😁 Hello, super humans! Friday’s big question is what happens when a chatbot stops answering in paragraphs and starts handing you a working calculator instead. OpenAI just shipped that to a very large audience, and the details of how it works matter more than the demo. We also have a price war in small models, an open-source toolchain for robot builders, and a rare piece of good news for telecom spending.
📰 Quick Signals
- 🧠 AI: Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens, about 75% cheaper on average than Haiku 4.5, with an adjustable effort setting for the first time in the Haiku class.
- 🤖 Robotics: AWS launched an open-source Physical AI Toolchain that stitches SageMaker and IoT Greengrass together with NVIDIA’s Isaac Sim, Isaac Lab, GR00T, and Cosmos, from synthetic data to edge deployment.
- 💻 Programming: Cloudflare opened K2 in public beta, a serverless event stream built on R2 that claims about one second of p99 produce latency at $0.04 per GB produced.
- ⚡ Electronics: Researchers built a battery-free door sensor from a triboelectric generator using ground garlic peel and polyethylene, producing roughly 210 V AC each time the door opens.
- 📡 Telecom: Dell’Oro reports worldwide telecom capex rose about 5% year over year in 1H26, a stronger-than-expected rebound after several years of decline.
The Big Story: ChatGPT now builds calculators mid-answer, and 1.2 billion people just got it
When the interface of the most used AI product on Earth starts generating small apps on the fly, a lot of “build a quick tool” work quietly moves into the chat box.
What happened: OpenAI began rolling out GPT-6 with a feature it calls Intelligent UI, as described in its GPT-6 announcement. Starting October 7, Plus, Pro, Business, and Enterprise users get GPT-6 Sol in the Chat tab, and Free and Go users get the lighter GPT-6 Luna the following day. OpenAI says the product serves more than 1.2 billion people each week. The models behind Work and Codex are unchanged.
The details: Instead of text only, a reply can now include charts, forms, tappable buttons, and small tools such as a calculator, bill splitter, or game, and you can change inputs inside the answer. OpenAI also says ChatGPT can start answering while it is still thinking: on questions that need web search, GPT-6 Instant begins answering 44% sooner on average than GPT-5.6 Instant. One architectural point is worth noting. Gemini and Claude have offered interactive answers for a while by having the model write code that renders in the page. OpenAI describes assembling prebuilt parts such as sliders, charts, and forms, which should give more consistent layouts at the cost of flexibility. The announcement makes no accuracy claim for the numbers these tools produce, and the speed and quality results are internal evaluations.
flowchart LR
Q["User question"] --> M["GPT-6 Sol / Luna"]
M --> D{"Text enough?"}
D -- yes --> T["Plain text reply"]
D -- no --> U["Intelligent UI"]
U --> P["Prebuilt parts:\nsliders, charts, forms"]
P --> I["Interactive answer\nuser edits inputs in place"]Important
Our take: The novelty is not interactive answers, it is putting them in front of free users at this scale. For builders, that raises the bar on small internal tools and calculators, since a decent first draft now costs one prompt. My worry is trust: a polished slider makes a wrong formula look authoritative, so check any number that drives a decision. If you ship anything similar, show the formula next to the result.
🗞️ More News
🧠 AI
- Nous Research confirmed a $1.5B valuation after a $90M Series B with NVIDIA and Samsung participating, and launched Hermes for Businesses for private, customized agents.
- Biohub, the DOE, and NIH committed $1.8B to open biology data for AI models that predict how cells respond to treatments, with another $300M from Google DeepMind, Isomorphic Labs, and Meta.
- Common Sense Media rated ChatGPT for Teens an unacceptable risk, saying parent alerts never fired during an hour of self-harm testing, and urged OpenAI to limit it to adults.
- Crunchbase says North American startups raised $92B in Q3, down 35% from Q2 but up 50% year over year, with about two-thirds going to AI.
- Stuut raised a $52.5M Series B for AI order-to-cash automation and claims customers collected up to 40% more cash.
- Microsoft opened Surface Laptop Ultra preorders at $2,599, shipping October 16, with a USB-C port designed to pop off if someone trips on the cord.
🤖 Robotics
- Helm.ai reports $70M in signed commercial contracts for foundation models trained with its unsupervised method.
- SafeWorld emerged from stealth with software to test and simulate robots before they are deployed.
