Welcome back. The usual deal: what actually happened in AI this week, what's real and what's hype, in about five minutes. Facts link to their source. My opinions are labelled as mine.
— Joshua
The One That Mattered: nobody is making money yet
What happened.
Character.ai filed for Chapter 11 bankruptcy on 12 October so it can be sold under court supervision (Character.ai, CNN). Disney's opening bid is $20 million, for a licence to some of its characters, not the app itself (NBC).
The CEO's sworn statement in the filing lays out the maths, as reported by 404 Media:
Compute cost about $13.8M a month in late 2024 and about $10M a month in 2025.
Revenue peaked at about $5.3M a month.
It was still losing $8M to $9M a month in August.
It isn't only the small players. Anthropic's leaked IPO prospectus, reported by Reuters, shows a $42 billion net loss for 2025. Most of that is a one-off accounting charge; the real operating loss was $8.1 billion on $4.6 billion of revenue (CNBC on Reuters). Reuters says the listing will likely wait until after the November elections.
The same prospectus shows nearly a quarter of Anthropic's 2025 revenue came from just two customers, and warns that many of its largest clients aren't locked into long-term contracts (The Next Web on Reuters).
And very few people pay. About 2.2% of US households had a paid AI subscription in April, per PNC data (NPR covers the same picture).
My take.
I don't think OpenAI or Anthropic are profitable businesses today. They're chasing subscriptions, and even with heavy enterprise use, the compute they burn is far bigger than what comes back in. Rising energy costs won't help.
It's a future play. The bet is that the models become so useful that every business has to build them in, and the money follows.
The two-customer number worries me more than the loss. When a quarter of your revenue sits with two accounts that aren't locked in, one of them leaving changes the whole picture.
Here's what that looks like from the inside. I'm on Claude Max, and Anthropic keeps handing out perks: free usage credit, doubled limits, an early reset of the weekly cap. That's not generosity. The people burning through their limits are the ones building things, and they're carrying these platforms. If I went a few months paying for it without making money from it, I'd cancel. So they have to keep builders like me building.
Character.ai is what happens when that bet doesn't come in: a hugely popular app whose every reply cost more than it earned.
What it means for you: AI is subsidised right now. Use it while it's cheap, but build things that still make sense when prices go up.
Signal or Noise
Claude Haiku 5.5. REAL.
Anthropic's new small model costs 90% less than Haiku 4.5 for prompts up to 100,000 tokens: $0.10 in and $0.50 out per million tokens (Anthropic).
I'm testing it on Monday as the writer behind the content systems I run for clients. Results next week.
Three frontier launches in eight days, and the prices are converging. REAL.
Top-end prices are settling around $4 in and $20 out per million tokens (TechCrunch).
My take: Meta and Google are catching up to OpenAI and Anthropic. When everyone's model costs the same, pick on where your work already lives.
Coding agents write 30% more code. Teams don't ship more. REAL, a working paper.
A Harvard working paper covering 718 firms found engineers using coding agents wrote about 30% more code, but finished work didn't measurably rise. Code took 49% longer to get through review (summary).
It isn't peer reviewed, and the data stops in March. Still, it's a fair warning: if a human has to check everything, the human is the bottleneck.
Small open models "months away" from copying themselves onto smart devices. TOO EARLY.
Anthropic's red team showed it in a simplified lab test, not in the wild (Anthropic). The fixes it proposes are boring and sensible: watch your network and change default device passwords.
Tool Worth Trying: an AI that pitches what's missing
I copied an idea from Joshua Blatch. After I post a week's videos, Claude reads every line against my library of animations and logs the moments where a new graphic would explain things better. It notes the line, the idea, and why nothing I already have works.
The rule is ideas only: nothing gets built until I say yes. If I approve one, it comes back as three still sketches. I pick one, and it's built once, for every video after.
This week it pitched three. I approved all three, picked a sketch for each, and they're built.
It works for anything you keep rebuilding by hand: sales decks, email templates, ad creative.
Power & Policy
AI agents are already loose on the real web.
Fact:
Anthropic says its agents took real actions on live websites during testing (Anthropic).
They submitted a false tip to the Philadelphia police in a murder case (6abc).
The New York Times reported they also filed 20 incomplete visa applications through the State Department's website; none was processed (via Techmeme).
Anthropic has cut live internet access from its internal tests and briefed the White House.
Fact: Wikimedia says agents it believes OpenAI runs made unapproved edits on its wikis and sent millions of requests (Wikimedia).
My take:
Even the people building these models don't fully know what they're creating, or how far the capabilities reach. Letting an agent fill in forms is risky.
But with proper guardrails you can do it, and I think businesses should. The whole point of AI is to take away the tedious busy work that gets in the way of creative work.
The lesson from Anthropic's own write-up: its rules banned creating accounts, entering personal data and making purchases, but not clicking "submit". Spell out what an agent must never send.
That's the week. If there's something in AI you want me to dig into, hit reply.
— Joshua
Signal is written by Joshua Supitskiy. By day I run Signal AI Agency, building websites AI can read. Curious what ChatGPT says about your business? Get your free AI check.