Summer releases and strategic pivots
Publicly available AI models made qualitative leaps in capabilities this summer. Anthropic launched Claude Fable 5 on June 9, suspended access on June 12 following a US export-control directive citing national security, and restored general access on July 1. In early September, OpenAI released GPT-5.6 Sol and then GPT-6 Astra, both on par with Fable in many capabilities. Other labs’ proprietary and open-weight models generally trail GPT and Claude.
Over the past year, AI labs have focused increasingly on software design and implementation, driven by pressure to monetize. Building and serving frontier models costs hundreds of billions of dollars. Anthropic and OpenAI both confidentially filed for initial public offerings in early June. Coding, among AI’s most lucrative applications, is where they have concentrated their efforts.
AI-driven software development, once met with deep skepticism, is now the norm. Engineers now set goals and constraints for AI agents instead of writing and debugging code by hand. This hits junior roles hardest. Their core tasks (debugging, small patches, routine implementation, and testing) are now done by AI.
Software, a lucrative market for AI, is the first field where AI does much of the work. The tasks it absorbed once trained new engineers; without new on-ramps that teach directing and checking AI, the senior pipeline will thin, an early warning for other white-collar jobs.
Breakthroughs in mathematics and engineering
The most dramatic recent development has been in pure mathematics. This year, AI models began solving open problems that had stymied mathematicians for decades. In September, OpenAI published a proposed solution to the Navier–Stokes Millennium Prize problem, accompanied by a proof formalized in Lean. The Clay Mathematics Institute subsequently said the problem appeared to have been settled, while emphasizing its evaluation process. OpenAI released more than 700 mathematical manuscripts, with results at different stages of verification. The field is being rapidly upended, and how mathematicians adapt may show how other professions will adjust as AI overtakes their work.
This matters for most technical fields, because mathematical modeling underpins all science and technology. Widespread access to fast, reliable AI modeling will likely drive sweeping scientific and technological advances. The biggest immediate impact may again be on software development, which remains largely trial and error. The methods that solve math problems can also prove that code will work or generate it from a description of what it must do. Formal verification can establish specified properties, while specification checks and integration testing remain necessary.
AI’s math breakthroughs show that more education no longer guarantees protection from automation. Research may accelerate, but technical work will shift toward framing problems and judging AI’s answers, and verified code could shrink testing jobs, a key entry point. STEM education must adapt.
Physical-world reasoning, biotech, and cybersecurity
AI also made a major leap in modeling the physical 3D world. OpenAI reports that Astra achieves 95.9 percent geometric overlap on BenchCAD, which tests reconstruction of objects from images by generating CAD code, compared with 83.3 percent for GPT‑5.6 Sol. Online demos showed Astra, guided by a few reference images and simple prompts, recreating landmarks and building working mechanical contraptions.
3D reasoning is also aiding biotechnology. In August, Claude Mythos Preview and Opus 4.8, operating specialist design tools, produced binders for 14 of 15 evaluable targets, with 22–35 percent of designs binding depending on the setup.
In cybersecurity, AI is speeding up both the discovery and repair of software flaws and their exploitation. On October 6, Anthropic reported that its Glasswing partners had used Claude to find at least 129,000 verified vulnerabilities based on partial partner reporting between April and July. Fixing flaws quickly matters more as software spreads into cars, robots, and lab equipment. But the same skills that detect threats can power attacks, and while the most capable models remain with top labs and their partners, open “unaligned” models are advancing.
As AI moves into the physical world, drafting, 3D design, and lab research come into scope, with work shifting toward real-world testing of AI’s designs. In cybersecurity, workers who can triage and quickly fix AI-found flaws grow more valuable as attackers gain similar tools. AI is not only a white-collar issue.
Looking ahead
The final quarter is typically busy for major AI releases, so more-advanced models are likely coming soon. This summer’s advances share one thread: a shift from humans doing the work to humans directing AI and checking its output. Workers with the support and ability to adapt will be best positioned for what comes next.
The most valuable skills will be directing and checking AI, which few training systems teach and vanishing entry-level tasks no longer build. Priorities include new early-career on-ramps and faster data on how AI changes job tasks, as AI moves in months and policy in years.