ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

November 5, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Revolut says making self-hosted GLM-5.3 its default cut coding inference costs about 90% for roughly 2,000 engineers, while Apple has reportedly cut engineers' daily AI cap from $300 to $200. Today is about the two answers to a CFO who asks what AI costs: own the model or ration the seat. Neither figure has been audited.

By the numbers

  • About 90% Revolut's infrastructure lead says so, as relayed by The Register; GPUs and staff are not priced in.
  • $300 to $200 TechCrunch, citing Bloomberg, reports it; some allowances may fall to $100 a day.
  • 70% AWS says U Mobile's helpdesk agents route them by keyword match, the cheapest cost control in the post.
  • 4 rides Uber's CFO said so against about 320 million trips and orders; plan agent channels on that base.
Own it or ration it

Revolut says self-hosting GLM-5.3 cut its coding inference costs about 90%

Revolut has made self-hosted open-weight models the default for coding, and says that cut its inference costs by roughly 90% in under a year. The claim comes from David Abitbol, the fintech's ops and infrastructure lead, who spoke at the AI Engineer conference; The Register reported it on 3 November 2026. Abitbol said engineers on commercial tools had been spending $1,000 to $10,000 a month with little return, and that GLM-5.3, from Z.ai, now serves about 2,000 engineers on 56 Nvidia B200 GPUs.

Why it matters: Abitbol said flat, cheap tokens changed behaviour: engineers stopped rationing and handed the model debugging and Jira and Confluence tickets, which he called the biggest gains. Revolut's earlier route, strict budget limits, left engineers short of tokens. That is the same squeeze behind the daily caps described below.

The catch: The Register reports that self-hosting GLM-5.1 at $0.30 per million tokens cost about a tenth of Codex or Claude Code. That compares token prices, not total cost: the report gives no price for the 56 GPUs, power, hosting or the engineers who tune the cluster. It also gives no test of output quality against Claude Code and no control group for the 60% coding-time and 70% lead-time gains Abitbol cited. GLM is text-only, so vision queries still go to Gemini or GPT Luna.

What to do: Pull one month of tokens per engineer from your current coding-assistant bill. Price a self-hosted cluster against it, including an owner for it, then run 30 of your own tickets through the cheaper model before moving any default. Decide in advance which work, such as image reading, stays on a commercial model.

Our coverage →The Register: British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted ↗

Daily caps tighten

Apple has reportedly cut its engineers' daily AI spending cap from $300 to $200

TechCrunch reported on 4 November 2026, citing Bloomberg, that Apple lowered the daily AI spending cap for engineering teams in October from $300 to $200. Bloomberg's sources reportedly say some allowances will fall to $100 a day, and that outside engineering Apple is weighing a limit as low as $25. TechCrunch says it asked Apple for comment; Apple has not confirmed anything in the sources we can read.

The catch: This is a secondhand report. It does not say what the caps cover, whether they apply per person or per team, or how many staff are affected.

What to do: Work out your own number before a bill sets it: spend per person per day, by team, against what each team shipped. At $200 a day, one engineer could spend about $4,000 over 20 working days; that arithmetic is ours, not Bloomberg's. A cap limits the worst day, so set it per person per day and review it monthly.

Our coverage →TechCrunch: Apple is reportedly cutting budgets for its internal AI usage (citing Bloomberg) ↗

Seventy percent need no model

AWS says U Mobile's helpdesk agents answer 70% of requests without calling a model

An AWS post of 4 November 2026 describes AiDaH, a helpdesk built by Malaysian carrier U Mobile with AWS and partner Axrail. AWS says about 70% of requests match a keyword and run with no model call, 20% go to a lightweight classifier and 10% reach the full orchestrator. U Mobile's internal figures for January to September 2026 show provisioning and network tickets down 84.5% and billing tickets down 48.7%.

The catch: The post is vendor-published and gives no starting ticket counts, cost per request or error rate, so the percentages cannot be turned into hours or money. A person confirms any change to a live subscriber account.

What to do: Take 100 recent tickets from your service desk and mark how many a keyword rule could answer. That share is the cost you could skip before any model is called.

Our coverage →AWS Machine Learning Blog: How U Mobile built agentic telco operations with Amazon Bedrock AgentCore (vendor-published) ↗

What else

  1. 01TechCrunch reports Uber's CFO said 4 rides and 500 Uber Eats orders came through AI agents in Halloween week; Ars Technica reports DoorDash counts several hundred orders since launch. Our coverage →
  2. 02Moonshot AI says Kimi K3.1 Pro's open weights are free for internal use, but model-as-a-service businesses above $20 million revenue need a separate agreement. Our coverage →
  3. 03OpenAI says it will acquire Pierre Computer Company to improve how Codex builds and reviews code; the deal needs regulatory approval and no price is given. Our coverage →
  4. 04Collinear AI's public benchmark has DeepSeek V4.1 Flash passing 57.0% of 100 security-patching tasks at $0.04 per attempt, against 67.8% for Gemini 4 Argon. Our coverage →
  5. 05AWS says Smart Approvals, on by default in Amazon Quick, auto-approves eligible agent actions and sends the rest back to the user; check what counts as eligible before leaving it on. Our coverage →
  6. 06AWS says its open-source Deep Insight agent saves about 95% of tokens per script by layering context; a vendor figure, not an independent test. Our coverage →

Stat of the day

About 80%

The Decoder reports that about 80% of agent failures on CloudOpsBench came from misjudging a task's scope or stopping early, and about 20% from wrong actions, some irreversible. The University of Michigan team behind the 45-task AWS suite measured it; the best model solved 60% of tasks. The README says 42 tasks lack full controls. What it changes: before an agent gets delete rights, rehearse a decommission task in a sandbox account and check it removes everything it should and nothing else. The Decoder: New open-source benchmark reveals reasoning as the biggest bottleneck for AI in cloud operations ↗

Field note

An AWS post on a Postgres semantic layer for agents reports a concrete failure: without one comment on the order-total column, the agent multiplied it by quantity and reported $502,529 for the top customer instead of $27,991.66. The post says the agent reads business meaning from database comments and obeys column-level GRANTs. This is a vendor-published demonstration, not a field measurement. Before an agent queries finance data, check that every column it can reach has a comment.

One question

If tokens cost your engineers nothing at the point of use, what would you want them to hand to the model that they ration today?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →