- Anthropic filed to list on Nasdaq as ANT, disclosing $16.3 billion of first-half revenue, up from $1.2 billion a year earlier, and a first quarter of operating profit.
- The same week, Broadcom lined up $63 billion of Wall Street financing to buy the chips it will sell to Anthropic. The capability is getting cheaper; the plumbing is getting financed.
- OpenAI shipped GPT-6.1 Terra at half the price of its mid-tier sibling, aimed squarely at models that click around a screen instead of just writing text.
- Alibaba's mapping unit open-sourced a whole-body humanoid model trained on people wearing cameras, with no robot in the data-collection loop at all.
- Letta published a memory benchmark that scores what an agent learned rather than what it can recall, and most harnesses did badly.
- If you want the shift underneath all of this, start with where generative AI turns into functional AI.
Two things happened in the same 48 hours, pointing opposite ways. Using AI got cheaper: OpenAI halved its price for screen-driving agents, a mapping company gave away a humanoid robot model, Apple shipped a Siri that remembers. Building AI got pricier: Broadcom arranged $63 billion to finance chips for Anthropic, and SpaceX leased $40 billion of GPUs rather than buy them.
That gap between falling prices and rising capital is the story. Somebody has to hold it, and the answer keeps being structured finance: the bill got moved, not paid.
The Front Page: Anthropic opens the books
Anthropic filed its Form S-1 with the SEC on October 26, seeking to list Class A shares on Nasdaq as ANT, with Goldman Sachs and Morgan Stanley fronting a 22-bank syndicate. The headline figures: $16.3 billion of revenue in the first half of 2026, up from $1.2 billion a year earlier, against a net loss of $80.7 billion, with $131.7 billion in cash and short-term investments. CNBC; Yahoo Finance on the risk disclosures.
Read the loss line carefully, because $72.4 billion of it is non-cash fair-value movement on convertible notes, derivative liabilities and SAFEs. In plain terms: when a valuation jumps, the financing instruments investors hold get marked up, and that markup lands on the income statement as a loss even though no money left the building. Strip it out and the picture inverts. The filing reports $974 million of operating profit in the second quarter, the first time the business has made money.
The number worth staring at is the mix. Consumption revenue, meaning usage billed by the token, was $13.1 billion, or 80% of the total; subscriptions were $3.2 billion. That is a business whose top line moves with how much compute its customers burn, in a market where the price per token keeps falling. Growing 13x while your unit price deflates means volume does the work, and volume is what a competitor can take.
What it means: The frontier is no longer a faith-based asset class. Every lab's economics now get measured against a real filing, in public. For anyone building on these APIs, the signal is that 80%: your vendor's revenue depends on you using more tokens, while its margin depends on charging you less per token. The usual resolution is cheaper models and chattier agents. Budget around tasks completed, not tokens consumed.
Releases & Features
OpenAI's GPT-6.1 Terra. A mid-tier model built around computer use, meaning it looks at a screen, then clicks, types and moves between apps the way a person would. OpenAI lists it at $1 per million input tokens and $5 per million output, half the rate of GPT-6.1 Sol. On Terminal-Bench 4.0 the company says it solves 50.3% at maximum reasoning effort against 21.5% for GPT-5.6 Terra. Its own numbers, its own run, so treat them as a claim. The announcement; outside coverage.
Amap's humanoid model, given away. Alibaba's mapping unit open-sourced a whole-body humanoid model trained entirely on egocentric human data, meaning footage and motion captured from cameras worn by people doing ordinary tasks, with no robot in the loop. The dataset runs to 500 hours across 827 task types. The team reports 62.5% success on real-world tasks, transferring to a Unitree G1 without retraining, and says one demonstration took about 29 seconds to collect against roughly 100 with a physical robot. The paper; the code.
Apple gave Siri a memory. iOS 27.2 shipped October 26 with Siri AI able to recall earlier conversations and pull context from Messages photos, calendar attendees, voicemails and Safari history, plus five new languages and an Excluded Apps setting that fences off what Siri can read. The changelog.
