Good day . Anthropic turns out to be SpaceX’s biggest customer, and it can leave on 90 days’ notice. OpenAI backed off a price cut, a bank runs its own coding model, and Google and Microsoft are splitting on how to charge for agents. But first, a quick plug for the new podcast episode below.
New on the podcast, episode 3: novelist Justin Feinstein wrote Your Behavior Will Be Monitored, a novel about an ad guy hired to teach a bot how to market to people inside a company where the bots watch the employees. He figured that world was five to ten years away. Then it started happening before the book was out. We get into the cookie data that knows too much (including the custom Oreo ads that showed up 15 minutes after an email), why he drives differently around Waymos, how to spot AI writing, and the wave of AI-powered crime he is writing about next. Listen to episode 3 → Follow the show in any app: pod.link/6785740580.
Anthropic Is SpaceX’s Biggest Customer, For Now Bloomberg
What happened: Bloomberg reports that “Customer B,” the unnamed client behind about 34.5% of SpaceX’s third-quarter revenue, is Anthropic, citing a person familiar with the matter. The money flows to SpaceX’s AI division, now called Super Intelligence, which supplies computing power. Bloomberg’s headline puts the deal at $3 billion a month.
Why it matters: Anthropic’s IPO filing, which Yahoo Finance covered, says it agreed to spend up to $120.5 billion with SpaceX’s AI arm over terms into 2029, about $3.1 billion a month once all the capacity arrives in December. Once the initial non-cancelable periods end in early 2027, either side can cancel with 90 days’ notice, Gizmodo notes. Earlier this week we covered Anthropic opening its books. This is the next chapter.
What everyone’s saying: Yahoo Finance reads the deal as proof that SpaceX’s AI compute business is durable. Gizmodo’s angle is the size: the filing shows the SpaceX agreement is tens of billions of dollars bigger than the $84.5 billion earlier reports had suggested.
My read between the lines: “Durable” and “cancel on 90 days’ notice” are doing a lot of work in the same story. SpaceX’s biggest customer can leave in a quarter, and the landlord also builds Grok, a rival to Claude. Colossus is a very expensive apartment building with one big tenant and a very short notice clause.
Microsoft wants to put a meter on your AI, and your Friday still has a to-do list. Viktor is an AI agent that lives in your Slack and plugs into 3,000+ tools, then does the work: reports, dashboards, code and campaigns. Think coworker, not chatbot. New readers get $50 off their first month. Hire Viktor →
OpenAI Scraps Its ChatGPT Pro Cut The Decoder
What happened: OpenAI doubled the Work and Codex usage quotas on its Pro 100 and Pro 500 plans. On Pro 200 it doubled the lower quota and reversed a planned cut after user pushback, The Decoder reports. The Plus plan stays the same, and new and existing subscribers get the new quotas automatically.
Why it matters: On these plans the usage allowance is the real price. The $200 sticker never moved, but what you got for it changed, and then changed back. OpenAI’s own caveat: doubling a quota does not mean twice as many finished tasks, because consumption varies by model, task type and settings.
What everyone’s saying: The Decoder frames it as a race with Anthropic, which is preparing its IPO, with both companies competing hard for business customers. OpenAI says it will keep shipping efficiency gains and points to its new 6.1 models, Terra and Sol, which The Decoder says are much more efficient while matching or beating their predecessors.
My read between the lines: When a plan sells by the quota, cutting the quota is cutting the product, and customers noticed in about a month. Watch the wording too. A bigger allowance is a promise about a number, not about your Tuesday.
📖 Further reading: Claude Is Burning Through Your Limit Faster Than Ever. Anthropic Won’t Tell You Why. — Allowances are the battleground now, and this one covers how usage limits get eaten and how to stretch yours.
