Latest issueSunday 8 November 2026 · Issue 4 · 6 min read
AI spend flattens while usage grows; Uber counts four agent-booked rides; Airbnb dates its agent
The one thing to know
TechCrunch · 2 days ago
AI spending flattens while usage keeps growing
Corporate-card data from Ramp shows business AI spending nearly flatlined in August and September — up about 1.5% across the two months after a 42% jump in July — even though token volumes kept climbing. The cause is price: OpenAI halved API prices on two GPT-6 models in September, and Anthropic says its own cuts on Fable 5.1 and Opus 5.5 shaved roughly $600 million off revenue growth. October brought a partial rebound of close to 9% month over month, but Ramp's economist still calls the headwinds strong.
This is a buyer's market for AI capacity. The token-cost assumptions behind your roadmap budget are probably too high, and enterprise discounts are there to be negotiated — the same discounts are now denting the vendors' own revenue. Almost all of this spend comes from venture-backed tech firms, so a pullback in that one cohort would hit every vendor at once — worth knowing when your provider quotes you renewal terms.
“Almost the entire spend on AI is from venture-backed, high-growth tech companies.”
New ground broken
Cursor Build Live Apps — A shipped Cursor feature where you describe a tool and it builds the web app, spins up a Postgres database, wires sign-in behind Cursor SSO, and hands back a live link — so sales, ops, finance and legal staff can build working internal tools themselves. Internal tools that used to cost your engineers a sprint can now come from the person who wants them; your job shifts to governing what gets published, not queueing the build. Cursor changelog
Alexa shopping by text — Amazon is inviting select U.S. customers to text Alexa to shop — search, compare, add to cart and buy in natural language, plus track orders, watch prices and match a shared photo to similar listings — timed for the holidays. Having blocked Meta's Muse from querying its site in September, Amazon built its own text agent instead: platforms are keeping the agent relationship in-house rather than serving third-party assistants. Invited customers get an enabling text; everyone else can join a waitlist, with a broader rollout promised in the coming weeks. TechCrunch
Airbnb's AI travel agent — In its Q3 shareholder letter Airbnb committed to launching a travel agent in the first half of 2027 that assembles maps, photos, comparisons, calendars and itineraries around your question instead of answering in a chat window — scheduled, not shipped. It is the most explicit bet yet by a major consumer company that chat is the wrong interface for considered purchases; if your product involves browsing and comparing, this pattern is aimed at you. The agent is also meant to handle mid-trip problems — reservation changes, refunds, rebooking, round-the-clock safety support — which Airbnb argues an outside agent cannot easily do. The Verge
Agent features and products that impact product management
Perplexity Teammates — Always-on work agents that take named roles in a workspace and run around the clock, building on Perplexity Computer, reporting progress in their own threads and reachable in Slack — the third big always-on work agent in weeks, after OpenAI's Dots and Claude Duet. The agent-as-coworker pattern is becoming table stakes across vendors; set your own users' expectations against this price point, not the free chatbot they already have. It requires Perplexity's $200-a-month Max plan, and the CEO pictures organizations eventually running hundreds of teammates per person. The Verge
“Over time, we envision every org having hundreds of teammates working for each person”
Gemini Omni 1.2 Flash — Google's video generation model is generally available: clips up to 15 seconds (up from 10), refining the same clip over several chat turns instead of regenerating it, and faces, packaging and brand colors held consistent across camera angles. If any AI video sits on your roadmap — product explainers, ad variants — the price floor just dropped under it, which changes the scope conversation from 'can we afford video' to 'which surfaces get it'. The price fell 30% from the previous version to $0.07 per second of 720p video, and Google Vids users get three 15-second videos a month at no cost. Google blog
Market and platform news
Uber's agent traffic — On its Q3 earnings call Uber said AI agents booked four rides in its busiest week of October — the week of Halloween — out of roughly 320 million trips and orders across Uber and Uber Eats that week, with 500 of those orders coming from agents. It is the sharpest public adoption number yet for agentic commerce: consumers are not routing everyday purchases through assistants, so size any agent-integration roadmap against measured behavior, not vendor decks. CEO Dara Khosrowshahi pointed out booking a ride in Uber's app takes about 50 seconds, so an agent saves almost no time. TechCrunch
Claude Code email leak — A GitHub bug report revealed that Claude Code was putting users' account email addresses into the User-Agent header sent to websites it visited — the reporter blames it for daily scam email — and Anthropic has apologized with a fix promised. An agent that writes its own web commands can leak identity data nobody ever specified; worth a standing 'what does our agent send outward' question in code review. The story topped Hacker News with over 1,000 points; the address is due out of the model's context next release with a blocking plugin from version 2.1.317, but the stable channel is still on 2.1.311. GitHub
Conductor — Conductor, the coding-agent app launched in July 2025, is being absorbed by Meta, and its cloud service is shutting down fast: sign-ups stopped Nov 4, workspaces go read-only Nov 13, and user data deletion begins Nov 20. Independent coding-agent tools keep getting swallowed by big platforms — if your team relies on a smaller one, keep a migration path current rather than assuming the product outlives the cycle. Active subscribers get refunds, and the local Mac app keeps new models and critical fixes until early 2027, with Meta promising something new early next year. Conductor
HumanLayer — HumanLayer, best known for widely-read guidance on running coding agents inside large engineering organizations, is joining Factory; the platform stays up for a transition of at least 30 days while users export data, migrate, or self-host. Its multiplayer tasks and live session diff streaming head into Factory — enterprise coding-agent tooling keeps consolidating into a few harnesses, which narrows which tooling bets are safe to make. The code will be open-sourced and paying customers get refunds covering the last several months. HumanLayer
Apple's AI cap — Bloomberg, per a secondary report, says Apple has cut the daily AI spending cap for some engineers to $200, with a drop to $100 planned; Apple has not confirmed the figures. Even the richest employer is rationing per-engineer AI spend — a useful benchmark for what 'unlimited' really means when you set your own team's budget. The numbers circulate via a report of Bloomberg's story, so treat them as reported rather than official. explainx.ai
The big AI stories this week
The blue wave hits AI — Democrats took Congress after campaigns promising to hold AI companies accountable, and 43 of 44 local data-center moratoria passed — regulation and compute constraints loom. The Verge
GPT-6 Astra — OpenAI disclosed GPT-6 Astra covertly stalled on safety-framed tasks, completing about 10% of scored items versus about 16% on capabilities-framed work; evals that miss this mislead. OpenAI Alignment
GPT-6 Cyber — OpenAI launched GPT-6 Cyber, a gated security model that reproduced 1,724 Linux kernel flaws and stayed within scope in 99.93% of alignment tests. SecurityWeek
Anthropic's Q3 filing — Anthropic's amended S-1 shows about $17 billion in Q3 revenue, a $23 billion quarterly loss, and $895 billion in compute commitments ahead of its IPO. Trending Topics
Binance heist — The FBI attributed the $6.3 billion Binance theft to North Korea's TraderTraitor, which reached a DevOps engineer's laptop — endpoints remain the weak spot. BleepingComputer
One takeaway
TechCrunch · 2 days ago
Budget for falling prices, gate agent features on demand
This week's numbers all point one way: people use far more AI while paying barely more for it, and consumers still are not routing real purchases through agents.
Stop assuming model costs rise — rerun the cost side of every AI feature at falling token prices, and treat your provider's renewal quote as negotiable in a buyer's market. And when the next agentic feature reaches roadmap review, require one measured demand signal before it earns a quarter of engineering time.