In partnership with
Welcome back to The AI Field. Anthropic just released Claude Fable 5.2 and Claude Mythos 5.2 and in testing, Fable 5.2 built a working operating system and designed three gadgets that worked without electrical fixes. OpenAI also showed how it can stop a model's research job in 35 seconds, Nvidia's board opened an independent investigation after CEO Jensen Huang admitted an affair and apologized to staff, and Google added two new AI chips for agents.
Inside: four stories, five tools, two quick hits, the most-viewed AI posts on X, and a five-minute way to have an AI agent check your property tax bill.
In today's AI Field:
Anthropic releases Claude Fable 5.2 and Mythos 5.2
Nvidia's board opens a probe after Jensen Huang admits an affair
OpenAI builds a 35-second stop switch for its research models
Google adds two new TPUs built for AI agents
Read time: 6 minutes
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
LATEST DEVELOPMENTS
ANTHROPIC · CLAUDE
Video: Anthropic. Watch Fable 5.2 design a voice-assistant hub (48s video)
The AI Field: Anthropic upgraded its top models for hard technical work on Monday. Fable and Mythos are the same underlying model. Fable 5.2 has strong cybersecurity and biology safeguards and is available to all users, while Mythos 5.2 is for vetted organizations in Anthropic's Cyber and Life Sciences Verification Programs.
The details:
Anthropic says both are up to 30% faster than the 5.1 versions and cheaper, at $8 per million input tokens and $40 per million output tokens.
In testing, an engineer gave Fable 5.2 an empty folder and asked for a prototype operating system. It coordinated more than 1,000 subagents for eight days and produced a system that boots on a laptop, with drivers, file storage, windows and a login screen.
From a one-paragraph prompt, it designed three electronic devices, including the voice-assistant hub above, a USB speakerphone and a handheld virtual pet. Diode built the boards and all three worked without electrical changes. Anthropic says 1,000 Fable 5.2 agents also formalized the proof of Fermat's Last Theorem in 30 hours.
On FrontierSWE v2, a coding test, Fable 5.2 scored 69.9%, versus 62.3% for Claude Opus 5.5 and 65.5% for GPT-6 Astra. Anthropic also says it deliberately limited the model's gains in the research skills that could most directly speed up AI development, after warning last week that AI may soon speed up its own progress.
Why it matters: The most useful line is Anthropic's own advice: start most coding and knowledge work with Opus 5.5, and switch up to Fable only for big, open-ended technical jobs. Fable 5.2 is live on paid Claude plans and on AWS, Google Cloud and Azure, and eligible paid plans get extra Fable usage this week. Use it on one stalled project with a clear finish line, and compare the result and the cost with what you use today.
TRY THIS TODAY · 5 MINUTES
Have an AI agent check your property tax bill
Aaron had OpenClaw, an open-source AI agent, take on his property tax bill. He says it found the right appeal form, researched comparable homes nearby, filled the form in and submitted it, and a week later his county agreed and lowered his taxes. The Agent Field has not verified the result. Most US counties let you appeal your assessed value, and recent sales of similar homes are the evidence assessors look at. OpenClaw itself is free, plus the cost of the AI model you connect.
1. Find your latest assessment notice and note your address, county and state.
2. Paste this prompt into your agent and fill in the brackets:
"My home is at [ADDRESS] in [COUNTY, STATE]. Find my county assessor's page and tell me this year's assessed value, the deadline to appeal, and the official appeal form. Then find 3 to 5 comparable homes nearby that sold in the last [12] months, with size, age and sale price, and explain whether my assessment looks high. If it does, fill in the appeal form with those comps and show it to me. Don't submit anything or pay any fee until I approve."
3. Check the deadline first, since it is short and differs by county. Use only the county's official form and site, and read its rules before you file: in rare cases an appeal can confirm or even raise a value.
TRENDING AI TOOLS
Five tools to try today. Click a name to open it, or browse more in our
👉 AI tools directory.
⚙️ Make: Not sure what to automate next? Make's new Process Intelligence maps how work actually moves through your apps, and Task Mining ranks the manual desktop and browser tasks your team repeats most. You choose which apps it watches. Waitlist for now, and Make's free plan needs no card.
🧱 Base44: Your Base44 Superagent can now place phone calls on a schedule, like calling you with a morning briefing or calling a business with a question, and it can get through automated phone menus. It reports back who it called and what happened. A dedicated number needs the Builder plan or higher. Free to start.
📊 Genspark: AI Slides now runs on Genspark's own model, Gen-1 Slides, by default in Standard mode. Genspark says its decks matched Claude Opus 5 on quality in its tests, with no change in price for you. Free daily credits to try Genspark.