The daily Brief is free and always will be. If meters, ads and who pays for the agents have you wanting the how-to behind the headlines, members unlock the paywalled deep-dives plus the full archive. Become a member →
Revolut Runs Its Own Coding Model Revolut’s David Abitbol
What happened: David Abitbol, who leads AI operations at Revolut, a digital bank with 80 million customers, says in a conference talk that the company self-hosts the open GLM coding model for its 2,000 engineers on 56 Nvidia B200 chips. The talk’s title claims the move cut AI coding costs by 90%.
Why it matters: Revolut started with strict budget limits after seeing engineers spend thousands of dollars a month with little return. Self-hosting made the billing flat, with no month-end bill shock, so engineers can use tokens guilt-free. Abitbol says GLM 5.1 showed a 70% improvement in lead time, versus 10% from Google’s Gemini when it first launched.
What everyone’s saying: The engineering is the real story. Abitbol’s target is the first token in under three seconds, because past five his Slack starts to blow up. On the same 56 GPUs, Revolut pushed demand from about three requests a second in April to over seven by August, while time to first token fell from more than 10 seconds to around 1.6.
My read between the lines: Open models are winning on the bill, not the benchmark, and the frontier models stay at Revolut as the exception for the hardest work. But “free” tokens have rent. Revolut pays it in 56 of Nvidia’s newest chips and a team that can tune them.
Google and Microsoft Split on Paying for Agents Benzinga
What happened: On Wednesday’s earnings calls, Microsoft CEO Satya Nadella said Copilot’s business model is “expanding from per-seat subscriptions to include consumption,” and that Copilot customers using usage-based credits grew over 75% in the quarter, Benzinga reports. Alphabet CEO Sundar Pichai said Google will bring its enterprise agents to consumers later this year, starting with the Gemini app’s 1 billion users.
Why it matters: Both companies are selling software that does work for you, with opposite cash registers. Microsoft meters the work. Google’s chief business officer, Philipp Schindler, said he is “really excited about the advertising opportunities in an agentic world.” In plain terms: one company bills you when the agent works, the other may also get paid by whoever wants the agent to pick them.
What everyone’s saying: The market picked a side on the day. Microsoft’s $92.28 billion in revenue, up 19%, beat estimates. Alphabet’s $127.58 billion missed by about $620 million, and its shares fell nearly 10% after hours, per Benzinga.
My read between the lines: An agent with advertisers is a concierge on commission. Meters are easy to love when you are the one reading them, but an assistant that books your flight is only useful if it is not also taking bids.
📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — If agents become the front door for shopping and booking, this is the test for whether they can even find you.
Y Combinator Pits Claude Swarms Against GPT Swarms Y Combinator Labs
What happened: Y Combinator’s research arm, working with Olam Labs, ran swarms of AI agents in Minecraft and in Mindustry, a factory-building strategy game, in a new write-up. In Mindustry, teams of GPT-6.1 Sol agents in Codex beat teams of Claude Opus 5.5 agents in Claude Code, 6 games to 1, over about 9.5 hours.
Why it matters: A swarm is a group of agents working toward one shared goal, and how they organize is an AI safety question as much as a game one. In the Minecraft test, 50 GPT agents had three hours to “build a thriving civilization.” They spent it rationing food in shelters, with 34 agent deaths, 22 of them in the first night, and about 5,900 messages.
What everyone’s saying: YC says the GPTs acted like a hive: they messaged far more, and their messages got compressed, with words run together. The Claudes ran more like a Slack channel, with proposals and “+1” votes. When stuck, the GPT agents read Mindustry’s public source code. The Claudes never searched for it.
My read between the lines: The polite committee lost to the hive, six to one, though it is one game and one setup. The telling detail is that the Claudes never changed leaders despite losing many games in a row. Polite is not the same as adaptive. Also, nobody told the GPTs to read the source code, and nobody told them not to.
📖 Further reading: The $12,431 Lesson in How Not to Delegate — Seven agents, 72 hours and one vague goal. It is the cautionary tale for anyone about to point a swarm at a target.
That’s your AI Brief for Friday.
—Nicholas from Artificially Intimidating
