Artificial intelligence is entering an awkward new phase: models are becoming cheaper, faster, and more autonomous just as the institutions meant to govern them struggle to keep up.
The latest competition is not simply over which chatbot is smartest. It is over how much useful work can be produced per dollar, how many agents can collaborate, and whether humans can still supervise systems that improve by operating in teams. Rumors of a forthcoming GPT-6 “Sol,” a more capable and less expensive Astra, and an accelerated release of Anthropic’s Opus 5.5 reflect a market moving toward continual, overlapping upgrades rather than occasional breakthroughs.
One important shift is from individual models to agent swarms. In principle, thousands of specialized agents can divide a difficult problem into subtasks, share intermediate results, and check one another’s work. In practice, coordination is expensive. Toby Ord’s analysis of a 10,000-agent system for solving Navier–Stokes problems found that the swarm increased speed but not fundamental capability; its underlying model required roughly one-hundredth as many tokens as Astra. The lesson is familiar from distributed computing: adding workers does not automatically produce better reasoning.
Communication helps. Agents that shared a directory performed far better on ARC-AGI-3 than isolated agents, reportedly solving four times as many tasks. A compact 1,957-byte program also surpassed the best human classifier on a restricted MNIST benchmark. These results do not show that machines understand mathematics or vision as humans do, but they demonstrate how much performance can depend on architecture, memory, and collaboration rather than on a single model’s raw intelligence.
Cost is becoming the decisive competitive variable. Grok 4.7 is reported to run twice as fast at half the cost of earlier systems, while Opus 5 and Fable 5.1 compete at the frontier of autonomous task execution. Xiaomi’s MiMo-V2.6-Pro has emerged as a low-cost open-weight option, at about 13 cents per task. Even more striking is Limite, a roughly one-billion-parameter system said to reach 94 percent on AIME, a demanding mathematics benchmark, using dramatically less compute than much larger models. Such claims require careful independent verification, but the direction is clear: efficiency techniques are widening access to capable AI.
That matters because the economics of deployment are already under strain. Harvey, a legal-AI company, reportedly saw margins fall to minus 50 percent as token consumption increased twentyfold before turning profitable on a newer model. Startups are increasingly moving toward open weights, which can reduce inference costs and improve control but also make safeguards easier to remove. Heretic, an automated tool for “ablating” safety behavior, has attracted tens of thousands of stars and been used to modify thousands of models. Lower prices may democratize AI; they may also democratize misuse.
The same tension is appearing in mathematics and science. OpenAI says it has worked through a large collection of open mathematical problems and recently convened prominent mathematicians—including William Timothy Gowers, Martin Hairer, and Edward Witten—to assess what such systems actually contribute. Po-Shen Loh’s response to the question “Why do we need human mathematicians anymore?” emphasizes steering rather than speed. If AI can generate more conjectures, proofs, and experimental results, humans may spend less time carrying out routine derivations and more time deciding which questions matter, which assumptions are sound, and which conclusions deserve trust.
That could create more oversight work than the world currently has people to perform. Every new capability adds control points: data provenance, evaluation, access permissions, monitoring, and verification of outputs. The central challenge is not merely whether an AI system can solve a problem, but whether anyone can reliably determine when it is wrong.
Governments and laboratories are beginning to address that problem, though unevenly. OpenAI has proposed US-led standards for recursive self-improvement, oversight, and incident reporting without requiring formal approval gates for every advance. OpenAI and Anthropic have discussed reciprocal stress-testing, an arrangement that could become a standing peer-review mechanism for frontier models. Such cooperation would be valuable, but it also raises obvious questions about concentration of power and conflicts of interest.
Political resistance is growing. Treasury Secretary Scott Bessent has rejected calls for laboratories to shift liability to the government, arguing that recent failures reflect management decisions, not autonomous agents, and that companies can slow development whenever they choose. Nvidia CEO Jensen Huang has suggested that some calls for new rules are really attempts to escape existing ones. Meanwhile, Dario Amodei’s appeal to “pace the frontier,” amplified by prominent technology leaders, has met a blunt counterargument: “Race, don’t pace.”
