Anthropic IPO skepticism meets the token cost problem

A research firm, New Constructs, called Anthropic's IPO an "Unprecedented Test of Investor Gullibility." Commenters noted the firm said the same about Uber and was partly wrong, but the core question stuck: how does Anthropic become profitable enough to justify its spend unless it stops training new models or open source stops catching up? One reader suggested spinning out the Opus/Sonnet business for developers and leaving the Mythos and projects hype behind.

The LLM-capability-gap thread attacks from the cost side. For models to be as useful as the believers claim, token counts need to rise by orders of magnitude, which means token prices have to fall just as far. Human review is already falling behind, with people still doing the "archaeology" on an OpenAI incident from months ago. A user in the 'replace Claude with a local model' thread reports Qwen 3.6 27B dense roughly matching Haiku 4.5.

The pattern: the argument has moved from "is AI good?" to "who gets paid for it?" Frontier labs are priced like monopolies while open and local models keep closing the gap.