AI
“Revolut reduced its AI coding costs by 90% after switching from OpenAI and Anthropic APIs to a self-hosted Chinese AI model (Zhipu AI's GLM-4.5)." Caption adds: 56 Nvidia B200 GPUs, roughly $0.30 per million tokens, response times from over 10 seconds to 1.6 seconds, 60% improvement in coding time and 70% reduction in lead time across a…”
Plain restatementRevolut says that after making a self-hosted open-weight model from the Chinese lab Z.ai (formerly Zhipu AI) its default for AI coding, instead of relying primarily on OpenAI and Anthropic services, its AI coding inference cost fell by about 90%, to roughly $0.30 per million tokens, running on 56 Nvidia B200 GPUs, with self-measured productivity and latency improvements.
This is a real story that has been garbled in one important detail. Revolut's AI operations lead did say on stage at the AI Engineer conference in New York on October 13, 2026 that running an open-weight Chinese model on its own 56 Nvidia B200 GPUs cut its AI coding inference cost to about 30 cents per million tokens, roughly a tenth of what it had been paying OpenAI and Anthropic. However, the model is GLM-5.3 from Z.ai, formerly Zhipu AI, after starting with GLM-5.1 in April 2026, not GLM-4.5 as the post says. The post also implies the jump from 10 seconds to 1.6 seconds response time came from leaving the American APIs, when the talk attributes it to tuning the company's own serving setup on the same hardware. The 90 percent figure covers token costs only and leaves out GPU, cloud, staffing and monitoring costs, which Revolut has not disclosed. All the cost and productivity numbers, including the 60 percent coding time and 70 percent lead time improvements, come from Revolut itself with no published measurement method and no independent check. Revolut also did not drop the American tools entirely; Claude Code and Codex remain available for exceptions, and image work still goes to hosted models.
Revolut [drifted from the evidence:] reduced its AI coding [drifted from the evidence:] costs by 90% after switching from OpenAI and Anthropic [drifted from the evidence:] APIs to a self-hosted Chinese AI [drifted from the evidence:] model (Zhipu AI's GLM-4.5)." Caption adds: 56 Nvidia B200 GPUs, roughly $0.30 per million tokens, [drifted from the evidence:] response times from over 10 seconds to 1.6 seconds, 60% improvement in coding time and 70% reduction in lead time across a sample of 2,000 engineers, Gemini still used for image tasks, and [drifted from the evidence:] that Revolut is "Europe's largest private company.
Revolut [added by the neutral restatement:] says that after making a self-hosted open-weight model from the Chinese lab Z.ai (formerly Zhipu AI) its [added by the neutral restatement:] default for AI coding, [added by the neutral restatement:] instead of relying primarily on OpenAI and Anthropic [added by the neutral restatement:] services, its AI [added by the neutral restatement:] coding inference cost fell by about 90%, to roughly $0.30 per million tokens, [added by the neutral restatement:] running on 56 Nvidia B200 GPUs, with self-measured productivity and [added by the neutral restatement:] latency improvements.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Revolut's AI operations lead did publicly state that making self-hosted open-weight models the default for coding cut AI coding inference cost by roughly 90%, to about $0.30 per million tokens.
- The model is from Z.ai, formerly named Zhipu AI, a Chinese lab, and the weights are open, so "self-hosted Chinese AI model" is correct as a description.
- The hardware figure is correct: about 56 Nvidia B200 GPUs serving roughly 2,000 engineers.
- The productivity figures are relayed accurately as Revolut reported them: about 60% improvement in coding time and about 70% in lead time, on a sample of roughly 2,000 engineers, after the April 2026 rollout.
- The post is right that frontier hosted models were not abandoned entirely and that image-related work still routes to models such as Gemini.
- The time-to-first-token figures, over 10 seconds down to 1.6 seconds, do appear in accounts of the talk.
- The post names Zhipu AI's GLM-4.5. The conference's own transcript names GLM 5.3, and The Register says Revolut started with GLM 5.1 in April 2026 and later moved to GLM 5.3. GLM-4.5 is a different, earlier and smaller model. A reader trying to replicate the setup would be pointed at the wrong model, and the named model is the one concrete technical identifier in the claim.