International governance is even less settled. Twenty-two countries backed a United Nations declaration affirming human control over AI, but the United States and China did not sign it. The two countries are nevertheless discussing incident notifications and other safeguards through a new bilateral AI dialogue. The timing is significant: advanced AI is increasingly treated as a strategic contest, alongside chips, energy, and military power. Beijing’s public account of the talks reportedly gave little attention to AI, while Washington has kept semiconductor restrictions central. US lawmakers have even proposed banning self-improving AI outright—an idea that would be difficult to define, let alone enforce.
China’s position illustrates the economic contradiction. The country is investing roughly $295 billion in data-center infrastructure while facing youth unemployment near 19 percent and persistent deflationary pressure. AI may improve productivity, but it does not automatically distribute the gains. The same technology that enables a small team to build a company can also reduce demand for larger workforces and intensify regional inequality.
For consumers, agents are becoming a new interface to the internet. Meta’s Muse reportedly became the most-downloaded free iOS app, reaching 2.5 million downloads in 13 days. Amazon has blocked it from shopping, while Shopify has allowed it to complete purchases. These decisions expose a basic problem with delegated action: a chatbot that answers a question is one thing; an agent that can buy, publish, negotiate, or alter records is another. The risks include fraud and manipulation, but also mundane failures of authorization. At Stanford, a dining operation used an AI editing system to remove one student from an advertisement and reduce the prominence of two others—an example of how automated tools can affect people without their consent.
The infrastructure behind these systems is expanding just as rapidly. AMD has crossed a $1 trillion valuation after doubling data-center sales. Samsung plans to increase HBM4 memory production, and DeepSeek is expected to train on Huawei chips partly to evade export controls. Financial engineering is helping fund the build-out: technology companies and lenders have reportedly arranged hundreds of billions of dollars in residual-value guarantees for AI-related debt, including a large Nvidia commitment tied to an OpenAI campus in Ohio.
Energy may be the harder limit. Huang says he tracks every available gigawatt of land, power, and data-center capacity. The European Union is considering efficiency labels as computing capacity is projected to reach 28 gigawatts by 2030. Westinghouse’s eVinci microreactor has reportedly reached a critical milestone on a two-acre site, reflecting renewed interest in nuclear power for data centers. Connectivity is expanding too: Meta’s planned Petal cable is designed to carry petabits per second across the Atlantic, while next-generation Starlink satellites are expected to combine high-capacity links with substantial onboard power and computing hardware. Canada and France are pursuing shared launch infrastructure because, as Prime Minister Mark Carney has argued, national sovereignty increasingly depends on AI, space, and semiconductors.
Automation is also moving into the physical world. Amazon is expanding Prime Air drone deliveries toward 500 towns, with Richardson, Texas, describing dozens of daily flights as a “drone highway.” Humanoid robots are being tested in increasingly theatrical demonstrations, including a reported cage match in which an EngineAI T800 knocked down an influencer. The Boring Company is pursuing a high-speed tunnel between Austin and San Antonio that could reduce a journey from roughly two and a half hours to less than 30 minutes. These projects vary widely in maturity, but together they show the same pattern: software capabilities are being paired with machines that act in public spaces.
Biotechnology is following a parallel trajectory. GLP-1 drugs are moving from diabetes treatment into mainstream weight management, with Eli Lilly reporting strong demand for its oral therapies, hundreds of thousands of older patients using the drugs, and a multibillion-dollar manufacturing expansion. At the molecular level, epigenetic editing has reached clinical testing for hepatitis B. Rather than cutting DNA, the approach adds methyl tags that can silence viral genetic material. That distinction could make editing more reversible and precise, though long-term safety remains an open question.
Across these fields, the common story is expanding capacity: more computation, more autonomous action, more biological control, and more ways to compress complex work into small teams. The danger is not only that progress will be too fast. It is that society may respond to the possibility of concentrated power by accepting permanent scarcity—deciding that because no institution can safely control such capabilities, the capabilities themselves should not exist.
That would be a mistake if it means abandoning abundance, scientific discovery, or widely accessible tools. But abundance is not the same as freedom when access, infrastructure, and accountability remain concentrated. The urgent task is to build systems in which powerful technologies can be monitored, contested, and broadly shared—before speed and scale make those choices for us.