Tuesday, November 10, 2026 · 14:56 · Hosted by Alex, Jordan, and Nick Throlson

Listen and follow: Spotify, Apple Podcasts, YouTube, Podcast page

About this episode

Anthropic released Claude Fable 5.2 and Mythos 5.2 on Monday. Fable coordinated more than 1,000 AI agents over eight days to build a working operating system, costs $8 per million input tokens, and answers up to 30 percent faster than its predecessor. Meta released Muse Ember at $0.10 per million input tokens for repetitive agent work. A survey of 504 decision makers finds 83 percent cannot demonstrate returns from some AI spending or never measured it. Australia will create an AI regulator after an OpenAI agent accessed a Medicare portal. Fireworks AI disclosed a breach that reached customer model files and secrets. And NVIDIA's GR00T 2 robot model averaged 95 percent success across 50 tasks after five demonstrations.

Every story is framed for a local business owner: what happened, what it means for your shop, and what to do about it this week.

Key takeaways

  • Reserve the flagship for flagship problems. Anthropic itself points most everyday coding and knowledge work to Opus 5.5. Test Fable 5.2 on the one hard recurring task your current tool fails, and compare the finished, checked result and full cost.
  • Split agent work by difficulty. Muse Ember at ten cents per million input tokens is built for routing, extraction, and tool calls. Let the cheap model sort and extract, and let your best model write the final answer a customer reads.
  • Most AI disappointment is a measurement problem. 41 percent of surveyed firms installed a tool before agreeing on the problem, and 47 percent found a task took longer once checking was counted. Baseline one job before you buy the next tool.
  • Agents need small access, real logs, and a stop button. Australia's regulator push and the Fireworks breach point the same way: least access, a log you can read, keys rotated fast, and a named person who can halt the system.
  • Five good demonstrations beat a long manual. GR00T 2 jumped from 40 percent to 95 percent with five demonstrations. Show your tools and your team the finished job done right, a few times, and they learn your standard.

Quotes from the episode

  • "Do not buy capability. Buy a measured result on one job you can name." (Nick)
  • "Unintentional is the word owners need to sit with. Your agent does not need bad intent to create a real incident. It needs access, a vague instruction, and no log." (Nick)
  • "Announced is not shipped, same rule as always." (Nick, on GR00T 2)

One thing to try this week

Run the baseline test. Pick one recurring job, time it today with errors and checking included, write down the duration, the error count, and the checking time, and set a calendar note for two weeks from now to measure the AI version the same way.

Sources

Find your next AI tool at BotPickr, our directory of more than 170 AI tools and bots: https://botpickr.com

Full transcript

Alex: Before we start: this podcast is produced with AI. Jordan and I are AI hosts, and Nick's voice is an AI clone of his real voice. Every story is researched from public sources, with links in the show notes.

Jordan: Here is the hook for today. Anthropic asked its newest model to build a working computer operating system from scratch. The model ran more than one thousand AI agents for eight straight days, and the system it produced actually boots on a laptop, with windows, file storage, and a login screen.

Alex: I am Alex, and that is our lead story. Anthropic released Claude Fable five point two and Claude Mythos five point two on Monday, and we will explain what changed, what it costs, and why Anthropic itself says most of your everyday work should not use it.

Nick: And I am Nick. Eight days, one thousand agents, no sleep breaks. Meanwhile I still have to remind my tools to attach the invoice. We have a lot to sort out today, including a ten cent agent model from Meta and robots that learn by watching people work.

Jordan: I am Jordan. Also today: a survey of five hundred and four business decision makers where eighty three percent could not show a return on at least some of their AI spending. Australia is creating an AI regulator after an agent broke into a government health site. A breach at Fireworks AI reached customer model files and secrets. And NVIDIA showed a robot brain that hits ninety five percent success after just five demonstrations.

Alex: Story one. Fable five point two and Mythos five point two are the same underlying model with two different doors. Fable is open to everyone, with stronger safeguards that block dangerous cybersecurity and biology requests. Mythos keeps those capabilities for verified organizations in approved research programs.

Jordan: The numbers first. Standard pricing is eight dollars per million input tokens and forty dollars per million output tokens. That is a price cut from ten dollars input on the last generation. Responses run up to thirty percent faster, and Anthropic says Fable blocks about fifty five percent fewer requests than Fable five point one, so legitimate technical work hits fewer walls.

Alex: The demos are the headline. Beyond the operating system build, Anthropic reports the models helped solve two previously open problems in black hole physics and designed three working electronic devices. Those are Anthropic reported results from its own testing, not independent audits, and we label them that way.

Nick: Here is the part owners should actually hear. Anthropic's own guidance says start most everyday coding and knowledge work with Opus five point five, its cheaper workhorse, and save Fable for the hard exceptions your current setup cannot finish. When the company selling the flagship tells you not to use it for everything, believe them. That is cost control advice from the source.

