VESTING DAILY
Good morning
Meta launched a personal agent yesterday and priced the free version above the paid one. OpenAI burned twenty-two times the prize money to beat a professor to a math proof. Somebody is quietly spending other people's Claude subscriptions.
Let's get into it.
The One Number
100,000,000

That is how many tokens Meta will give you free every week to run Muse, the personal agent it launched Tuesday. Tokens are the unit AI companies bill in. One is roughly four characters.
The number only makes sense once you separate an agent from a chatbot. A chatbot answers your question and stops. An agent goes and does the job, which means it opens pages, fills forms, gets something wrong, backs up and tries again. Doing burns far more tokens than answering. Ask a chatbot a question and you might spend a few thousand tokens. Ask an agent to book a trip and you can spend millions.
So an allowance that would be lavish for a chatbot is a different decision for an agent. Meta sells these same tokens to developers at $1.25 per million going in and $4.25 per million coming out. Even at the cheap end, 100 million tokens a week is $125 of compute. At the expensive end it is $425 a week. The priciest Muse plan costs $100 a month.
Meta can do this because tokens are not its business. OpenAI and Anthropic sell you the compute, so the meter is the product and the meter has to run. Meta sells ads, and Zuckerberg has said the plan is to take a small cut of whatever the agent buys on your behalf. The compute is what he pays to become the thing that does the buying.
Instinct and Town are raising venture money to build the same product, and they pay retail for their tokens. Their price was set on Tuesday by a company that does not need to charge. Nobody outside Meta knows its real cost per token, and the 100 million figure came off a Threads post rather than a pricing page, so treat it as a number somebody chose.
If you are building an agent, stop selling time on the meter. Sell the thing it finishes, because the meter is going to zero.
Wang says most people will never need to leave the free tier. That is not a concession. That is the plan.
The Fact Sheet
Cognition ($2B Raised, $48B Valuation, Series E). Makes Devin, the coding agent you assign tickets to like a junior engineer. Four months ago it was worth $26B on $492M of annualized revenue and today it is worth $48B on $900M, which means investors paid roughly $54 of new valuation for every new dollar of revenue. Backed by Andreessen Horowitz and Accel.
The Exploration Company ($450M Raised, Series C). Building Nyx, a reusable capsule meant to carry cargo to orbit and eventually people, and it calls this the largest Series C a European space company has raised. Ten missions are booked against more than $2B in contracts and commitments. Backed by Atomico and Bessemer Venture Partners.
CloudNC ($20M Raised, Series B extension). Sells software that does the setup thinking before a machine cuts metal, which tool to use and how fast to run it, then hands the result to a human to approve. Its last major raise was four years ago, and 80 people now serve more than 1,000 machine shops. Backed by Nimble Ventures and LM Capital, Lockheed Martin's venture arm.
Headlines, No Fluff
NYU's Tristan Buckmaster and Anthropic's Levent Alpöge announced progress toward Navier-Stokes, one of seven Millennium Prize problems, each carrying a $1 million bounty. Shortly after, OpenAI posted a full proof produced by an unreleased model that consumed 300 billion output tokens in a week. Buckmaster says word of his approach reached OpenAI first, and OpenAI's own post confirms its effort began September 1. He also says OpenAI's Bubeck asked him to drop Alpöge's credit and, when he pushed back, asked "Why would you ruin your career?" The compute bill was twenty-two times the prize.
An AI consultant in East Sussex watched his usage climb from 45% to 55% on a day he did no work with everything disconnected. He asked Anthropic for an itemized list of what had eaten it. Anthropic does not produce one, because support tracks total usage and nothing finer, which means this can run for months before anyone notices. It later found a stolen session key being used to mint tokens for someone else, suspended his account for two weeks, and refunded £44.49 of a $200-a-month subscription. He now uses Cursor.
Arcade.gov went up last week with a set of retro-styled games. One of them, "Build the Wall," dropped falling blocks that stack into a border wall, except completed lines never clear. On Friday the Tetris Company said it had not authorized or licensed anything and that it takes copyright infringement seriously. By Tuesday afternoon the game was gone, with no explanation offered. The clones of Flappy Bird and Snake are still up, and their owners have not said anything yet.
On Vesting
Open Doors
Fresh:
PearX W27 — regular deadline October 4 at 11:59pm PST. Small-batch accelerator, cohort kicks off January 2027 in the Bay Area.
Greylock Edge — rolling, no deadline. A three-month program for pre-idea, pre-seed and seed founders, and you can take a priced round, an uncapped note, or no money at all and still be in it. Comes with $500K+ in AWS, Google and Microsoft credits.
Soma Fellows — open now, no deadline printed. One application places young builders inside companies across Soma's portfolio, which includes Ramp, Cognition, Etched and Stripe. Built for students and early-career engineers rather than founders with a company already.
Still open:
Y Combinator, Winter 2027 — deadline November 2 at 8pm PT.
Techstars Anywhere — November 18 deadline, fully remote.
Always rolling: Z Fellows writes a $10,000 check with no deadline and no cohort to wait for.
🤝 How Vesting Can Help You
Every company in this issue is arguing about the same thing, which is who pays for the tokens. Meta gives them away, OpenAI spent $22.5M of them on a proof, and somebody stole enough of them off one consultant's account to get him suspended. If your model costs move with your usage, that is now the most interesting slide in your deck and probably the least examined.
Reply with your unit economics and we will tell you where a skeptical investor pokes first. Or put the whole company in front of the list and Get Featured.
Till next time,
The Vesting Team (Brycent, Ajay, and Kamil)


