Independent analyst Martin Alderson argues that open-weight models are crossing a threshold this summer, not because of a new benchmark release, but because pricing and capacity dynamics are turning against labs that sell proprietary access. Alderson, who writes at martinalderson.com, frames the moment as a mirror of last winter, when a frontier model release drove a boom in coding agent usage. His case rests less on which model is smartest and more on who can afford to sell inference cheaply.
OpenAI has cut the price of its fastest, cheapest tier by 80 percent and trimmed its flagship tier by 20 percent, Alderson notes, citing the company’s own pricing updates. Meta has gone further: it now prices its open Muse Spark 1.2 model at roughly a tenth of its standard rate on a contributor tier where usage may feed back into Meta’s training data. Alderson says that tier’s cache-read price undercuts every other API he tracks.
Anthropic has not matched those cuts. Alderson points to a Financial Times report on weak uptake of Fable 5, Anthropic’s top-priced tier at $10 per million input tokens and $50 per million output tokens, under the headline “Anthropic’s best AI model struggles to attract users as cheaper tools thrive.” That framing is the FT’s reporting and Alderson’s reading of it, not a disclosed usage figure from Anthropic. He adds that Anthropic’s own developer-relations account has been extending weekly usage limits while warning that capacity stays tight, which he reads as evidence of compute scarcity rather than weak demand.
Beyond Meta, Alderson counts at least five labs outside OpenAI, Anthropic and Google shipping models he considers strong enough for serious agentic work: Z.AI, DeepSeek, Kimi and xAI’s Grok, plus a widely rumored open release of Meta’s own frontier system. He also flags an unreleased model called Ox Alpha that has drawn attention purely on early buzz, without confirmed benchmarks.
Alderson’s sharper claim concerns serving cost rather than raw capability. He cites investor Gavin Baker’s point that GPU capacity contracted years ago near $2 per hour is rolling off those terms and could reprice toward $4, roughly doubling underlying compute costs across the industry. Whoever serves open models most efficiently, not whoever trained the smartest one, currently holds the pricing advantage, Alderson argues. He names Fireworks, Together and Cloudflare, three inference hosting providers, as firms competing directly on that margin.
Alderson’s case is strongest on the pricing evidence: the rate cuts from OpenAI and Meta are documented moves, and Anthropic’s public statements about tight capacity support his compute-scarcity reading. It is weaker where he leans on the FT’s framing of Fable 5 without independent adoption numbers of his own, and he concedes that a sudden leap in frontier intelligence or token efficiency could reopen the gap he says is closing.
Teams evaluating inference contracts for the next two quarters should weight token-efficiency benchmarks alongside raw capability scores. Alderson’s argument implies that the operators with spare compute, not necessarily the best model, will set the market price through the rest of this year.
Martin Alderson, “The summer of open weights,” martinalderson.com, published August 23, 2026.
