Anthropic's Fable 5 stalls as corporate AI spending hits a ceiling
Spending data is questioning the core bet of the frontier labs — that each new, pricier model sells itself — while Morgan Stanley warns compute stays scarce for years and the milestones of automated AI research keep falling.
Anthropic's most capable model is its slowest seller: Fable 5 accounted for just 6 percent of the tokens companies purchased from Anthropic in its first month, according to spending data from fintech Ramp. By comparison, OpenAI's flagship GPT-5.6 Sol captured 25 percent of tokens and 23 percent of spending at OpenAI, and Fable 5 brought in only about 75 percent of the model-related revenue Sol generated — despite costing roughly twice as much, at $10 per million input tokens and $50 per million output tokens. Ramp economist Ara Kharazian reads the pattern as a new ceiling on corporate willingness to pay: the extra capability simply isn't worth the premium when the performance gap barely shows up in daily work. The caveat is that Ramp's sample skews toward tech companies, so real-world adoption may be even lower.
The numbers matter because the entire frontier investment thesis rests on willingness to pay for each new generation — and they land alongside signs that the top of the market is saturating: Ramp puts Anthropic ahead of OpenAI in overall US corporate adoption (43.5 percent of companies vs 39.7 percent), yet its priciest tier is the weak spot, while advanced users drift toward open models that now trail frontier quality by only a few months. Total AI spending is still climbing — the top 1 percent of companies spent a median of $7,400 per employee in July — but buyers are increasingly spending more overall while balking at the top price point. That is the quiet version of a problem the labs don't want to name: if a 2x price only buys a barely measurable capability edge, pricing power has a ceiling even as demand grows.
Morgan Stanley strategist Michelle Weaver warns the AI buildout now faces a compute bottleneck that capital alone cannot fix, with shortages expected to persist for years. Enterprise AI adoption is accelerating — about 25 percent of S&P 500 companies can now quantify returns on their AI investments, up from 14 percent a year ago — but supply is still "severely insufficient," in Weaver's assessment. The constraints are no longer chips or money (Nvidia-linked financing plans are seeking up to $500 billion): they are the labor shortage in data-center construction and an electricity supply gap of roughly 10 to 20 percent even counting fixes like converting Bitcoin mining sites and deploying fuel cells, plus permitting delays and political headwinds. The takeaway is that compute is becoming a constrained resource priced for scarcity, and the industry's growth curve now bends around power lines and construction crews, not GPUs.
An assessment from IAPS fellow Severin Field, who interviewed 25 researchers at OpenAI, Anthropic, Google DeepMind, Meta, and US universities about recursive self-improvement, argues the automation of AI research is no longer hypothetical — several predicted milestones have already fallen. Twenty of the 25 rated automating AI research among the most severe and urgent risks, and Field points to what has happened since the interviews: OpenAI and DeepMind reached gold-medal level at the Math Olympiad, Sakana's "AI Scientist" produced a peer-reviewed workshop paper, Andrej Karpathy built an agent setup that runs its own training cycles, and Anthropic reports Claude now writes more than 80 percent of its production code. Just 4 of 20 respondents expect research-capable models to ship as public products — half expect them to stay internal, an "incentive flip" where keeping a model secret becomes more valuable than selling it. It's the latest turn in a debate we've tracked since a think tank made the case for pacing AI self-improvement earlier this month.
What to watch: whether Anthropic responds to Fable 5's slow start with pricing changes or bundling — and whether the compute bottleneck starts showing up in real deployment delays.
Would you pay double for a frontier model whose edge you can't measure at work? Tell us in the comments.
