Frontier AI model prices are collapsing, and the timing is brutal for Anthropic and OpenAI. In eight days, Meta, SpaceX, and China's Moonshot each released models that directly undercut the U.S. labs on cost. Meta's Muse Spark 1.1 matches Anthropic's Opus 4.8 on select benchmarks at a fraction of the price. SpaceX's Grok 4.5 is less than half the price of Opus 4.8 and competitive on coding benchmarks. Moonshot's Kimi K3, a 2.8-trillion-parameter model, is going open-weight, which could destroy pricing power entirely.
The structural problem is market shape, not just individual competitors. Atreides Management's Gavin Baker argued publicly that a world with only 2 to 3 dominant frontier labs running 90% inference margins is bad for every downstream AI layer: semiconductors, hyperscalers, software. The labs built their entire capital strategy on owning the frontier. That moat is eroding. Instead of 2 to 3 competitors, the market is trending toward 6 or more frontier developers, which compresses margins across the board.
The original piece is worth reading for the pivot it tracks inside Anthropic itself. Boris Cherny, head of Claude Code, confirmed in a May interview that products now constitute a much larger share of Anthropic's revenue than API access did a year ago. That shift is the real story: the labs must now win on product, competing with every software company on Earth, while simultaneously burning cash on model development and watching inference margins shrink. The full article maps exactly where that leaves OpenAI and Anthropic as their IPO windows open.
[READ ORIGINAL →]