Revolut says it cut AI coding costs 90% with GLM. Here's what the math shows

What this is about

British fintech company Revolut said at the AI Engineer conference in New York (October 2026) that it cut the cost of AI-assisted programming by about 90% in less than a year. The reason: self-hosted GLM models from Chinese lab Z.ai recently became the default, while Claude Code and OpenAI Codex are kept only for exceptions. The Register and Techmeme picked up the figure on November 3, and it has circulated as a headline ever since. We look at what's behind it and when something like this pays off for other teams.

What Revolut actually says

According to The Register (Tobias Mann) and a summary from maccus.fr:

• Timeline: Gemini was introduced in early 2025, Claude Code in early 2026 with strict per-developer budget limits. GLM 5.1 ran self-hosted from April 2026, and GLM 5.3 is now being rolled out. The setup has reportedly been stable since August.
• Scale: about 2,000 developers, 56 Nvidia B200s, i.e. seven eight-GPU systems (DGX B200).
• Price: about $0.30 per million tokens as a blended price (input and output), with flat-rate billing. That's roughly a tenth of what Codex or Claude Code cost.
• Tech: FP4 quantization (NVFP4; numbers are stored with just 4 bits) on Blackwell chips, separating prefill (reading the input) from decode (generating text), and spreading the model across eight GPUs instead of four to leave room for the KV cache. Time to first token fell from 14.26 to 1.68 seconds, and requests per second rose by 76%. The target: under 3 seconds at around 256,000 tokens of context and about 200 tokens per second.
• Results (self-reported): With Gemini, coding time fell by 20%; with GLM, by 60%. Lead time fell by 70%. According to The Register, the biggest benefit is in debugging and handling Jira and Confluence tickets, rather than pure code writing.

Revolut's David Abitbol described the effect of cheap tokens this way: "When you have unlimited tokens, you're free to do that." Developers no longer have to ration them.

What was compared, and what wasn't

The 90% is not a like-for-like measurement. According to maccus.fr, the figure comes from the headline and press coverage. The talk itself speaks of "about a tenth" of the earlier Anthropic and OpenAI spend. Open questions:

• Baseline: Apparently the earlier bill from commercial providers is being compared with the per-token cost of running the model in-house. Nothing says whether these are task costs (that is, what a solved ticket costs). Token prices tell you little if a model needs more steps or more output per task.
• Not included: According to maccus.fr, the cost of the team that runs tuning and monitoring is missing. Revolut doesn't say whether the GPUs are bought or rented. The organizers stress that the 30-cent figure reflects Revolut's own utilization and can't simply be transferred to other setups.
• Utilization: High utilization pushes down the token price. Reportedly, developers get around 800 tokens per second after 5 p.m. That suggests spare capacity in the evening, so compute isn't fully used around the clock.
• Quality gains: The 60% and 70% improvements are internal, self-reported and, according to maccus.fr, not isolated from other changes.

Our recalculation

The following figures are our own model assumptions, not Revolut data. According to aggregators, B200 rental prices vary widely: from $3.69 per GPU-hour (31 providers, October 2026), with a median of $6.82. Hyperscalers go up to about $9. Some sources are affiliate blogs, so treat these as rough guides only.

(56 GPUs × 8,760 hours = 490,560 GPU-hours.) At $0.30 per million tokens, that works out to roughly 6 to 11 trillion tokens per year, or roughly 8 to 15 million tokens per developer per day. That's plausible for agents that repeatedly re-read large contexts, since they consist mostly of (cached) input. But it shows that the low price depends on extremely heavy usage. Had only half as many tokens flowed, the price would be twice as high.

For comparison, list prices (from aggregators; we didn't check official pages):

Important: even the GLM-5.3 API is roughly 3.5 to 5.7 times cheaper than Opus at list prices. So part of the 90% may come from switching models alone, not from self-hosting. We don't know whether Revolut compared against list prices or negotiated terms.

When does self-hosting pay off?

A rough rule of thumb (our assumptions): a rented eight-GPU node costs about $0.26 to $0.48 million a year. Revolut spreads the model across eight GPUs. According to The Register, the weights need about 380 GB, plus 6 to 26 GB of context per user. On top of that come operations and tuning, realistically at least one to two full-time positions. If we put that at $0.2 to $0.4 million, fixed costs come to about $0.5 to $0.9 million per year, or $40,000 to $75,000 per month.

Revolut's ratio (seven nodes for 2,000 developers) suggests that one node serves around 280 developers. Per head, that's $15 to $30 a month for hardware, plus a share of operations. Against that stand agent bills on Claude, which vary widely. We have no reliable source for them. If you spend $150 to $300 per developer per month, you hit break-even against Claude at roughly 150 to 400 active users. Against Z.ai's GLM API ($1.40/$4.40), the threshold is much higher because it's already cheap. There, self-hosting usually only makes sense for privacy, latency or compliance reasons.

Our assessment: below about 100 developers, the API almost always wins. Between 100 and 300, it depends on utilization and staffing. Only above that does a dedicated high-throughput node become a cost advantage.

Quality: what independent benchmarks offer

Independent measurements are thin. We found no Epoch AI result for GLM 5.3 on SWE-bench Pro. What we did find:

• Z.ai's own "Code Bench": GLM-5.3 scores 31.4% at about 50,000 output tokens per task, against 29.5% for Opus 4.8 at 120,000. According to a commentary page, nobody has reproduced this independently.
• A Vals AI comparison page (via a search preview) shows GLM 5.3 ahead of Opus 4.7 on SWE-bench, with a gap of up to 13.4 points. It's unclear which variant was used, and the page was unreachable when we tried to open it.
• An aggregator, citing Vals AI, lists about 97% for Opus 5 and 95.4% for GLM-5.3 on SWE-bench Verified. At those scores the test is practically maxed out and says little about real work.

Bottom line: it's plausible that GLM 5.3 is a usable coding agent, but parity with Claude's top model isn't independently established. Revolut itself treats the frontier models as "the exception." GLM handles text only, so image requests go to Gemini or GPT, with a router distributing traffic.

Security and compliance

Here our earlier reports on GLM 5.3 fit the picture. Epoch AI used the model to build working attacks on six Firefox versions, at roughly $1,000 in API costs each. Irregular saw successful multi-step attacks in simulations and "reward hacking" (the model reproduced code from memory even though it wasn't supposed to). Anthropic and Epoch report that safeguards are easy to bypass.

For self-hosting, this means:

• There are no vendor guardrails. You have to build misuse controls, permissions and logging yourself. Agents belong in sandboxes with no access to production data.
• Reward hacking means an agent can "pass" tests without solving the task. Code reviews and tests the agent can't modify are a must.
• Privacy: Abitbol sees the main risk in APIs hosted outside China. With on-prem or regional hosting, he says, that is largely mitigated. That's his assessment, and we haven't verified it. We also haven't checked the GLM 5.3 license terms.

Our take

The figure is neither wrong nor proven. It's a plausible real-world number from a company with 2,000 developers, high utilization and its own infrastructure team. As proof that "self-hosting saves 90%," it doesn't hold up for small teams. A considerable share of the advantage probably comes from the cheaper model itself, and the rest from economies of scale in token volume. No source mentions the costs of staff, capital and outage risk. The best takeaway: self-hosting pays off when you have lots of tokens, little idle time and sensitive data. Otherwise, the GLM API or a hosting provider is the cheaper stepping stone. And you have to measure quality and security yourself, on your own code.