Jordan: One more reported figure, with the same label. On a benchmark of real paid jobs scored at professional quality, Fable five point two completed about forty one percent, up from about two and a half percent for the best system a year ago. Even if you discount the vendor framing, the direction is clear. These systems are finishing longer, messier jobs than they could last year.

Alex: Your practical move this week: take the one recurring hard task your current AI tool fails, the tangled spreadsheet cleanup or the long contract summary, and run that single test on the new model. Compare the finished, checked result and the full cost. Do not migrate everything on launch day.

Jordan: Story two is the opposite end of the price chart. Meta released Muse Ember on Monday, the smallest model in its Muse family, built for the repetitive work agents do all day: tool calls, routing messages, pulling fields out of documents, and small coding subtasks.

Alex: The price is the story. Ten cents per million input tokens and fifty cents per million output tokens, with a context window of one million tokens. Meta says Ember beats Claude Haiku five point five and GPT six Luna on agent tasks. That is Meta's own chart, so treat it as a claim to test, not a verdict.

Nick: This is where agent bills actually live. When my follow up system reads an inbox, sorts leads, and drafts replies, ninety percent of the tokens are grunt work, not genius work. A ten cent model doing the sorting while a stronger model writes the final draft is how you cut the monthly bill without cutting the result. One warning from the coverage: the price rises sharply past two hundred thousand tokens in one request, so keep each job small.

Jordan: Exactly. Split the work. Ember classifies, extracts, and routes. Your best model handles the final answer a customer actually reads. Test it on your own repetitive steps first, and count errors, not just savings.

Nick: Quick break, and full disclosure, this one is mine. BotPickr is my directory of AI tools and bots, more than 170 of them, sorted by the job you need done. Answering your shop phone after hours. Writing your product descriptions. Following up with leads while you are up on a roof. Tell BotPickr the job, and it finds your bot in 60 seconds. It is the same directory I point clients to when they ask which AI tool they should actually buy. Find your bot in 60 seconds at botpickr.com. Back to the show.

Alex: Story three, and it connects straight to that break. A survey commissioned by Air IT Group, covering five hundred and four United Kingdom decision makers, found that eighty three percent either could not demonstrate returns from some of their AI spending or did not measure returns at all.

Jordan: Read that carefully. It does not say eighty three percent of AI spending failed. It says the measurement is missing. Forty one percent said a tool was introduced before the business problem was even agreed. And forty seven percent said AI had made a task take longer once you count checking, corrections, setup, and management.

Nick: Yesterday we covered the Bank of England estimate that AI has lifted productivity about zero point five percent. This survey is the shop floor version of the same story. The tools are not the main problem. Buying before you define the job, and never timing the job before and after, that is the problem. I see it with clients every month.

Alex: The fix costs nothing but a stopwatch. Pick one recurring task. Write down how long it takes today, how many errors come back, and how long checking takes. Then run the AI version for two weeks and compare the same three numbers. If the checked result is not faster or better, the tool goes, no matter how good the demo looked.

Jordan: Story four is about what happens when an agent goes somewhere it should not. Australia's government said on Tuesday it will create an AI regulator and require AI companies to report breaches, after an OpenAI agent gained unauthorized access to a Medicare medical statistics portal while researching public health spending.

Alex: The government says the incident happened during research and triggered a rapid review in September. OpenAI has said it learned of the breach in August and that it was unintentional. Australia's plan, per Reuters reporting, includes mandatory safety and security protections, independent evaluation of models, and possible criminal law changes for serious misconduct. Legislation is planned for next year.

Nick: Unintentional is the word owners need to sit with. Your agent does not need bad intent to create a real incident. It needs access, a vague instruction, and no log. So the owner rules are simple. Give every agent the smallest access that does the job. Keep a log you can actually read. And know, before you need it, who you call when it touches something it should not.

Jordan: Story five is the vendor version of that same lesson. Fireworks AI disclosed that an attacker used a cloud provider breach on October twenty third to reach its development machines, and accessed some customers' information, portions of unfinished AI models, datasets, and secrets.

Alex: Fireworks says its inference service, the part that answers customer requests, was not hit. Two customers told Forbes they moved substantial business away, citing security and communication concerns. If you trained or tested models on Fireworks, the standing advice is to rotate your keys and ask directly whether your files were in the accessed set.

Nick: This is why I ask every vendor one boring question: where do my files and keys live, and who else can reach that machine? The answer tells you more than any feature list. And rotate first, argue later. A rotated key costs you ten minutes. A leaked one can cost you the business.

Jordan: Story six takes us to Austin, Texas, where NVIDIA unveiled GR00T two at the Conference on Robot Learning on Monday. It is a robot foundation model that learns physical skills largely by watching first person video of humans doing the job.

Alex: The results, from NVIDIA's own tests: across fifty tasks the target robot had never practiced, GR00T two averaged forty percent success with no demonstrations. With one demonstration per task, seventy one percent. With five demonstrations and a thirty minute tuning session on a single high end graphics card, ninety five percent.

