Welcome to the first issue of Signal. The idea is simple: every Thursday, what actually happened in AI, 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: the AI price war is on
What happened.
On 22 September OpenAI launched GPT-6 Sol and GPT-6 Luna, with API prices cut by about half against GPT-5.6 (OpenAI).
The same day, Anthropic released Claude Opus 5.5, which it says costs 40% less to run than Opus 5 (Anthropic).
A week later, at DevDay, OpenAI added GPT-6.1 Sol, which its changelog says offers "near-Astra performance for complex work at a lower cost than Astra" (changelog).
The two biggest labs cut prices in the same week.
My take. The people building on top of these models are winning, not the labs. That's just how competition works: more labs fighting, including cheaper Chinese open models pressing on the American companies, means more options and lower prices for everyone downstream.
What's interesting is that more options doesn't mean I actually switch. I mostly work in Claude and Claude Code. I've moved some workflows onto GPT, and GPT will sometimes ship a model that beats anything Claude has. But I've learned that Anthropic will ship something comparable, or better, soon after. So I stay put. Moving context and rebuilding workflows between tools is tedious enough that it's rarely worth it.
That's the part to watch. The price war is real and it's good for you. But the labs aren't really competing on price. They're competing to become the place your work lives, because once it lives there, you don't leave.
Signal or Noise
"One prompt, one hour, a finished motion reel." REAL, but oversold.
A prompt has been going around that gets Opus 5.5 to write the code for a motion-graphics showreel (Latent Space).
I used Opus 5.5 to make Signal's 30-second brand film, so here's what the posts leave out:
The viral examples are "make me something cool". A business video has to be specific: your brand, your style, and facts you can stand behind.
Mine took two to three days of rounds. That covered picking the music, lining up every sound effect with every cut, and checking every claim on screen.
By hand, it would have taken far longer, so the tool is genuinely good. But "prompt it, walk away, come back to a finished video" isn't what happens when the video has to be right.
Google's Gemini will phone businesses for you. TOO EARLY, but pay attention.
Google is testing "Call for Me": Gemini calls a business to check stock, book, or reschedule, and you watch a live transcript (TechCrunch).
For now it's US Pixel 11 owners on paid plans only.
Still, the next caller to a salon or a dentist might not be a person. If your hours and booking details are wrong online, that call goes somewhere else.
ElevenLabs Eleven v4 and v4 Turbo. REAL.
New voice models in 90+ languages. Turbo is built for low-latency voice agents (ElevenLabs).
AI voices on the phone just got harder to tell from people.
Opus 5.5 across "12 laptops" making an editable After Effects video. HYPE.
It's a slick demo from Higgsfield (product page). The setup is gated behind commenting on the post, and nobody has reproduced it independently.
The tell: a launch video with no way for you to try it.
A security hole in the official MCP Python SDK. REAL, act on it.
If you plug "connectors" into AI tools: versions 1.9.1 to 2.1.1 of the official SDK let a malicious server steal OAuth secrets. It's fixed in 1.30.0 and 2.2.0 (advisory).
Check your versions.
Tool Worth Trying: Claude Code with Opus 5.5
This is the tool I actually use every day, and the one I made the film with.
What it's good for: long, multi-step builds where the AI has to hold a lot of context: a website, a video, a codebase.
What it isn't: a vending machine. You still direct it, check its work, and go several rounds. That's also why I stick with it: the more you build in one place, the better it knows your work.
Power & Policy
Going in-house.
Fact: Latham & Watkins, one of the largest US law firms, bought Nvidia servers and is fine-tuning Nvidia's open-weight models in-house. It cited data sensitivity and not wanting to be locked into one vendor (FT, via Latham).
My take: I think this is where businesses with sensitive data end up. The big labs constantly change subscription plans, usage limits and API limits, and model quality isn't consistent from one week to the next. Some days a model is great. Other days the same model just doesn't work the way it should, and you never know whether that's the lab or the model. If you can run a lightweight, open-weight model built for your own use case, on compute you control, it can be exactly as smart as you need. And nobody can change the terms on you.
"Superintelligence," by executive order.
Fact: On 29 September the President signed orders renaming "Artificial Intelligence" to "Super Intelligence" across the executive branch. He also launched America.gov, an AI chatbot for government services (White House).
My take: it's a naming convention. It doesn't change what the technology is or what it can do. And a government chatbot is hardly news: every website has one now.
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. Free 20-minute AI audit if you're curious.