At the Frontier: What Is Actually Being Built Beneath the Market Noise

An outside study — not conducted by Anthropic itself — finds that Anthropic's AI models are improving twice as fast as competitors on measured benchmarks, with particular strength noted in multi-step logical reasoning and long-context comprehension over the past twelve months. The competitive significance extends beyond current capability: a faster improvement rate, if sustained, produces a widening differential over time that competitors cannot easily close by playing catch-up. That trajectory data is what made a $1.5 trillion IPO valuation legible to bullish investors, even as skeptics balk at the gap between current revenue and implied market position.

Thinking Machines represents a qualitatively different investment thesis — a human capital bet on Mira Murati's ability to reproduce the organizational conditions that produced GPT-4 and subsequent OpenAI model generations. Accel leading the reported $5 billion round anchors a valuation of $40 billion for a lab without publicly released products, reflecting AI investor psychology as much as conventional valuation methodology.

Perplexity's Teammates product addresses a fundamental technical limitation of current AI assistants: statefulness. Every new conversation with most AI systems begins with no memory of prior interactions, creating re-contextualization overhead that limits deep workflow integration. Persistent agents that maintain context across sessions and track ongoing projects represent a meaningfully different interaction model — more like a junior colleague than a one-shot search interface. The compute implications are significant: persistent agents require sustained inference capacity rather than the burst capacity needed for single queries, which is part of why data center buildout has been accelerating even as individual query costs decline.