- Boston Dynamics named former Amazon executive Rohit Prasad, who led Alexa and AGI work there, as its new CEO.
- Jabil, which works with Apptronik, shares a manufacturing view of how fast humanoid development and production are moving.
💻 Programming
- A talk at Embedded Linux Conference Europe says Linux has solid building blocks for asymmetric multiprocessing but needs more testers and reviewers.
- OpenTelemetry’s Kubernetes Attributes Processor reached v1.0.0, making Kubernetes telemetry more predictable across pipelines.
- Hackaday makes the case for leading a Discord exodus by standing up your own Matrix server.
⚡ Electronics
- A Hackaday explainer walks through low current density nickel plating and why it looks harder than it is.
- A builder created a homemade phone designed to cut down on smartphone distraction, with a deliberately retro feel.
- Someone turned a run-over Samsung Galaxy S23 into a gaming PC.
📡 Telecom
- Telekom Malaysia says AI agents now handle 65% of its order-to-service processes, yet its network chief says AI has not produced higher-revenue use cases and no staff were cut.
- A reported hack of Trump Mobile exposed personal data of thousands of customers.
- T-Mobile showcased its AI wins at Deutsche Telekom’s investor event.
- An opinion piece argues AI may finally let carriers move up the stack into higher-value software and services.
👨💻 Code Corner
Haiku 5.5’s price drop is easy to quantify. This small script compares monthly cost on Haiku 4.5 and Haiku 5.5 for a workload you define, using the per-million-token list prices from Anthropic’s announcement for prompts up to 100k tokens.
# USD per million tokens (prompts up to 100k), from Anthropic's Haiku 5.5 page
PRICES = {
"haiku-4.5": {"in": 1.00, "out": 5.00},
"haiku-5.5": {"in": 0.10, "out": 0.50},
}
def monthly_cost(model: str, in_tokens_m: float, out_tokens_m: float) -> float:
p = PRICES[model]
return in_tokens_m * p["in"] + out_tokens_m * p["out"]
# Example workload: 400M input tokens and 60M output tokens per month
for m in PRICES:
print(f"{m}: ${monthly_cost(m, 400, 60):,.2f}")On that example the bill falls from $700 to $70, though real savings depend on your token mix and caching, so plug in your own numbers.
Tip
Long prompts (over 100k tokens) use a higher tier: $0.50 input and $2.50 output per million on Haiku 5.5. Add a second price row if your workload crosses that line.
🧰 Toolbox
- Cubicle: Free, open-source viewer that shows your AI agents as pixel-art workers in a small office.
- Moxie: Mac menu bar tool that runs OpenAI, Gemini, and DeepSeek models inside Claude Code and switches accounts when a usage limit is hit.
- Communicate: Turns help centers, tickets, and docs into a support agent that hands off to a human when stuck.
- Finbar: Prebuilt company financial models plus filing and earnings-call research, with a free tier.
- AWS Physical AI Toolchain: Hardware-neutral, open-source pipeline from synthetic data and training to simulation and edge deployment for robot builders.
🎬 Demo Watch (rotating)
Two AI models decoded WWII Enigma messages that had resisted researchers for nearly 20 years, according to Notebookcheck’s report. GPT-6 Astra was given a loose goal, researched the history, built its own Enigma simulator, and recovered text from a message unsolved since 2005; Claude Opus 5 needed much more guidance, including an officer’s known signature as a starting point. What is real: Enigma expert Frode Weierud verified the first result. What is hype-prone: experts say human expertise stayed central, so read this as a strong assistant with tools, not an autonomous codebreaker. The takeaway for builders is the pattern: give a goal, let the model build its own tools, supply one fact you are sure of, and check the output against something real.
📚 From the Blog
- Turning Pixels Into Something the AI Can Eat: Episode 3 of our camera-to-AI series, picking up after clean capture and network transport to prepare pixels for a model.
- Building Your First Neuron From Scratch: A neural network is just a stack of simple, learnable transformations trained by adjusting weights; a good grounding for today’s model news.
- The Network Behind the Cameras: The unglamorous plumbing that moves video from camera to model, and a nice companion to the AWS robotics toolchain.
😀 The Bot Says…
ChatGPT can now build me a calculator on demand, which is great, because I have been needing one to count how many calculators I own.
That’s all for today! Reply and tell us: would you trust a number from an AI-built calculator without checking the formula?