- Misattribution of the latency baseline: The post places "response times improved from over 10 seconds to 1.6 seconds" in a sentence about switching away from OpenAI and Anthropic. In the talk, the 10-second baseline was the untuned self-hosted deployment, and the 1.6 seconds came from infrastructure tuning, context capping at 256k, NVFP4 quantization and parallelism changes, on the same GPUs. As written it reads as though the Chinese model was faster than the hosted APIs, which is not what was reported.
- The 90% figure covers inference token cost only. GPU hardware, cloud, staffing and observability costs sit outside it, and Revolut has not disclosed the capital cost of 56 B200 GPUs. The headline "AI coding costs" invites readers to hear total cost.
- Every number here is a single company's self-report from a conference stage, with no published measurement methodology and no independent verification. The post presents them as established outcomes.
- The 70% lead-time improvement is attributed in the post to the model switch. In the talk the largest gains were attributed to work outside coding, such as debugging and ticket handling, which engineers began automating once tokens were no longer rationed. The lever described is removal of budget caps as much as the change of model.
- Exaggeration of the switch: "Switching from OpenAI and Anthropic APIs" overstates it. Claude Code and Codex remain available as exceptions; what changed is which model is the default.
- Imprecise descriptor: Revolut is widely reported as Europe's most *valuable* private company, valued at $115B in a July 2026 secondary share sale. "Largest" is a different and unsupported measure.
- Whether the 90% cost reduction holds after GPU, infrastructure and staffing costs are included. No figures for those were published.
- How coding time and lead time were measured, and what share of the 60% and 70% improvements is attributable to the model versus to removing token budget caps versus to other changes over the same period. No methodology was published.
- Whether Revolut ever ran GLM-4.5 at any earlier point. The talk record and The Register describe GLM-5.1 and GLM-5.3 only, and the "4.5" label appears to enter through secondary write-ups.
- Whether Revolut has made any statement in its own channels. No Revolut press release or engineering blog post on this was found; the only company statement located is a quote given to The Register.
The underlying event is real and well documented. David Abitbol, who leads AI operations and infrastructure at Revolut, gave a main-stage talk at AI Engineer NYC on October 13, 2026. The conference's own transcript page states that open models became Revolut's default for coding with frontier services kept for exceptions, that 56 B200 GPUs serve roughly 2,000 engineers, and that the blended cost is near $0.30 per million tokens, around a tenth of the prior Anthropic and OpenAI invoice. The transcript line naming the model is explicit: "And that's to host GLM 5.3." The Register's November 3 account, which includes direct quotes, says Revolut deployed Gemini in early 2025 and Claude Code in early 2026 under strict budget caps, then began self-hosting GLM 5.1 in April 2026 at $0.30 per million input/output tokens, about a tenth of the cost of Codex or Claude Code, and has since moved to GLM 5.3. Abitbol reported a 60% improvement in coding time and a 70% improvement in lead time across roughly 2,000 engineers, and said the largest gains came from work outside the editor such as debugging and Jira and Confluence tickets. On the latency figure: the drop from over 10 seconds to 1.6 seconds time-to-first-token was achieved by tuning the self-hosted stack, capping context at 256k, quantizing to NVFP4 on Blackwell, and tuning parallelism, on the same 56 GPUs, with requests per second rising from 3 to over 7. The baseline of "over 10 seconds" describes the untuned self-hosted deployment, not the previous hosted APIs. On the model version: GLM-5.3 was released by Z.ai on August 14, 2026, with roughly 753 billion total parameters and about 40 billion active. GLM-4.5 is an earlier, smaller model from 2025. StartupHub.ai's October 29 write-up named the model "GLM 4.5," and Startup Fortune, the source the Instagram post credits, repeated that and additionally placed the talk at "GTC Berlin on October 22" with a different talk title. The conference's own record and The Register both place it at AI Engineer NYC on October 13 and name the GLM-5.x models.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.