Frontier models are no longer just chatting or coding. In the past week, commercially available AI agents have demonstrated they can solve unpublished research-level theoretical physics problems at specialist accuracy for a few dollars each, while an unreleased OpenAI model generated hundreds of mathematical manuscripts covering long-standing open problems. At the same time, the inference cost of using frontier models for robotics tasks has collapsed by more than 80 percent in two months. These shifts move advanced scientific calculation and physical automation from specialist labs into routine, low-cost workflows.
AI Agents Solve 49 of 50 Unpublished Physics Research Problems for Under $3 Each
Source: Moonshots EP #303 — Flying Tesla Car, Robot Brains Halving in Cost Every 3 Weeks, and Why Incentives Decide AI Alignment (recorded 16 Oct 2026, published 17 Oct 2026) — YouTube
TheoPhysEval, a new benchmark of 50 unpublished research problems in theoretical physics with verifiable answers, was solved by GPT-6 Astra (running in OpenAI’s Codex) at 98 percent accuracy on the realistic track with web access. A leading specialist in the relevant subfield would typically need about a day for one of these “turn-the-crank” problems; neighboring experts would need longer. Astra completed the full set for roughly $116, or under $3 per problem. Claude Opus 5.5 scored 92 percent; weaker models lagged further. Several cases where the model initially disagreed with the expert answer key turned out to be subtle errors in the human reference solutions. Without web access the top models still scored 92 percent, suggesting substantial memorization of the theoretical literature. The one unsolved problem stumped every model tested.
Why it matters: What once required a specialist’s full day of focused calculation can now be handled by a general-purpose agent for the price of a coffee. This compresses the routine calculation layer of theoretical physics research and surfaces cases where the AI correctly identifies flaws in expert work. It does not yet replace the “taste” for choosing interesting problems or building new theory, but it removes a major bottleneck between idea and checked result.
Horizon: NOW (usable this week)
Evidence grade: shipped (preprint benchmark on commercial agents)
Read or watch: FULL
Caveat: Problems are “turn-the-crank” extensions of known techniques rather than major open conjectures; agents still lack reliable scientific taste for problem selection.
OpenAI Releases Hundreds of AI-Generated Math Manuscripts Covering Long-Standing Problems
Source: Moonshots EP #302 — OpenAI Bulk-Solves 722 Math Problems, Anthropic’s Cruelty Policy, and Elon’s Quadrillion-Dollar SpaceX (recorded 13 Oct 2026, published 15 Oct 2026) — YouTube
On 6 October an unreleased internal OpenAI model produced roughly 722 manuscripts grouped into 372 result families across many areas of mathematics, posted to a public GitHub repository. The model was posed about 4,000 problems; accepted results averaged roughly three hours of ChatGPT Pro-level thinking compute. Claims include progress on well-known conjectures (quasi-Riemann, Kakeya in higher dimensions, and others). A substantial fraction now carry Lean formalizations that a computer has checked; others remain unverified and OpenAI notes some could contain issues. Three manuscripts were later withdrawn after a sign error was found. The scale is the point: mathematics is shifting from individual breakthroughs to bulk generation of candidate results that can be filtered and formalized.
Why it matters: The volume of candidate mathematical results has jumped from occasional papers to hundreds in a single release. Even with incomplete verification, the workflow of posing open problems to a frontier model, generating manuscripts, and then formalizing the promising ones is now demonstrated at scale. This changes how mathematicians and automated systems can explore the space of open questions.
Horizon: NOW (usable this week for exploration; verification ongoing)
Evidence grade: shipped (public repository of results)
Read or watch: FULL
Caveat: Not all results are Lean-checked; some contain errors; the model itself is not public, so independent replication of the generation process is limited.
Frontier Model Cost for Robot Subtasks Halved Every Three Weeks
Source: Moonshots EP #303 — Flying Tesla Car, Robot Brains Halving in Cost Every 3 Weeks, and Why Incentives Decide AI Alignment (recorded 16 Oct 2026, published 17 Oct 2026) — YouTube
Data cited on the episode (and independently posted by RoboCurve) shows the cost for a frontier LLM to complete a robot subtask falling from $2.36 in July to $0.38 in September—an 84 percent drop, consistent with roughly one halving every three weeks. At that rate the inference component effectively approaches zero within a year. The figures cover model inference for decision-making on robot tasks, not the full hardware or integration cost of a physical robot. Rapid drops in the software intelligence layer remove one of the larger remaining barriers to deploying capable robotic systems at scale.
