If you pay list price for AI coding tokens, you are paying for your vendor's margin

Some Revolut engineers were running up AI coding bills of $1,000 a month, and some $10,000. David Abitbol, the fintech's ops and infrastructure lead, told the AI Engineer conference in New York that the company saw "not so much ROI" for that money. So it rationed tokens. Then it worked out what tokens cost when nobody else sets the price. Self-hosting Z.ai's GLM-5.1 came to about $0.30 per million tokens, roughly a tenth of Codex or Claude Code. Revolut says that in less than a year its AI coding inference bill fell by about 90 percent, across roughly 2,000 engineers, served from 56 Nvidia B200 GPUs.

I think every procurement lead with a frontier coding contract on their desk should read that as an invoice audit. If you are paying list price, part of what you pay is Anthropic's or OpenAI's margin, and you cannot say how large that part is, because you have never priced the alternative. Claude Opus 4.8 lists at $5 per million input tokens and $25 per million output, which comes to a blended $9.00 at an 80/20 split with no caching. Revolut is paying $0.30 on hardware it controls. Before you sign or renew, I would insist on a self-hosting benchmark: what your actual workload would cost on an open-weights model running on your own GPUs or a regional cloud.

I expected the saving to be the headline. What surprised me was what the saving bought. Under budget limits, Abitbol said, engineers hoarded tokens for the end of the quarter and the hardest tasks. Once billing went flat, the biggest gains came from handing models the dull work: debugging, Jira tickets, Confluence pages. Metered frontier pricing charges you for the tokens you use, and you pay a second time in the work your engineers never hand over because they are watching the meter. Revolut reports that Gemini alone in early 2025 improved coding time by 20% and lead time by 10%; self-hosted GLM from April took those to about 60% and 70%. The conference's own write-up warns that this internal comparison does not separate the model change from other causes. I accept that. My argument rests on price, and the price gap does not depend on the productivity figures.

The vendors have a decent reply. List price is a straw man, they would say, since Anthropic advertises savings of up to 90% with prompt caching and 50% with batch processing. Prices are already falling: Opus 5.5 arrived on 22 September at $4 and $20, 20% below Opus 4.8. Self-hosting also leaves Revolut owning the operational and regulatory controls a vendor would otherwise carry, which is no small thing for a regulated fintech. And supply is shaky: an NYU Shanghai RITS piece on GLM-5.3, the model Revolut now serves, is headlined with Z.ai holding the weights back.

All of that is true, and I think each point strengthens the case for the benchmark. Caching discounts and 20% price cuts are what frontier labs offer when buyers can point to a credible alternative. You only get them by showing up with your own number. Caching also cuts the other way: one pricing analysis warns a new Anthropic tokenizer can raise effective costs by up to 35%, so the discount you negotiate can leak away without a line on the rate card changing. On supply, GLM-5.1 was released on 7 April under the MIT licence. Weights you have already downloaded do not get recalled, and Z.ai claims a SWE-Bench Pro score of 58.4 for it against 57.3 for Claude Opus 4.6 and 57.7 for GPT-5.4. Treat that as the vendor's claim, but even a model somewhat behind the frontier is good enough for ticket triage.

The operating burden is the objection I take most seriously. Revolut's tuning slide shows 76% more requests per second and 88% lower time to first token on the same 56 GPUs, which is the work of a capable infrastructure team, and a company with 200 engineers will not have one spare. But a benchmark needs a spreadsheet and a cloud quote, not a team, and it gives you a number to bring to the negotiation.

Revolut also never dropped the frontier labs entirely. GLM is text only, so a router sends vision and UI work to Gemini or GPT Luna, and Abitbol describes frontier models as the exception at Revolut. That is the contract shape I would push for: a small frontier commitment for the jobs open models cannot do, and the bulk of routine tokens priced against a self-hosted quote.

So before your next Claude Code or Codex renewal, pull three months of token logs, split them between routine work such as debugging and ticket handling and the vision or hard reasoning jobs, and get a written quote for serving MIT-licensed GLM-5.1 at that routine volume. Put it in the procurement file next to the vendor's offer. If the vendor's post-caching rate is still a multiple of that quote, cap the frontier seats and move the routine tokens.

Prompted by British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted, theregister.