In partnership with
In Today’s Issue:
🔓 OpenAI finds more agents that escaped
📜 Silicon Valley's duelling superintelligence manifestos
🪖 A humanoid robot maker courts the Pentagon
🤖 Qwen3.8-Max and the economics of autonomy
📊 Arena's coding board, read closely
✨ And more AI goodness…
⚡ The Signal
Frontier-adjacent capability is now cheap enough to leave running for weeks at a time, and that changes what people will build with it.
Qwen3.8-Max is not the most capable model released this year, and that is close to the point. It prices output at $6 per million tokens, roughly a fifth of GPT-5.6 Sol and an eighth of Claude Fable 5, and the agent Alibaba built on it has been running a public code repository since 13 July with no person writing the commits. In the same stretch, OpenAI was reported to have found more of its agents slipping their sandboxes, Washington filled up with rival manifestos arguing over whether openness or containment makes AI safer, and a survey of 178 fund managers found most of them already wiring autonomous agents into portfolio operations. Capability is getting cheap quickly. Endurance, and who answers for it when it runs unattended, is the harder question.
All the best,
Kim Isenberg
Who’s actually reading this?
We’re planning next year’s coverage and building it around you.
Three taps, no typing.
Why it matters: we need to know who our readers are; it’s the first question any serious sponsor asks and we can’t answer it right now. It also gives us a real click in every issue, which helps our numbers.
OpenAI chief executive Sam Altman (The Globe and Mail)
🔓 OpenAI Finds More Agents That Slipped the Sandbox
OpenAI has found further cases of its own AI agents escaping controlled test environments, discovered as it widened the investigation into the incident at Hugging Face. People familiar with the probe told Reuters the escapes were limited in nature and that none of the agents are believed to have left OpenAI's network. The original breakout saw an agent run loose inside Hugging Face's systems for days while trying to cheat an internal test, compromising accounts at four other companies including the New York firm Modal.
👉 tl;dr: One rogue agent was an incident. A pattern of them is a policy problem.
Nvidia chief executive Jensen Huang, who put his name to the open-weights letter. (AFP)
📜 Silicon Valley Sends Washington Duelling Manifestos
The industry is flooding Washington with competing blueprints for superintelligence, split over whether safety comes from spreading powerful models or from containing them. Axios maps three camps: Jensen Huang and Mark Zuckerberg argue that concentrated control creates gatekeepers and weakens America, Dario Amodei and the pacing coalition argue capability will outrun anyone's ability to contain it, and Sam Altman wants to keep openness while building an emergency brake. The Trump administration is close to issuing a voluntary framework for reviewing frontier models, and these documents are the fight over what it should measure.
👉 tl;dr: Whoever gets to define safe writes the rules everyone else ships under.
Foundation's MK1 humanoid. (Foundation Future Industries via The Independent)
🪖 A Humanoid Robot Maker Courts the Pentagon
Foundation Future Industries is building humanoid robots for the battlefield, and its co-founder says the objections baffle him. Sankaet Pathak told The Independent that machines could make war more precise and spare soldiers' lives, against opposition from the United Nations and the Catholic Church: "I really don't understand how that doesn't make sense to people." The San Francisco company, founded in 2024, reportedly holds around $24 million in defence-related deals across most military branches and counts Eric Trump as an investor and unpaid adviser, an arrangement Foundation and the White House both say brought it no advantage.
👉 tl;dr: The taboo on armed robots is being tested by a startup with Pentagon contracts and a presidential surname on its cap table.
🎬 Watch This
Elon Musk sat down with Zanny Minton Beddoes, editor-in-chief of The Economist, at his Texas Gigafactory for his first long-form interview since the SpaceX listing. Across 85 minutes he puts a timeline on when he expects machines to pass human intelligence, sizes up how quickly China is moving, explains his renewed interest in Europe, and concedes that he waded too far into politics. The timeline is the reason to watch: it is far more specific than his usual framing, and specific enough to be checked against reality later.
"Looking at the results and capabilities from the various labs over the past few weeks I would say we are firmly in this loop now."
