In partnership with
In Todayโs Issue:
๐ SpaceX buys Cursor for $60 billion
๐ Why switching off Anthropic could backfire
๐ Only 6% of companies make AI pay
๐ GLM-5.2, the open model at #2
๐ง DeepMind maps the road past AGI
โจ And more AI goodnessโฆ
โก The Signal
The edge in AI is shifting from owning the smartest model to being allowed to use one.
This week an open-weight Chinese model, GLM-5.2, climbed to #2 on a closely watched coding leaderboard, beating every version of Claude Opus while costing roughly one-sixth as much to run. It lands in a strange moment: Washington has suspended access to Anthropic's two most powerful models, so part of the US frontier is switched off even as Chinese labs ship freely downloadable weights. The money is flowing the other way, with SpaceX paying $60 billion for the coding startup Cursor. The defining question of 2026 is no longer only who trains the best system, but who can deploy one, and at what price.
All the best,
Kim Isenberg
SpaceX acquires Cursor maker Anysphere for $60 billion (TechCrunch)
๐ SpaceX Buys Cursor for $60 Billion
SpaceX has agreed to buy Anysphere, the maker of the AI code editor Cursor, in an all-stock deal valuing it at $60 billion. The merger, signed on June 16, folds one of the fastest-growing coding tools (around $2.6 billion in annualized revenue, from a 2022 start) into Elon Musk's rocket company, which had circled Cursor for months. SpaceX is betting that owning the tools developers live in is as strategic as the models underneath.
๐ tl;dr: The most valuable real estate in AI may be the editor, not the model.
Anthropic CEO Dario Amodei testifies on AI in Washington (Getty Images / Fortune)
๐ Switching Off Anthropic Could Backfire on US AI
Washington's decision to restrict access to Anthropic's most powerful models could dent the entire U.S. AI industry, analysts warn. OpenAI's and Anthropic's valuations rest on global adoption, and enterprises are wary of building on a model the government can switch off. "You can't rely on something that could be switched off," warned Deutsche Bank's Jim Reid. The bigger risk, says the Peterson Institute's Martin Chorzempa, is that export controls hand the market to open models: "That's a big advantage to open models."
๐ tl;dr: Cut off your own labs, and you push customers toward open weights.
Scale AI, which ran the enterprise AI survey (Scale AI)
๐ Only 6% of Companies Have Made AI Pay
Just 6% of enterprises have deployed AI across the business and seen real returns, a new Scale AI survey of nearly 500 senior decision-makers finds. The result echoes an earlier MIT study that put the success rate at 5%. What sets the 6% apart: they reach measurable results within six months, trust their data foundations, and lean on specialist vendors rather than off-the-shelf tools alone.
๐ tl;dr: The AI divide is no longer about access to models, it is about execution.
Stop paying frontier prices for work an open model can already handle.
Why it helps: Most teams default to one expensive model out of habit. A quick head-to-head often surfaces an open model that is 80 to 90% as good for your task at a fraction of the price, especially on long-context work.
Try this: Paste this into a long-context model (open or closed): "Here is my entire [codebase / contract / research folder]. Act as a senior reviewer. List the ten biggest risks or weaknesses, ranked, each with its exact location and a one-line fix. Be specific and skip the praise."
With GLM-5.2 and DeepSeek V4 now matching closed models on long, messy coding tasks at a fraction of the cost, the cheapest upgrade this week is a five-minute bake-off before your next big job.
๐ฌ Watch This
Andrej Karpathyโs introduction to Large Language Models is already two years old, but still one of the clearest explanations of how systems like ChatGPT, Claude, and Gemini work. In one hour, he covers inference, training, hallucinations, finetuning, RLHF, tool use, multimodality, customization, and the idea of LLMs as a new kind of operating system - before turning to security risks like jailbreaks, prompt injection, and data poisoning. The field has moved fast since 2023, but many of todayโs core AI debates are already laid out here with remarkable clarity.
Open source AI models will soon become unbeatable. Period.
โ Yann LeCun, Chief AI Scientist at Meta (X, October 2023)
Microsoft appears to be having it both ways with China's hottest open model. Even as it bans the consumer DeepSeek app on employee devices over data-security concerns, the company is reportedly testing a Microsoft-hosted version of DeepSeek V4 inside Copilot Cowork, wrapped in Azure's compliance controls and sold to enterprises. The move reads as a quiet test of how far Washington's patience stretches: US lawmakers have pushed to add DeepSeek to the Commerce Department's blacklist, even as the government has so far held off. Hosting a Chinese frontier model while restricting an American one would be the kind of contradiction only 2026 could produce.
GLM-5.2: An Open Model at the Frontier's Door
The Takeaway
๐ GLM-5.2, a freely downloadable Chinese model, reached #2 on Arena's frontend coding leaderboard (an Elo of 1,595), behind only the suspended Claude Fable 5.
