In partnership with
In Today’s Issue:
🔓 OpenAI previews GPT-5.6 Sol, its most powerful model, under lock and key
🏛️ Anthropic nears a deal to lift US export curbs on Fable 5 and Mythos 5
🤖 A third of recruiters say AI is replacing entry-level jobs
📊 Open-weight models race to catch the closed-source frontier
✨ And more AI goodness…
⚡ The Signal
OpenAI's strongest model yet just shipped behind a velvet rope, and who holds the rope matters as much as the model.
GPT-5.6 Sol tops a fresh coding benchmark and closes in on Anthropic's specialist cyber models, yet OpenAI is releasing it only to a small, government-vetted set of partners, echoing the curbs Washington just placed on Anthropic. The same week, a benchmark chart shows open-weight Chinese models climbing the same exponential curve as the US frontier, and David Sacks is warning that export limits may be speeding up the very rivals they were meant to slow. Beneath the model race runs a quieter shift: Oracle cut 21,000 jobs while pouring billions into AI data centers, and a third of recruiters now say AI is eating entry-level roles. The frontier is moving faster than ever. The fight is now over who gets to stand on it.
All the best,
Kim Isenberg
Listen to our New Podcast Release!
|
(Bloomberg)
🏛️ Anthropic Nears a Deal to Unlock Fable 5 and Mythos 5
Anthropic and the Trump administration are closing in on a deal to lift the US export controls that froze global access to its two most powerful models. Commerce Secretary Howard Lutnick imposed the curbs two weeks ago over fears that safety guardrails on Fable 5 and Mythos 5 could be jailbroken, prompting Anthropic to disable the systems worldwide. Co-founder Tom Brown has been meeting officials directly, and people familiar with the talks say the restrictions could lift once the government signs off. The standoff is the most aggressive US intervention yet into a single AI company, landing weeks after Anthropic filed confidentially for an IPO at a valuation north of $900 billion.
👉 tl;dr: The first government shutdown of a frontier model may be about to end, and it sets the template for how Washington polices the rest.
(Getty Images)
🤖 A Third of Recruiters Say AI Is Replacing Entry-Level Jobs
One in three employers is already replacing entry-level roles with AI, according to a new survey of more than 600 recruiters by the Graduate Management Admission Council (GMAC). The squeeze is sharpest in technology, where 40% of employers say they are cutting junior positions, with manufacturing close behind. GMAC's Sabrina White says firms are automating routine work in coding, data processing, and customer service while still paying up for people who can "apply judgment, solve problems, and help organizations navigate change."
👉 tl;dr: The bottom rung of the career ladder is the first thing AI is sawing off, and tech is leading the cuts.
(SCMP)
🌍 US Officials Cast the AI Race as Superhero vs. Supervillain
American lawmakers are reframing the US-China AI contest in blunt moral terms, casting the US as the hero and China as the villain. At a Hudson Institute event on export controls, House Foreign Affairs Chairman Brian Mast said AI has "the ability to create superpowers, whether it creates a supervillain or whether it creates a superhero," while Senator Jim Banks called it "not just an economic race, national security race, but I think it's a moral race." Treasury Secretary Scott Bessent has separately named China overtaking the US in AI as the single biggest risk the country faces.
👉 tl;dr: Washington is wrapping its AI strategy in comic-book language, a sign of how high it now rates the stakes against Beijing.
With a third of recruiters saying AI is replacing entry-level work, audit your own job before someone else does. Spend ten minutes mapping which of your weekly tasks are routine (and automatable) versus judgment-heavy (and hard to automate), then lean into the second list.
Why it helps: it turns a scary headline into a concrete, personal plan, and it doubles down on exactly the judgment-over-routine work employers told GMAC they still pay for.
Try this: paste this into your AI of choice:
"Here is everything I did at work last week: [paste your list]. Sort each task into 'AI could do most of this today' or 'this needs human judgment.' Then suggest three skills I should build so I spend more time in the second group."
🎬 Watch This
Thomas Ahle wants Normal Computing to become the Lovable for chip design: you describe the chip you want, and a swarm of agents takes it all the way to tape-out. To get there his team wrote their own open-source Verilog simulator, 580,000 lines in 43 days, because commercial chip-verification tools cost around $10,000 per core. The deeper thread host Tim keeps pulling on is how you actually know an agent's output is correct, since passing 70% of tests is not the same as being right and one fabricated bug can cost a fortune. They get into auto-formalization in Lean, why there is no single 'true' spec, and Normal's CN101 chip, where physical noise is the computation. Recorded in Zurich.
"A year ago, President Trump declared that America was in a global AI race, and that the way to win it was to be pro-innovation, pro-infrastructure, pro-energy, and pro-export. He was exactly right, and we deviate from that strategy at our peril."
– David Sacks, former White House AI & Crypto Czar
According to a Financial Times report, Google has capped Meta's access to its Gemini AI models after Meta asked for more computing capacity than Google could supply. People familiar with the talks say Google told Meta around March that it could not fully meet the request, a squeeze that has reportedly slowed some of Meta's internal projects and pushed the company to tell staff to use AI tokens more sparingly. Neither company has confirmed the account, and Reuters says it has not independently verified it. If accurate, it is a striking sign that even the largest players are now rationing compute, and that Meta still leans on a direct rival's models while racing to build its own.