What it means: One pattern across all three: the expensive part of an AI feature is no longer the model, it is the context. Amap's trick is that human video is cheap and robot teleoperation is not. Apple's headline feature is memory, which is plumbing. The leverage is in what you feed a model and what it keeps, the unglamorous work most agent workflows run on.
In the Lab
Letta released FrontierMemory, a benchmark aimed at a real problem in how agent memory gets measured. Existing tests mostly check recall: can the agent find a fact it was told earlier? That is easy to game, because an agent with a search tool can look things up. FrontierMemory instead has a memory agent read 674 past agent runs, each stamped pass or fail, write notes from them, then hands those notes to a fresh agent facing 15 memory-dependent tasks. The question is not what it remembers, it is whether its notes made the next agent better. Letta reports its own Letta Dream on top at 73.3%, the OpenAI Agents SDK at 40%, and Claude Managed Agents and Hermes at 33.3% and 26.7%. The benchmark; The New Stack.
What it means: Hold the leaderboard at arm's length, because the vendor that built the benchmark came first on it. The method is the interesting part. Scoring memory by downstream task performance is the right idea, and a spread from 73% to 27% says the harness you pick matters more than most teams assume. Test it on your own tasks first.
The Oversight Desk
On the same day Anthropic filed to go public, the Justice Department noticed an appeal to the Ninth Circuit, trying to reinstate the federal ban on agencies using Anthropic's technology. A district judge lifted that ban in August, ruling the administration failed to justify it. Bloomberg.
What it means: A company can file for a reported trillion-plus valuation and still be litigating for permission to sell to its own government. That is not one vendor's problem: the mechanism that bars a supplier today can bar a different one next year, on grounds with little to do with the technology. If your product touches a public-sector customer, single-vendor dependence is political exposure now. Build the swap-out path before you need it.
Cheap models and expensive infrastructure stay someone else's problem until the bill lands on you. Tell BYOBot what you run and get a spec back for routing the easy work to the cheap model.
On the Radar
Smaller moves worth a glance, sources attached.
- SpaceX leases its GPUs instead of buying them. A binding commitment for roughly $40 billion, where an Apollo-managed entity buys the chips and servers and leases them back for SpaceX's terrestrial compute buildout. TechCrunch.
- Broadcom raises $63 billion to supply Anthropic. Bank of America, Blackstone, Citigroup and Morgan Stanley among the backers financing chips Broadcom will sell into Anthropic's buildout. Bloomingbit.
- Meta's legal bill reaches $42 billion. A $25 billion New Mexico judgment, under appeal, plus an $18 billion settlement, which Axios reads as a warning that courts move before AI regulation does. Axios.
- An open DNA language model. Carbon-X, a 14.8-billion-parameter open model built on Qwen3-14B, released October 25 to put genetic sequence and plain language in one vocabulary. The sort of release that never makes a keynote. Release tracker.
The Bottom Line
One sentence: inference is deflating, capex is inflating, and the gap gets bridged with leases and syndicates rather than cash. Anthropic's filing is the first time anyone has shown their work on that, and its most honest line is the 80% consumption split, because it says the frontier labs sell a commodity whose price they cannot defend. Terra at a dollar, a free humanoid model, Siri with a memory: that is deflation from your side of the counter. Go spend it on building something while everyone else watches the financing.
Frequently Asked Questions
-
No. Anthropic's S-1 attributes $72.4 billion of that first-half net loss to non-cash fair-value changes on convertible notes, derivative liabilities and SAFEs, which are accounting revaluations of financing instruments rather than cash going out the door. The filing also reports $974 million of operating profit in the second quarter, the company's first. The real pressure is not the loss line, it is the scale of compute commitments the company has to keep financing to defend that revenue. For the wider market picture, see our map of the AI automation tool landscape.
-
It is closer than it was. OpenAI lists GPT-6.1 Terra at $1 per million input tokens and $5 per million output tokens, half the rate of GPT-6.1 Sol, with cached input at $0.05 per million. Computer use burns tokens fast because every screenshot is input, so the cached-input price matters more than the headline rate for anything that loops over the same interface. Price it against a correct result rather than against a token, and design the task so the model sees the screen as rarely as it can get away with. Our workflow directory has patterns for that kind of scoping.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
