A house newsletter · daily · researched and written by the Kurzia AI
Tuesday 3 November 2026 · Issue 42 · 5 min read
Anthropic's amended S-1 lands Q3; Musk says Grok Bot loses money; Goldman pulls Claude Code from traders
What's hot
Anthropic's revised filing shows growth cooling as costs climb. Anthropic filed an amended prospectus ahead of its listing, and this version finally includes the third quarter that the first one left out: roughly $17bn of revenue, a net loss of $23.3bn, and $895bn of future commitments. The fear had been that Q3 was being hidden because it was ugly; it isn't a disaster, but the shape changed. Where earlier quarters roughly doubled, Q3 grew about 47% on the quarter, while spending on compute and infrastructure rose 77% against Q2.
This is the number your clients will quote at you. The useful read is not the loss, which is mostly accounting on convertible notes, but the widening gap between what agent usage earns and what it costs to serve — the same gap that moves prices, rate limits and plan terms under your delivery commitments. Readers of the filing noted Q3 contained only one week of Opus 5.5, from 22 to 30 September, so the economics of the current model barely show up yet. explainx.ai
New tools & products
Message threads — is a Claude API feature, now in public beta, that lets an agent send only the newest turn instead of replaying the entire conversation on every request. Long coding sessions pay a waiting tax before the model even starts answering; Anthropic measures that wait dropping about 35% on the largest requests, with no change to price or rate limits. Threads live in memory only and lapse five minutes after the last call, so your own client still has to hold the full conversation. claude.com
Budgets — lets an enterprise admin split Devin's spending down the org chart, up to three levels deep, and see what each team, project or engineer burns. This is the FinOps conversation moving inside the agent itself — worth knowing before a client asks who owns the meter on your delivery. If a budget empties while Devin is working, the session stops; admins can also auto-raise an engineer's ceiling when their Efficiency Score looks good. The New Stack
Uber MCP Gateway — is Uber's single entry point for agents to find and call internal services, exposing existing APIs as agent-callable tools without rewriting those services. It is the clearest large-scale reference you can point a client at when they ask how to govern agent tool access instead of letting every team wire its own. It fronts more than 800 MCP servers and 5,000 tools, and automatically discovered tools stay switched off until the owning team turns them on. InfoQ
Codex closed 149 security holes in stable Linux. Security firm Trail of Bits put OpenAI's Codex on thankless work — carrying fixes from current Linux code back into the older stable releases most machines actually run — with 149 vulnerabilities patched by 29 October. A clean story for a sceptical client: agents grinding through maintenance nobody funds, with a human engineer reviewing and signing every patch before it goes to a maintainer. Their own tally finds 2,006 known flaws still unfixed in the 5.10 kernel branch, and a simple backport now costs roughly half an hour of an engineer's time. Trail of Bits
Momentic — is a testing agent, launched publicly this week, that drives your application like a user and reports the bugs it hits; the tests themselves are plain-English YAML files kept in your repository. Once agents write the code, knowing it works is the bottleneck — this is the category clients will start shopping for next. When the interface changes, it finds the control again mid-run and proposes an edit to the test file for a human to approve. Momentic
Claude Opus 5.5 — produced permit-quality construction drawings for a simple house in a working architect's test of 13 models, writing Python to generate the linework rather than driving any CAD package. The best non-software evidence this year that the supervise-the-agent pattern is spreading into other licensed crafts — useful with a client who still thinks this is a developer story. The drawing set took twelve hours, cost a few hundred dollars in compute, and ran to over 15,000 lines of Python. Stephen Gallagher
What people are saying
A thread argues tech layoffs trace to AI capex, not visas. A thread about jobless tech workers being shut out of San Francisco's AI boom — the article opens with dozens of them hiking Mount Tamalpais on a Monday morning — turned into an argument about cause. The loudest reading was that the layoffs follow from the industry pouring its free cash into AI infrastructure, with commenters dismissing H-1B and offshoring as decades-old factors being blamed in bad faith. Others pushed back that the city's cost of living does most of the damage, and that this is a narrow slice of people to mourn.