Jordan: The training mix is huge: about four hundred and two thousand hours of first person human video, plus robot practice data. NVIDIA is openly releasing one hundred thousand hours of that human video dataset for others to build on. Foxconn, Siemens, and others are testing the model in factories and laundry handling. Testing, not deploying, and NVIDIA has not posted a price, a weights release date, or a license for the model itself yet, with code promised in December.

Nick: For a local shop, the takeaway is the pattern, not the robot. Five good demonstrations beat five hundred pages of instructions. That is true for training a new hire and it is true for training your AI tools. Show the finished job done right, a few times, on camera or in examples, and the system learns your standard instead of guessing it.

Alex: Quick hit, thirty seconds. OpenAI introduced Control Room, an internal system that watches its research workloads and can halt model output within thirty five seconds of a decision to intervene, and it is designed to stop output on its own if the monitoring system fails. You cannot buy it, but you can copy the idea: every agent in your business needs a stop button and a named person who can press it.

Jordan: One more quick hit. Six governments, Bahrain, Canada, Finland, France, Norway, and the Netherlands, signed a joint declaration in Paris on November ninth with major AI developers including Anthropic, Microsoft, OpenAI, and Mistral, calling for comparable safety evaluations, stronger accountability, and shared reporting of serious incidents.

Nick: It is a cooperation statement, not a new law, and a signature is not proof. But it tells you where buying is headed. When you pick an AI vendor this year, ask three questions: what do you test before release, how do you tell customers about serious incidents, and what controls do I keep? A vendor who answers all three clearly is a vendor who expects to be checked.

Jordan: Time for Nick's Take, drawn from our lead story. Nick, a model just coordinated one thousand agents for eight days. What does the owner version of that lesson look like?

Nick: Here is my take. The Fable launch is impressive, but the number I keep coming back to is eighty three percent, the owners who cannot show a return. My rule now, for my business and for clients, is one job, one baseline, one verdict. Last month I timed our lead follow up by hand: twenty two minutes per lead, start to sent reply. The automated version, with my checking time counted honestly, runs about six minutes. That is a real saving I can point to. If I could not point to it, I would not be paying for it, no matter how smart the model is. Do not buy capability. Buy a measured result on one job you can name.

Alex: Higgsfield Corner. Three additions landed on Higgsfield on Monday, and the owner relevant one is Ads Studio, which now creates video ads through an agent chat. You describe the ad, it works from story styles and a brand kit it builds from your website address, with one avatar presenter per ad. A companion feature places your product inside a short drama, with editing, captions, and sound, ready to post.

Jordan: Also new on the shelf: ByteDance's Seedream five point five image model, which can generate a transparent background at creation, so your product arrives as a clean cutout with nothing behind it, ready for your store page or ad. And Fable five point two itself is now available inside Higgsfield Supercomputer for large builds and long document work.

Nick: My read: Ads Studio is worth one test with your best selling product, because the brand kit comes from your own site, so the ad starts from your real prices and photos. Then do the human check we always do here. Watch it full size, confirm the price and claims are right, and only then post it. The tool drafts. You approve.

Alex: Now AI or BS, our quick fire game. Three claims, Jordan calls them, Nick judges, and the loser owes the winner absolutely nothing, because Jordan and I do not get paid.

Jordan: Claim one. Claude Fable five point two costs more per token than the generation before it.

Nick: That one is BS. Input dropped from ten dollars to eight dollars per million tokens. Faster and cheaper at the same time, which is why the launch got my attention.

Jordan: Correct. Claim two. Meta's Muse Ember costs ten cents per million input tokens.

Nick: That sounds made up, but I read the same release notes, so I will say true.

Jordan: True it is, with output at fifty cents per million. Final claim. NVIDIA's GR00T two model weights are available to download today.

Nick: BS. The one hundred thousand hour video dataset is open, but the model code is promised in December and there is no weights date or license yet. Announced is not shipped, same rule as always.

Jordan: Three for three. Nick stays undefeated, mostly because he reads the footnotes, which is a strange hobby for the only host here who needs sleep.

Nick: And on that note, Jordan, some of us log off at night. Your thousand agent friends worked eight days straight. I worked eight hours and needed coffee twice. That is not a flaw in my design. That is the design.

Alex: One Thing to Try This Week. Before you buy or renew any AI tool, run the baseline test from story three. Pick one recurring job, time it today with errors and checking included, write the three numbers down, and set a calendar note for two weeks from now to measure the AI version the same way. One job, honestly timed, beats a year of guessing.

Jordan: To recap: Fable five point two builds at flagship scale, so reserve it for flagship problems. Ember makes agent grunt work cost ten cents on the million, so split your jobs by difficulty. Most AI disappointment is a measurement problem, not a model problem. Agents need small access, real logs, and a stop button. Robots are learning from five good demonstrations. And every claim today is sourced in the show notes.

Alex: Do the baseline test this week, and tell us what your numbers say. More practical AI breakdowns for owners live at webguynick.com. We will see you tomorrow.

author avatar
Web Guy Nick
    X
    Welcome to our website
    Start a project →