hosted models first, so any hardware decision is based on your own numbers.

Will cheaper models hurt code quality?

They can on hard tasks. That is why we test per task type and keep the best model for planning and difficult work.

Start with the $150 burn-audit hour. One hour with an operator. You leave with a written map of the jobs, tokens and flows to cut or rebuild. Pay the hour, then we schedule the session. See a sample audit (example data) or try the free burn check.

Book the hour — $150 Operator pack — $625 depositAI Concierge — $2,000/mo

The pack is $1,250 for 10 hours; checkout takes the $625 deposit. Questions first? Email daniel@agenthiveinc.com or call our AI front desk demo at (320) 335-6186.

Related

Outside facts link to their sources and were checked in November 2026; product pricing and limits change, so confirm on the vendor’s page. First Deploy is built and run by Daniel Graham at AgentHive Inc, Palm Coast, Florida. We don’t promise a savings figure before we’ve seen your setup.