Signals

Artificial Analysis counted four frontier launches in eight days, with six labs now fielding a model above 50 on its Intelligence Index — a frontier that went from two labs to six in six weeks. The more consequential number is price: near-frontier intelligence became 2-3x cheaper over the same eight days. GPT-5.6 Sol lands a point below Claude Fable 5 at $1.04 per task versus $2.75; Grok 4.5 delivers 54 at $0.31, under a third of GPT-5.5's price; fresh 51-point models undercut last week's cheapest option within days. The leaderboard shuffle is the distraction. The signal is the rate of commoditisation. If the thesis holds, this is the onset of a Jevons dynamic — cheaper near-frontier intelligence expanding total demand faster than it compresses per-token revenue. A parallel note on frontier-lab economics puts the survival test bluntly: for any lab that doesn't own data centres or power, the only thing that matters is model demand, and models must be either the best or cheap and fast. Both ends of the smiling curve tighten at once: the model layer commoditises in public while value migrates to infrastructure below and applications above. For anyone repricing model-layer businesses on last quarter's assumptions, the market is moving faster than the spreadsheet.

Two data points this week point at energy — not silicon or capital — as the scarce input. Anthropic listed a Data Center Energy Lead for Australia, tasked with securing multi-hundred-megawatt power capacity at the speed and scale frontier AI development demands: a lab moving to lock up power directly. And in the US, New Mexico denied a gas-pipeline permit for an Oracle data centre, a concrete instance of the permitting friction now shaping where compute can physically land. The pairing sharpens a question worth taking seriously for the Australian buildout: how a host economy captures the value of siting frontier compute domestically rather than exporting the returns, at a moment when local coverage frames data centres mainly as money flowing offshore. The analytical point stands independent of any single project. As capable inference proliferates onto edge hardware and hyperscale training concentrates, secured power capacity becomes the input that decides infrastructure winners. Compute is increasingly a function of who can get megawatts approved and connected — and that is a slower, more local, more political variable than chip supply.

Cursor frames each jump in AI capability as raising the level of abstraction at which an engineer works, and offers a sharp analogy for multi-agent orchestration: a swarm resembles a compiler. A compiler lowers source code to machine code through intermediate steps; a swarm lowers intent — planners parse a goal into task trees, then step by step into executable work. The load-bearing difference is the whole point: a compiler preserves meaning at every step, while the swarm is probabilistic at every one. As Cursor puts it, everything in the system exists to close that gap. That reframes harness engineering precisely — it is the discipline of managing probabilistic lowering, the orchestration, verification and error-correction machinery whose job is to recover compiler-like reliability from stochastic parts. It is a clean statement of why value in agentic systems migrates into the harness rather than the model. The model is the transistor; the reliability comes from everything built around it. For anyone assessing where defensibility accrues in AI-native software, the compiler-that-lost-its-guarantees framing is a better guide than raw model capability.