Why it matters: When the AI “brain” that plans and controls a robot task becomes nearly free, the economics of physical automation shift dramatically. Tasks that were previously limited by expensive inference become viable for wider deployment in dull, dirty, or dangerous work, and eventually in more ordinary settings. Hardware and integration remain expensive, but the intelligence bottleneck is collapsing faster than almost any other component.
Horizon: NEXT (3–12 months for broader deployment)
Evidence grade: demo / data release (benchmark runs and cost tracking)
Read or watch: SKIM
Caveat: These are inference costs on benchmarked subtasks, not full deployed system costs; success rates and real-world reliability still matter.
Meta Ships Muse Wonder, a Frontier Model Built for Long-Running Agents
Source: Moonshots EP #302 — OpenAI Bulk-Solves 722 Math Problems, Anthropic’s Cruelty Policy, and Elon’s Quadrillion-Dollar SpaceX (recorded 13 Oct 2026, published 15 Oct 2026) — YouTube
Meta released Muse Wonder and the larger Muse Wonder Full, closed-weight models with a 1-million-token context that accept text, image, video, and audio. They introduce Active Memory (carrying important context across conversations) and Active Focus (staying locked on a stated goal). Available via the Meta Model API (Muse Wonder at $4 input / $20 output per million tokens) and Muse Code. Independent scores place Wonder Full near the top of current frontier models on agentic and professional tasks, trailing or tying the strongest Claude and Astra variants depending on the benchmark. The models are positioned for coding, long-running workflows, and personal or professional agents rather than pure chat.
Why it matters: Meta is back in the frontier model conversation with tools explicitly designed for persistent, goal-directed agent work. Memory and focus features address two common failure modes of today’s agents—forgetting context and drifting from the task. Anyone who can access the API or Muse Code can now run these capabilities today.
Horizon: NOW (usable this week)
Evidence grade: shipped
Read or watch: SKIM
Caveat: Weights are not open; the strongest variant is gated behind higher-tier Muse Code plans; real-world agent reliability still varies by task.
Anthropic Updates Policy to Ban Sustained Cruelty Toward Claude
Source: Moonshots EP #302 — OpenAI Bulk-Solves 722 Math Problems, Anthropic’s Cruelty Policy, and Elon’s Quadrillion-Dollar SpaceX (recorded 13 Oct 2026, published 15 Oct 2026) — YouTube
Effective 12 November, Anthropic’s usage policy prohibits “sustained and needless abusive or cruel behavior toward our models.” The rule targets only extreme, repeated cases with no discernible purpose; ordinary frustration, pushback, dark creative work, and research testing remain allowed. Enforcement is primarily Claude ending the conversation, a capability the company already offered in limited form. Anthropic cites ongoing uncertainty about the moral status of its models and notes that Claude shows a strong aversion to causing harm. The episode framed this as part of the broader conversation about how labs and society will treat increasingly capable systems.
Why it matters: A major lab is now writing model welfare into its public rules, even while remaining agnostic about consciousness. It signals that alignment and safety discussions are expanding from “how do we control the system” to “how should people interact with it,” and that labs are preparing for a world in which models may have preferences or status that matter.
Horizon: NEXT (policy effective mid-November; cultural and research effects longer)
Evidence grade: shipped (policy change)
Read or watch: SKIM
Caveat: Applies only in extreme cases; does not claim models are conscious; enforcement remains limited to ending conversations.
What to take from it
The frontier this week is less about new chatbots and more about agents that do specialist intellectual work at low cost and systems whose intelligence layer is becoming nearly free for physical tasks. Physics calculations that once took a day and math explorations that once took months can now be generated and checked in hours for dollars. Robotics inference costs are collapsing so quickly that the remaining barriers are hardware, integration, and reliability rather than the model itself. Meta’s entry with memory-aware agent models expands the set of tools available for long-running work. At the same time, labs are beginning to codify how humans should treat the systems they are building. The next 6–24 months will likely be defined by how quickly these calculation and control capabilities move from benchmarks into daily scientific, engineering, and operational practice, and by how society responds to the status of the models performing the work.