๐ It runs at about $1.40 per million input tokens and $4.40 output, roughly one-sixth the price of OpenAI's GPT-5.5.
๐ It ships with a 1-million-token context window and an MIT license, so anyone can run, inspect, or fine-tune it.
๐ The gap between open and closed frontier models is now measured in a handful of leaderboard points, not generations.
An open-weight model from a Chinese lab has, for the first time, muscled into the top two of a frontier coding leaderboard. GLM-5.2, released this week by Chinese lab Zhipu, sits at #2 on Arena's Code Arena Frontend ranking with an Elo of 1,595, ahead of every Claude Opus variant and behind only Claude Fable 5, which is not currently being sampled after Washington suspended access to it.
The model is a 744-billion-parameter Mixture-of-Experts system (roughly 40 billion parameters active per token) with a 1-million-token context window, built for long, messy coding-agent runs. Zhipu did not publish official benchmark scores at launch, but independent testers report it edging out OpenAI's GPT-5.5 on several long-horizon coding tasks (for example 62.1 versus 58.6 on SWE-bench Pro, a real-world software-engineering test) at roughly one-sixth the cost: about $1.40 per million input tokens and $4.40 per million output, under a permissive MIT license that lets anyone download and run it.
The timing makes it land harder. The most capable American models are getting harder to access, not easier, while a model anyone can grab for free is matching them on the work developers care about most.
Why it matters: When a free, open model trades blows with the best closed systems at one-sixth the price, the question for most companies stops being which lab is smartest and becomes how much they are willing to overpay for the last few percent.
Sources:
๐ VentureBeat coverage
๐ LLM Stats
No theory. No slides. Just pipeline.
Most founders know their product. Few know how to get it in front of the right people. In this hands-on session, Clay + HubSpot for Startups walk you through ICP definition, prospect list enrichment, and AI-personalized outreach. You launch your first sequence before the session ends. June 18. 11am ET / 4pm GMT.
The chart: Arena's Code Arena Frontend leaderboard ranks models on real frontend-coding tasks by human preference, scored Elo-style. GLM-5.2 (Max) lands at #2 with 1,595, ahead of every Claude Opus variant, Qwen-3.7 Max, and Gemini-3.5 Flash, and behind only Claude Fable 5 (1,654), which is flagged as not currently being sampled. Its own predecessor, GLM-5.1, sits back at #9 (1,531).
The lesson: An open-weight model is now essentially tied with the best closed systems at frontend coding. The distance from #2 to the top of the actively-sampled field is a few dozen Elo points, not a generational gap.
The caveat: This is one narrow benchmark, frontend coding decided by human-preference votes, and the #1 model is offline, so the leaderboard's peak is provisional. A single Elo chart is a snapshot, not proof of all-round superiority.
๐ง What Comes After AGI? DeepMind Maps the Next Step
โก Bottom line: A new Google DeepMind paper sketches how AI might move from human-level AGI to a superintelligence that outthinks entire expert organisations.
๐ก Why it matters: It is one of the clearest attempts yet to define ASI and the concrete routes there, written by researchers building frontier systems.
๐ What it means: Superintelligence may arrive as a series of societal shifts, not a single overnight event, which changes how we should prepare.
Most AI debates stop at AGI, loosely, a system as broadly capable as a competent human. A new paper from Google DeepMind, titled From AGI to ASI and posted this month, asks what lies beyond it. Its authors, including DeepMind co-founder Shane Legg and AI theorist Marcus Hutter, define artificial superintelligence (ASI) as a system "more intelligent and cognitively capable than large organisations of humans", not merely better than one person at one task.
Shane Legg, Google DeepMind co-founder and an author of the paper (Shane Legg / TED)
The authors map four routes from here to ASI: scaling compute, data, and models further; an algorithmic paradigm shift beyond today's transformers; recursive self-improvement, where AI starts accelerating its own research; and group agency, where many AGI-level agents combine into something smarter than any one of them.
Just as important is what could stop it. The paper names six bottlenecks: a data wall as high-quality training text runs out; compute, energy, and capital demands rising faster than supply; the risk that the current neural paradigm is simply insufficient; research getting harder as the easy gains are spent; an abstraction barrier past which more intelligence stops helping; and a deliberate slowdown if society chooses caution. Progress, they argue, could stall as easily as it could explode.
(Google DeepMind)
Its most grounding point is that superintelligence may not announce itself. Rather than one dramatic switch-on, the authors expect "a series of transformative societal changes," and they frame preparing for it as "a massively interdisciplinary endeavour of global scope." Or, as the paper opens, quoting Turing: "We can only see a short distance ahead, but we can see plenty there that needs to be done."
โWho is this person again?โ
Youโve had that moment. Walking into a call, scrolling through old emails, trying to remember what you promised. Lindy texts you a brief 15 minutes before: attendee context, past discussions, open items, talking points. All pulled automatically. Try Lindy free.