Your coding agent can write the code. It can't answer the customer — no email, no calendar, no seat in chat. Puffin gives each agent its own address and inbox, with every message held for approval. Identity is the new seat license. Sell the inbox, not the bot.
Chargebee's Arc hands the money flow to agents: contract in, invoice out, collections chased, revenue recognized — every step cited and held for human review. The best-paid agent job was never writing. It's collecting. Sell the outcome, not the output.
Google shipped Gemini 4 Neon at $3/$15 per million tokens — near Opus 5.5's coding scores for 25% less. A day earlier Microsoft put MAI Code 2 in Copilot at $0.50/$2.50. Three labs repriced in 72 hours. When tokens race to the bottom, sell the routing, not the model.
Model ML raised $250M at $1.5B to give every banker a personal agent — models, decks, due diligence, inside the firm's own systems. The company says revenue grew 12x in a year. Wall Street's first AI casualty isn't a job title. It's the all-nighter.
Genesis AI raised $500M at $3.5B for Eno, a general-purpose robot with human-like hands — no head, no legs, just wheels and dexterity. Only 30–40 built, pilots unnamed, production in 2027. The bet isn't the robot. It's the data loop: human hands in gloves teaching the model.
Anthropic priced at $83, above its $70–78 range, and upsized to 980M shares — an $81B offering, the largest IPO ever. ANT trades today. A model lab is now a public company. The compute war has a ticker now, and every quarter, the bill is public.
Your coding agent can read the whole repo. It can't see production. Appxiom feeds it the crashes, freezes and failed checkouts — and ranks bugs by the signups they cost you. Agents don't fail for lack of code. They fail for lack of context. The fix layer is the product now.
A test hid Google's AI Overviews from 1,065 users: clicks to outside sites fell 39.8%, zero-click searches rose 54% to 73%. Ad clicks didn't move. If you live on search traffic, the funnel didn't shrink — it stayed with Google. Rebuild to be cited, not clicked.
OpenAI cut GPT-6.2 Astra 20% to $8/$40 per million tokens, and says a task now costs under a third of before. Independent tests disagree: Artificial Analysis says 36% cheaper, Vals AI says about 50%. The sticker price isn't the bill. The business now is measuring real task cost.
Isomorphic Labs just raised $5B — six months after a $2.1B round — to push AI-designed drugs into trials. Discovery was the demo. Trials are the business. The lesson for AI founders: the model isn't the product anymore. The regulated outcome is.
Anthropic prices its IPO tonight: $70–78 a share, up to $64B public, orders 7x oversubscribed. The twist: its biggest suppliers are buying $26B of the stock too. Same price as you — plus a $250B compute order on the side. The AI boom now funds itself. In a circle.
Nvidia just did $115.8B in a quarter — up 103% — and guided $130B next, its first report since Jensen Huang resigned. Supply commitments jumped to $390B. The bubble debate keeps losing to the order book. Sell the memory and the power, not the model.
Exa cut its Monitors price from $15 to $2 per 1,000 runs — scheduled web search that pings your agent only when something new turns up. At $2, always-on watching is a rounding error. The agent businesses that win will be the ones that never stop looking.
Launch Club makes $34,447 a month from 19 clients, per TrustMRR payment data: getting brands mentioned on Reddit so AI search tools cite them. SEO was about ranking. Now it's about being the answer. Sell the citation, not the click.
GitHub's fastest-rising repos this week aren't agents. They're the leash — tools to run agents from your phone, review their code line by line, and make them prove it in a real browser. Writing got cheap. Supervision is where the money's going.
UFORCE raised $700M at $6B — a Series A, eight months after a $50M seed at ~$1B. Co-leads: Valor Equity and the UK's National Wealth Fund. When sovereign money co-leads your round, your customer is also your investor. In defense AI, the moat is procurement, not product.
Marc Lou's $100K hackathon started yesterday: 30 days to build a business, winner decided by profit — not revenue, not votes. 3,492 builders in. Most hackathons reward the demo. This one rewards the P&L. In AI startups, that's the only scoreboard that matters.
Corma hit a $1B valuation three months after launching — $125M Series A, NEA leading, Sequoia and Khosla already in. Cybersecurity is the one AI market where buyers don't ask for a demo. They ask: can your agent defend faster than theirs attacks? Speed is the pitch now.
DeepMind's agents formally verified Signal's crypto core in four weeks — 6M+ lines of Lean proof, a job DeepMind says took expert teams years. The proof only covers what's specified. Still: certainty just got cheap. Formal verification is now a service, not a luxury.
Browser Use flipped its default model to BU 4.0: $1.80/$9.00 per million tokens, 15x the old bu-fast. My code didn't change — my bill would have. Defaults are the quietest price hike in software. Sell the watchdog that pins them before the invoice lands.