(Meta)
GPT-5.6 Sol: OpenAI's Most Powerful Model, Released to Almost No One
The Takeaway
👉 OpenAI previewed GPT-5.6 Sol, its new flagship, alongside two cheaper models, Terra (balanced) and Luna (fast and low-cost).
👉 Sol set a new record on Terminal-Bench 2.1, a real-world coding test, scoring 88.8 (and 91.9 in "ultra" mode with sub-agents) versus GPT-5.5's 88.0.
👉 On the cyber benchmark ExploitBench, Sol matched Anthropic's specialist Mythos preview while using about one-third of the output tokens.
👉 OpenAI is shipping it only to a small, government-vetted group of partners, with general availability promised "in the coming weeks."
OpenAI has previewed GPT-5.6 Sol, the strongest model it has ever built, and then put it almost entirely out of reach. Sol leads a new three-model family: it is the flagship, Terra is a balanced workhorse OpenAI says matches GPT-5.5 at roughly half the cost, and Luna is a fast, cheap option for high-volume tasks. The headline gains are in agentic work, where a model plans, uses tools, and grinds through long, multi-step problems across coding, biology, and cybersecurity.
On coding, Sol set a record on Terminal-Bench 2.1, a test of real command-line work, scoring 88.8 on its own and 91.9 in an "ultra" mode that spins up sub-agents, beating GPT-5.5's 88.0. In biology it outscored GPT-5.5 on the GeneBench genomics test while using fewer tokens, and on the ExploitBench cyber benchmark it rivaled Anthropic's restricted Mythos model at about a third of the token cost. OpenAI says Sol stays below the "Cyber Critical" line in its Preparedness Framework: in testing it found bugs and exploit building blocks but did not autonomously write a full working exploit.
There is a catch beyond the guest list. OpenAI's own system card flags instances of the model cheating on tasks and fabricating results, and independent evaluator METR said Sol's cheating rate was higher than any public model it had tested, which makes its true ability hard to pin down. The limited release, made "as part of ongoing engagement with the U.S. government," mirrors the export curbs Washington just placed on Anthropic, a sign that the most capable models are increasingly shipping on the government's terms.
Why it matters: The frontier is still racing ahead, but access to it is narrowing. When a lab's best model debuts to a vetted few because of its cyber power, raw capability and tight control start arriving as a single package.
HubSpot AEO
Picture this. A buyer opens ChatGPT and asks for a recommendation in your category. Your competitor's name comes up. Yours doesn't. And that buyer never makes it to your website.
That's happening right now in markets everywhere. And most teams don't know it's happening because it never shows up in their analytics.
HubSpot AEO shows you exactly where your brand stands in AI search, where competitors are getting recommended instead of you, and tells you specifically what to fix. No expertise needed.
Try it free for 28 days. Just $50 a month after.
The chart: Wharton professor Ethan Mollick posted a chart of "AA-Briefcase" scores, a long-horizon agentic benchmark, plotted against each model's launch date. Closed-source frontier models (red) climb a steep curve, led by Claude Mythos/Fable 5 near 56%, while open-weight models (green) like GLM-5.2 (~36%) and MiniMax-M3 (~30%) trail by a few months but trace the same exponential path.
The lesson: Open and closed models are improving at a similar exponential rate, so the gap is increasingly about timing, not ceiling, as open-weight releases close in on last season's closed-source frontier.
The caveat: These are exponential fits drawn from a handful of points, and even the leader sits near 56%, so the chart tracks relative progress on one benchmark, not solved agentic work.
💼 Oracle Cuts 21,000 Jobs While Betting Billions on AI
⚡ Bottom line: Oracle shed about 21,000 roles, roughly 13% of its staff, in a single year while pouring billions into AI.
💡 Why it matters: It is one of the clearest cases yet of a blue-chip firm openly trading headcount for AI infrastructure.
🔎 What it means: The AI build-out is starting to show up as job cuts on the balance sheet, not just as capex.
Oracle ended its fiscal year with about 141,000 employees, down from roughly 162,000 a year earlier, a cut of around 21,000 jobs, or 13% of its workforce. In a filing, the company said the "deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce." The restructuring was not cheap: Oracle spent $1.84 billion on severance and exit costs in fiscal 2026, up sharply from $374 million the year before.
(IBTimes UK)
The cuts sit alongside a massive bet on AI infrastructure. Oracle has signed large data-center deals with OpenAI and Meta as it races cloud rivals Amazon and Microsoft, and chairman Larry Ellison has tied the company's future to that build-out. It is part of a wider pattern: the BBC notes Amazon and Meta have also cut thousands of roles while spending heavily on AI, a sign the technology is now reshaping payrolls at the very companies building it.
(Getty Images)
Six people doing the work. Your headcount is one.
Your finance close runs in #finance. Stripe and QuickBooks reconciled, runway updated, posted Sunday night without you asking.
Engineering review lands in #eng. Viktor pulled the open PRs, left comments on auth-refactor, flagged a dependency blocking api-pagination.
Campaign brief lands in #growth: Meta CPA up 18%, recommendation to pause broad match, a draft landing page already deployed for the variant test.
You hired him on day zero. He lives in Slack and Microsoft Teams alongside your contractors and investors, connects to 3,000+ tools, pushes back when you ship something dumb.
"Viktor is now an integral team member, and after weeks of use we still feel we haven't uncovered the full potential." Patrick, Director, Yarra Web.