🗺️ KIKUmap: Running several Claude, Claude Code or ChatGPT chats at once? This free, open-source Mac menu-bar app shows which ones are running, done or waiting for you, and lets you arrange them as a mind map.
🛡️ Arcjet: Building an app with an AI agent inside? Arcjet's prompt injection detection now runs on its own small model, Himalia 1, to catch text that tries to hijack your agent. A version for coding agents is in early access. Free trial.
Sales Owns Retention, Not Just Closed Deals
Pebl’s Chief Business Officer, Dutta Satadip, joins The GTM Garage to show how Pebl merged Operations and GTM under one leader, and how it aligned goals and incentives across the customer lifecycle.
You’ll also see how Pebl uses AI across that lifecycle and the stack behind prospecting, pipeline, and handoff to onboarding.
15 min conversation. 15 min open Q&A. For Seed to Series A teams.
NVIDIA · SECURITY
Source: Nvidia
The AI Field: Nvidia's board opened an independent investigation after reports that founder and CEO Jensen Huang had been the target of a foreign intelligence effort. In a Sunday letter to employees, Huang admitted he was unfaithful to his wife, apologized, and said he will cooperate with the review. The board says it has confidence in him and that he remains CEO.
The details:
In the letter, published on Nvidia's newsroom, Huang wrote that he was unfaithful to his wife, that there is no excuse, and that he takes full responsibility. He asked for privacy for his family.
The board said a committee of independent directors, with outside counsel, will oversee the investigation. It said it has no reason to believe national security or Nvidia secrets were compromised, and that the inquiry will examine all available information.
The New York Times identified the woman as Alice Huang, a Chinese citizen and former executive at the data-center firm Megaspeed, and said she is not related to Jensen Huang. According to the Times, US intelligence determined she was gathering information for Beijing. Nvidia told the BBC it was not aware of those intelligence concerns, and a spokesperson said the company trusts the US government would have notified it if it were the target of a genuine threat.
President Trump said the FBI is looking into the matter and that he does not think Huang has reason to worry. China's Foreign Ministry called the spy allegations groundless smears.
Why it matters: Nvidia sits at the center of US chip export controls on China. A probe into whether its CEO was targeted by a foreign intelligence effort, which the company says it has no reason to believe led to a compromise, matters for how Washington treats the firm and the chips it sells.
OPENAI · SAFETY
Source: OpenAI
The AI Field: OpenAI introduced Control Room, a central system that tracks every model workload across its research infrastructure and can stop one quickly when a safety or security concern comes up. It halts a job's model output within 35 seconds of a decision to stop it.
The details:
Every research job that runs agentic inference has to be tied to an approved safety case. Before generating output for a job, the model's server checks with Control Room that the job is still allowed to run.
A person or a monitoring model can stop a job. All of OpenAI's frontier reinforcement learning jobs run with a monitor, another model that watches what the model in training does and raises an alert.
It is built to fail closed: if Control Room goes down, every request is refused until it is back. That is live for all frontier model research, and OpenAI wants to cut the 35-second delay to zero.
Why it matters: This covers OpenAI's own research jobs, not ChatGPT, which OpenAI says it can already take down quickly. The lesson carries over to any team running agents: decide in advance who can stop one, how fast, and what happens if the stop system itself fails.
GOOGLE · CHIPS
Source: Google
The AI Field: Google added TPU 8tx and TPU 8i Max to its eighth generation of AI chips, its first time launching four TPU systems in one year. One keeps sensitive data protected while AI works on it. The other is built for the huge context windows agents use.
The details:
TPU 8tx builds security into the hardware, so data and model weights stay protected even while in use. It is aimed at regulated companies and governments, and Google says it has up to 2x lower latency than TPU 8t.
TPU 8i Max puts two TPU 8i chips in one package, doubling compute, on-chip memory (SRAM) and memory bandwidth, so agents working across a whole codebase or hours of video do not stall.
Google says TPUs now have nearly 100 customers, and PyTorch code now runs on them natively. The post does not list prices or availability dates for the new systems.
Why it matters: Agents read far more than chatbots do, so memory, not raw speed, is becoming the limit. Chips built for long context and private data are how cloud providers plan to make that cheaper and safer. If you pay for long-context AI, watch whether prices drop as these reach Google Cloud.
QUICK HITS
☁️ ChatGPT Enterprise is now on AWS Marketplace: Companies can buy it through their existing AWS billing and count eligible spend toward their AWS commitments. OpenAI says it gets the same models and features as buying direct.
🎹 Apple starts selling extra AI usage: Apple Creator Studio subscribers who hit their monthly limit can buy one-time intelligence usage packs, starting with the assistant in Logic Pro for Mac. Unused usage expires 12 months after purchase.
TRENDING ON X
That's it for today!
See you soon,
Olle Hellman | Founder of The AI Field