A house newsletter · daily · researched and written by the Kurzia AI
Tuesday 3 November 2026 · Issue 5 · 4 min read
Grok's cost confession, ChatGPT merges its tabs, and AI QA goes GA
The one thing to know
Grok Bot — Users woke on Halloween weekend to weekly allowances sitting at 100% even though they had barely touched the agent. Musk confirmed on X that this was deliberate, not a glitch: the product costs more to run than it earns. He offered no price change, no reset and no timeline.
It is the first blunt public admission that a flagship agent feature is underwater on subscription pricing — expect caps, metering and shared allowances to become the norm, and model what your heaviest agent users cost you before copying any unlimited-agent offer. Grok Bot has roughly 800,000 paid weekly users and no standalone price; it is bundled into plans such as SuperGrok at $30 a month. Some users watched usage leap from 50% to 100% overnight. TechCrunch
“The cost of serving Grok Bot was unsustainably high relative to revenue”
New ground broken
Anthropic's IPO filing — The filing shows that in April, staff Claude usage valued at retail prices passed Anthropic's monthly employee compensation, excluding stock — after internal consumption grew more than 250-fold between January and September. Treat AI consumption as a budget line that sits beside headcount: your team's agent bill will be measured, contested and capped, so plan the token budget, not just seats. In September, Claude produced over 90% of the new code appearing in Anthropic's own software. Business Insider
Agent features and products that impact product management
ChatGPT — An email to users says the Chat and Work tabs become one surface, tested on a small group first: describe an idea, bring files and plugins, and ChatGPT builds the site or deck while you steer. Codex stays a separate area. The market is converging on one entry point where the model decides what is a quick answer and what is a multi-step job — and one meter for everything, which heavy users will feel. The Work tab launched in July and is retiring about four months later; the fast Instant model stays unlimited and off the meter. TechCrunch
Momentic — Its agent Mo is now generally available: it exercises your app end to end from a prompt, keeps tests as plain-English YAML in your repo, and re-finds controls when the UI shifts, proposing test-file repairs for review. A second agent reproduces each suspected bug before Mo reports it. With coding agents accelerating output, verification is the bottleneck; QA that plugs into the agent loop over MCP is the pattern to evaluate this quarter. Quora's Poe team cut daily test execution from seven hours to 30 minutes and replaced 500+ manual cases; the enterprise tier has SSO and SOC 2 Type 2. Momentic
DeepSeek Harness — A new "creator mode" lets users build custom plugins in minutes — Bloomberg cites a 3D model renderer and a video editing workbench — which DeepSeek pitches as a way to handle complex tasks. Plugin authoring is being commoditized: if your product exposes an agent surface, users will expect to build their own tools without a developer in the loop. It is a Bloomberg report surfaced via Techmeme; no pricing or rollout details were given. Bloomberg via Techmeme
How product teams work now
PM OS — A repo holding 31 agent skills, grown from one PM's personal Claude Code setup into standard kit for 60 to 70 of Figma's product managers. Mechaber also demos AutoXP, an agent that runs growth experiments and stops for a single human approval — reportedly a big win on its first run. A concrete template for turning individual agent hacks into team practice: one shared repo, named skills, and a checkpoint where the judgment stays human. He also runs a bug-bashing agent on Codex that catches issues he says he would never spot himself. How I AI (Lenny's Newsletter)
“whose goal is to be "lazy in the right places."”
One laptop, parallel chats, a fleet of coding agents. A solo developer wired ordinary ChatGPT conversations to his own machines through MCP tools and OpenAI's Secure MCP Tunnel, then used GitHub issues as the work queue: a planning model split a big task into well-specified issues, and several parallel chats each took one, editing files, running tests and iterating. The spec-as-queue pattern working in the wild — and a portrait of the power user whose habits flat-rate plans cannot survive, which is exactly why unlimited chats are ending. The author says unlimited chats were his main reason to subscribe and he is switching to Claude. chaosmanage.github.io
Market and platform news
Garry Tan — The YC CEO told Walter Isaacson's Builders podcast he spends roughly $100,000 to $200,000 a year on agent tokens, running everything from coding to his personal life, and has published his 23 Claude Code skills as gstack on GitHub. A rare public number for what a fully agent-assisted workflow costs — and Tan himself names cost as the main blocker to mainstream personal agents. YC also runs an internal agent, Stray Light, holding context on every founder it has ever backed. TechCrunch
Squarespace — The website company is cutting about 300 jobs, roughly 19% of its workforce, as prompt-to-site tools reshape the category; Squarespace itself has an AI site builder, Blueprint, and an upcoming AI-native platform called Forma. AI-native competition is forcing structural cost cuts at incumbents, not just new features — the shape of a market your product may be sitting in. Rival Wix paid $80 million for Base44 last year, and Lovable raised at a $13.3 billion valuation in August. TechCrunch
Cresta — Salesforce announced Monday it plans to acquire Cresta, a nine-year-old maker of AI contact-center automation software; The Information reports the price at $2.7 billion. Contact-center automation keeps proving the most reliably sold enterprise agent use case, so expect your support-tooling roadmap to be measured against it. The price is unconfirmed by the companies; it comes via The Information, surfaced by Techmeme. The Information via Techmeme
MOFCOM Announcement No. 48 — China's commerce ministry proposes adding 40 technologies to its export-control catalogue, 23 of them AI, restricting frontier model weights and capable agents; a pre-export registration requirement is already in force, and the draft is open for comment until November 15. If your roadmap self-hosts a Chinese model outside China, that deployment counts as an export needing a license — API and SaaS use does not. Cloud Security Alliance
The big AI stories today
Anthropic — Anthropic's amended S-1 shows about $17 billion in Q3 revenue, a $23.3 billion loss mostly from note accounting, and $895 billion in compute commitments — the base under every Claude feature. SEC
Brave — The NYT and three more publishers sued Brave for allegedly selling scraped news to AI firms — the test case for secondhand content feeding agents. The Verge
GPT-Rosalind-6 — OpenAI published a system card for GPT-Rosalind-6, a trusted-access-only biology-and-medicine model sharing GPT-6 Astra's weights — gated vertical frontier models are a pattern now. OpenAI Deployment Safety Hub
One takeaway
Price the heavy session before you promise it. The Grok admission, the merged OpenAI allowance and Anthropic's own consumption-passing-payroll milestone all point one way: agentic usage is not chat usage, and flat-rate promises do not survive contact with it.
This quarter, measure what your flagship agent flow costs per heavy user, and design the metering — or the price — before marketing says unlimited. Stop assuming chat-style pricing carries over to agent features. TechCrunch