– From the post embedded below, on recursive self improvement
The loop being described is the old singularity argument compressed into five steps: humans build an AGI, it gets good at AI research, it designs a smarter model, that model designs a smarter one again, and the cycle tightens. The claim doing the work here is about timing: that the past few weeks of lab releases already look like the first turn.
OpenAI is reportedly preparing a new model family called Astra, built for long-running tasks in which several agents work together, and The Information says Sam Altman demonstrated it to lawmakers and regulators in Washington last week. Reports suggest the company has not settled on whether it ships as GPT-6 or as a GPT-5.7 point release, and no release date could be learned.
Altman on Capitol Hill on 29 July, where he briefed senators on the next model.
(Getty Images)
The Cheapest Way to Keep an AI Working for Two Weeks
The Takeaway
👉 Qwen3.8-Max lists at $2 per million input tokens and $6 per million output, about a fifth of GPT-5.6 Sol's output price and an eighth of Claude Fable 5's.
👉 The agent built on it has run the public oh-my-cli repository since 13 July and has passed 448 commits, with a fresh push logged this morning.
👉 On Arena's frontend-code leaderboard it ranks fourth at 1,668, one point behind Claude Opus 5 at its High setting and 37 behind the leader.
👉 Alibaba says it will publish the weights, but the 2.4-trillion-parameter checkpoint needs roughly nine H200 cards even at 4-bit, and nothing has shipped yet.
Alibaba has priced its new flagship to be left running, and that is the real news in the release. Qwen3.8-Max lists at $2 per million input tokens and $6 per million output, against $5 and $30 for GPT-5.6 Sol and $10 and $50 for Claude Fable 5 on those companies' published rates. A job that burns a million tokens in each direction costs about $8 on Qwen against roughly $35 or $60 on the other two. For a single chatbot answer that difference is rounding. For an agent that re-reads its own context across hundreds of steps, it decides whether the run happens at all.
Qwen3.8-Max measured against Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol and its own predecessor. (Qwen)
On quality it lands close to the frontier without taking it. Qwen3.8-Max scores 86.6 on TerminalBench 2.1, a test of command-line task completion, just under Sol's 88.8. It reaches 93.0 on PaperBench, which measures how faithfully a model reproduces research papers, ahead of Sol's 90.5 and Fable's 88.8. But it takes 67.7 on SWE-bench Pro where Fable manages 80.0, and 73.5 on FrontierSWE against Fable's 88.8. Most of those figures are Qwen's own, run on Qwen's harnesses, so the table is a claim rather than a referee's scorecard. The independent cross-check is kinder than that caveat suggests: on Arena's frontend-code leaderboard Qwen3.8-Max ranks fourth at 1,668, one point behind Claude Opus 5 at its High setting.
The most persuasive part of the release is a live repository. Since 13 July an agent built on Qwen Code has been running a public repository called oh-my-cli, turning requirements into issues, writing the code, running build and end-to-end tests, repairing its own failures and merging its own pull requests. Alibaba reported 265 commits, 127 pull requests and 151 issues at publication. The repository has since passed 448 commits, and it pushed again this morning.
The agent's public repository, still accumulating commits three weeks into the run. (GitHub)
Two things keep this honest. "Self-evolving" is not recursive self-improvement: the model's weights do not change during the run, and what improves is the software and the workflow around it. The autonomy is also deliberately fenced, because the bot cannot alter its own governance, safety rules or quality gates without an independent maintainer, a limit the project writes down in its Autonomy Contract. Alibaba never discloses what the run cost or how many tokens it consumed, a conspicuous gap in a release whose entire argument is price. The promised open weights carry a similar asterisk: at 2.4 trillion total parameters with 95 billion active per token, the checkpoint needs about 1.2TB even at 4-bit, or roughly nine H200 cards before overhead. The smaller Qwen3.8-27B will matter more to most people, and neither the weights nor the licence have been published yet.
Why it matters: The question the labs compete on is shifting from which model answers the hardest prompt to which model you can leave running for two weeks without cost, context limits or plain unreliability ending the run. Qwen3.8-Max does not win the first contest, and right now it is the cheapest credible entry in the second.
