The day in one read

The GPT-6 Astra Release and the AI Capability Race

OpenAI’s release of GPT-6 Astra has intensified the competitive dynamic in the large language model market, marking a shift from efficiency gains to expanded capability sets. The model, built on a new pre-training base, achieved 57.6% on Terminal Bench 4.0 and 41.1% on Automation Bench, outperforming Anthropic’s Fable 5.1, which scored 55.8% and 31.4% respectively. While Astra initially trailed Fable 5.1 on the Artificial Analysis Intelligence Index, the latter’s release of index version 4.2, which emphasizes agentic tasks, suggests a recalibration of how these capabilities are measured. Greg Brockman, OpenAI’s co-founder, stated that he believes Astra represents Artificial General Intelligence, a claim that contrasts with the more cautious assessments of its limitations in front-end UI design and Python code quality. The rollout began with limited access for cybersecurity partners before broad availability, and the announcement video, viewed over 132 million times, highlighted ambient voice interaction and hands-free computer use. This release coincides with Anthropic’s delayed IPO, with secondary market perpetual contracts valuing the company at approximately 2 trillion dollars, a figure that may reflect strategic timing to coincide with the launch of its own Claude 6 model.

AI Infrastructure, Funding, and Credit Risk

The capital intensity of the AI sector is reshaping corporate finance and credit markets. Mistral raised 3 billion euros in a round led by Samsung Electronics, reaching a valuation of over 21 billion euros, a significant step for the European competitor aiming to close the gap with US rivals. This funding is part of a broader trend where AI companies are accumulating massive debt obligations; OpenAI’s compute commitments are estimated at 1.3 trillion dollars, while Anthropic’s are approximately 500 billion dollars. These long-term contracts now dominate the corporate bond market, with datacenter-related debt comprising a quarter to a third of new issuances. Bankers are reportedly lobbying rating agencies to grant OpenAI and Anthropic investment-grade ratings post-IPO to lower borrowing costs, a move critics argue is irresponsible given the companies’ lack of profitability and parallels to the 2008 financial crisis. Meanwhile, Qualcomm gained 4 percent after announcing a multi-year AI partnership with Amazon for custom chips, with Amazon holding options for 25 million shares tied to 60 billion dollars in orders. ASML secured orders for its new-generation high-end lithography machines from TSMC and Samsung, reversing TSMC's earlier stance that the equipment was too expensive, with each machine costing approximately 400 million dollars.

Apple’s Hardware Pivot and AI Integration

Apple is preparing to unveil its first foldable smartphone, the iPhone Ultra, at its September 9th event, marking a market entry more than five years after Samsung’s initial foldable device. The launch coincides with Tim Cook handing the CEO role to former hardware engineering vice president John Turnus. The device is expected to cost over $2,000 and features a 5.5-inch cover display expanding to 7.8 inches when unfolded. To maintain a sleek profile, the phone omits a telephoto camera and replaces Face ID with a side-mounted Touch ID button. Apple is also moving forward with a Gemini-powered Siri AI update and a HomePod refresh. Notably, researcher Aaron Paris discovered evidence in a macOS preview of AirPods with built-in cameras, which are expected to act as eyes for Siri to capture low-resolution visual information rather than for photography. This move aims to leverage Apple’s privacy-focused brand against competitors like Meta, which has launched its private AI agent Muse based on the Muse Spark 1.3 model. Meta also stopped using AI token consumption as a performance metric for engineers, citing "token maxing" behavior, a shift that reflects the changing nature of AI-driven productivity.

Enterprise AI Evaluation and Cost Management

The gap between public benchmarks and actual model capabilities is creating a demand for independent evaluation frameworks. VALS, founded in 2024 by Rayem Krishnan, provides independent AI model evaluations to address this gap. Krishnan noted that Meta’s Lama 4 underperformed on VALS’s private held-out benchmarks despite showing strong results on public, open-source benchmarks. For enterprises, Krishnan argued that AI token spend is becoming an existential cost line item that may eclipse salary spend. He cited a Fortune 10 company that initially budgeted $100 per day per engineer for Cloud Code, later increasing it to $300, noting that arbitrary limits create productivity dead zones when rate limits reset. VALS itself spent approximately $1.5 million in free tokens during a one-month experiment, with one engineer using six billion tokens in a single day. This spending was 10x the company’s employee salary costs for that month. The experiment revealed that Cognition’s Devon tool was more token-efficient than other options, informing VALS’s strategy to use subscription-based pricing where possible. Krishnan also emphasized the need for a shared language of evaluations to facilitate international verification of AI capabilities, drawing parallels to nuclear non-proliferation treaties.

Corporate Restructuring and Market Volatility

Volkswagen’s 147-page restructuring plan, leaked by journalists, confirms the company faces existential threats. The plan proposes cutting an additional 50,000 jobs on top of the 50,000 already agreed for 2030, with 15,000 at Audi and Porsche. Plants in Emden, Zwickau, Hannover, and Neckarsulm are flagged for closure due to lack of competitive follow-on production by 2031-2034. Production targets drop to 9 million vehicles annually, down from pre-pandemic levels. The return on capital target rises from 3.8% to 9% by 2030. Potential divestitures include Ducati, with a reported 2.5 billion Euro offer, and the Seat brand. The stock trades at a price-to-book ratio of 0.22, valued at 40 billion Euros, significantly lower than BMW and Mercedes-Benz. In the pharmaceutical sector, Novartis shares plunged 11 percent after its drug Pelacarsin failed to show statistically significant reductions in heart attacks or strokes in a key study. Helena Smolak, a pharma correspondent for Handelsblatt, argued that Novartis faces a significant M&A problem and that its CEO is under pressure. In the insurance sector, Allianz reported a 15.5% increase in core net income and a 20.7% return on equity in the first half, with its solvency ratio rising to 225%.

Also Notable

Brent crude approached 99 dollars following Houthi rocket attacks on Saudi oil facilities, up from 70 dollars in late June, fueling inflation expectations and higher interest rates. UK 10-year bond yields hit their highest since 1998, while German 10-year Bunds yielded 3.37 percent and US Treasuries 4.8 percent. A California security firm, Califf, disclosed a zero-click worm in WeChat that could have compromised over one billion accounts by exploiting a vulnerability found autonomously by AI. The exploit, developed in two days, allowed infected accounts to call contacts and hijack their devices on both iOS and Android. Tencent patched the flaw after initially banning the researchers' accounts. Genius Sports, the exclusive NFL real-time data provider through 2029, reported Q2 revenue of 196 million dollars and adjusted EBITDA of 53 million dollars. The company guides for 2026 revenue near 1 billion dollars and EBITDA of 285 to 295 million dollars. Analysts note that NFL data for prediction markets like Kalshi and Polymarket is not yet included in forecasts, though the NFL currently prohibits such sales. Germany’s draft crypto tax law proposes taxing gains as capital gains regardless of holding period, but includes grandfathering for assets acquired before December 31, 2026. The ministry projects 160 million euros in additional revenue by 2028, rising to 350 million by 2031.