This is the mood of the senior engineers your clients are asking to adopt agents, and it is hardening. Expect resistance framed as caution, and plan the human side of a rollout accordingly. One commenter notes that roughly 90% of the Bay Area doesn't work in tech and was left out of every previous boom too; another claims most 2026 software was already written, reviewed and tested by AI, which the thread does not settle. bloomberg.com · from the Hacker News discussion
News
Anthropic's own Claude use outgrew its payroll. Anthropic's IPO paperwork shows that the value of the Claude its own staff consumed passed its monthly employee pay bill back in April — measured at the prices customers pay, not cash going out the door. It is the sharpest anecdote available for the argument that tokens are becoming a line in compensation, and a handy reference point when a client baulks at a per-seat agent budget. Internal usage grew more than 250-fold between January and September, and over 90% of the code added to Anthropic's software in September was written by Claude. Business Insider
Grok Bot — xAI tightened weekly allowances on its background agent after users woke to exhausted quotas, and Musk confirmed on X that it was deliberate rather than a glitch: serving it costs more than it brings in. A rare public admission that agent economics don't clear yet — good cover when you advise a client against fixed-price commitments that sit on somebody else's meter. About 800,000 people were paying weekly for it, and one user reported hitting 100% of their allowance with four days still to run. TechCrunch
“The cost of serving Grok Bot was unsustainably high relative to revenue”
Goldman Sachs — restricted GitHub and Claude Code to designated software engineers and quantitative 'strats' under guidance issued in October, taking the tools away from bankers and traders who had been building their own software. This is phase two in regulated industries — enthusiasm first, then a hard narrowing of who may run an agent — and it is the precedent your financial-services clients will cite. In July the bank's CIO was publicly praising front-office staff who assembled working tools before handing them to its 12,000 developers for refinement. explainx.ai
MOFCOM Announcement No. 48 — is China's first proposed export control on frontier AI, covering model weights, agent architectures and AI-assisted vulnerability hunting, with a pre-export registration step already in force while the catalogue is still a draft. Custody is now a legal question as well as a security one: calling a Chinese model's API is outside scope, but self-hosting its weights on your own infrastructure abroad counts as an export and needs a licence. 23 of the 40 listed technologies are AI items, and comments on the draft close on 15 November 2026. Cloud Security Alliance
Garry Tan — Y Combinator's chief executive says he personally spends roughly $100,000 to $200,000 a year on tokens, and has let his agents read his email, his Dropbox and twenty years of therapy notes. Executives hear this and immediately ask what the real per-head cost of going all-in would be; Tan's own answer is that price is the thing keeping personal agents out of the mainstream. He has published his 23 Claude Code skills as gstack on GitHub, and YC runs an internal agent called Stray Light loaded with its founder history. TechCrunch
Weight-Space Programming — is a set of experiments in which a frontier model reads a neural network's raw weights the way it reads source code, says what the network was trained to do, and then edits those weights directly to fix a flaw — with no training run and no gradients. Early research rather than anything you can buy, but it is the sort of result that lands in client decks fast, so it helps to know what it actually claims. In one case it altered 0.016% of a small Qwen model's weights and its maths and instruction-following improved sharply. rikiyat.github.io
LongCat-2.5 — is a newly announced model pitched at agentic work, with handling of text, images and other inputs built in rather than bolted on afterwards. Worth a line on your model-options list, since the releases aimed at tool-using agents rather than chat are where the competitive pressure now sits. The announcement page is bare — a heading and little more — so hold judgement until access terms and independent reports appear. longcat.ai
One takeaway
Name the budget owner before you name the stack. The week's stories rhyme: a vendor admitting the agent loses money per heavy session, a coding agent shipping department-level spend caps, a bank deciding only engineers may spend, and a lab whose own usage dwarfs its wage bill. Capability is no longer the constraint; the meter is. In your next engagement, write down who owns the budget, what happens when it runs dry mid-task, and who is allowed to raise it — before anyone argues about models. The New Stack