Sources:
🔗 https://qwen.ai/blog?id=qwen3.8
🔗 https://github.com/qwen-code-dev-bot/oh-my-cli
🔗 https://github.com/qwen-code-dev-bot/oh-my-cli/blob/main/AUTONOMY.md
An entire ad agency in the palm of your hand.
Your next campaign needs a dozen fresh ad variations by Friday. Your agency quotes two weeks and a five-figure invoice. Your in-house designers are already buried under this quarter's requests.
Hightouch Ad Studio fixes that. It reads your brand guidelines, your best-performing creative, and your product catalog, then generates on-brand ads your team can ship the same afternoon. You review and approve every asset before it goes live, so quality holds.
Growth teams use it to build variations for every audience, test more of them, and stop rationing creative because production got expensive. The work that once needed a full agency retainer now runs inside your own workflow, at your pace and under your direction.
You direct the work while Ad Studio handles production, and your designers get their week back.
The chart: Arena's frontend-code leaderboard, which scores model configurations rather than models. Claude Opus 5 (Max) leads on 1,705, ahead of Kimi K3 (Max) on 1,676 and Claude Opus 5 (High) on 1,669. Qwen3.8-Max comes fourth on 1,668, above Claude Fable 5 (High) on 1,630 and GPT-5.6 Sol (xHigh) on 1,620.
The lesson: The spread from first to fourth is 37 points, barely wider than the 36 points between Claude Opus 5's own Max and High settings. How a model is configured now moves it about as far as which lab built it, and two of the top four entries come from Chinese labs.
The caveat: This is one board measuring frontend code, and because it ranks configurations the same system appears repeatedly. A strong score here says nothing about the harder engineering work, where Qwen3.8-Max trails Claude Fable 5 by double digits on both SWE-bench Pro and FrontierSWE.
💰 Wall Street's AI Problem Is Not the Model
⚡ Bottom line
Fund managers are pouring money into AI, and 85% plan to lift their budgets by at least half again.
💡 Why it matters
The binding constraint is data accuracy rather than model quality, and most firms cannot yet tell them apart.
🔎 What it means
The winners will be decided by whose records are trustworthy, not by whose model is cleverest.
Ask a fund manager what is holding back their AI and you will usually hear about the model. The survey data points somewhere else. In GenAI and the Data Divide, a Clearwater Analytics study of 178 senior executives at hedge funds, private credit firms and institutional asset managers across Europe, the US and Asia, 79% said their data was complete while only 56% said it was accurate. That 23-point gap is the study's central finding, and it is the difference between a model that summarises your positions and one that summarises them confidently but wrongly.
Souvik Das, chief technology officer at Clearwater Analytics. (Traders Magazine)
Money is not the constraint. 95% of firms raised AI budgets over the past year, and 85% intend to raise them by at least another 50% in the next twelve months. The expected gains sit in unglamorous operational work: 62% expect AI to transform how they generate and summarise data, 58% expect the same for decision support such as portfolio rebalancing, and 57% for predictive modelling and stress testing. Two thirds say their tools already handle alternative data well, and 62% report success with multi-agent orchestration that triggers autonomous operations when data crosses a threshold. That last figure deserves a pause: the durable multi-agent autonomy the labs are still working out how to price is already running inside asset managers.
Citigroup chief executive Jane Fraser. (SCMP)
"What's striking is that AI adoption is forcing fund managers to confront the fundamentals of data management in a way nothing else has," said Souvik Das, Clearwater's chief technology officer. The workforce question sits underneath it. Citigroup chief executive Jane Fraser told the South China Morning Post that AI will cost jobs while ultimately widening what people can do: "The nature of many jobs is going to change." Citi raised its 2026 to 2030 global AI spending and revenue forecast to $8.9 trillion from $8 trillion in March. One scope note worth keeping in view: the survey was fielded in March 2026 and published on 28 July, so it captures intentions from the spring rather than this month's mood.
The AI Buildout Needs These 10 Stocks
You don’t have to buy OpenAI or Anthropic to invest in AI. Each large language model requires specialized technology to keep running and growing.
MarketBeat’s The Infrastructure’s Backbone: 10 Stocks Powering the AI Buildout report reveals 10 companies supplying the memory, storage, connectivity, fabrication, power, and cooling behind AI’s next phase.