The data from Ramp, tracking 70,000 companies, tells the story: Fable 5's share of Anthropic spending plateaued at about 11%, while the cheaper Opus 5, released in late July, has already overtaken the flagship in enterprise outlays . This isn't about Fable 5 being a bad model. It's about enterprise customers no longer defaulting to the most expensive option. As Accel partner Miles Clements, whose firm invested nearly $1 billion in Anthropic, put it: the period when customers mostly chose flagship models was "not a sustainable era" .
What Changed: The "Always Buy the Best" Playbook Just Died
For years, the business model of frontier AI labs rested on a simple assumption: enterprise customers migrate to the most powerful model available, justifying the billions spent training each new generation. That assumption is now breaking.
The pricing gap is too wide. Anthropic's Fable/Mythos 5 charges $50 per million output tokens, while Claude Haiku costs $5, and Chinese open models can be even cheaper . Customers are realizing the premium doesn't always deliver proportional value. AI legal assistant Harvey cut its inference cost to one-third by combining Claude Opus with Zhipu's open model GLM 5.1, without sacrificing quality . Lindy shifted its entire workflow from Anthropic to DeepSeek V4, saving millions and reportedly improving performance .
Coinbase CEO Brian Armstrong put it simply: about 80% of AI workload doesn't need the frontier—it just needs "good enough" . If that holds, the economics of the AI industry shift fundamentally.
The Political Headwind Didn't Help
Fable 5's launch was disrupted by the Trump administration over national security concerns in early June, forcing Anthropic to withdraw the model before it was approved for release on July 1 . Government-mandated data retention rules also slowed enterprise adoption . But analysts note that price and performance now outweigh political factors in customer decisions.
The Other Side: Open Models Are Winning the Volume Game
As Fable 5 falters, open-weight models are winning on volume. According to OpenRouter data, Chinese models now account for roughly 63.5% of global model token volume, with U.S. models holding just 35.5% . DeepSeek V4 Flash, MiniMax M3, and Z.ai's GLM models consistently top the global token consumption charts . Alibaba's Qwen family crossed 2 billion downloads on Hugging Face, more than ten times Google's 4.18 billion and Meta's 2.27 billion combined . In H1 2026, Chinese open models accounted for 41% of downloads on Hugging Face, surpassing the U.S. for the first time .
A Kenyan developer described the choice simply: using Chinese models is like buying a house; using American models is like renting . The cost advantage is undeniable—according to some estimates, including infrastructure, the total cost of using Chinese models can be up to 90% lower.
What This Means for Anthropic's IPO
The stakes are high. Anthropic is preparing what could be the largest IPO in history, with bankers floating a $2 trillion valuation and raising over $100 billion . The company reported $65 billion in annualized revenue in July, missing the most optimistic $80 billion forecast . Its internal projection of $1.9 trillion to $2 trillion in ARR by 2028 underpins the IPO valuation—roughly 10 times forward revenue .
If enterprise customers are no longer willing to pay a premium for the frontier, that projection becomes harder to defend. Anthropic's own Claude Code product, which drove its revenue surge past OpenAI earlier this year, is also showing signs of deceleration . As Ramp's chief economist noted, "If you impute previous trends, you expect Anthropic to own the market. But because OpenAI's newest model was so good and Fable underperformed..."
The implication extends beyond Anthropic. The four largest cloud vendors have pushed annual capital expenditure to nearly $750 billion. If model-side revenue growth slows, the sustainability of that upstream spending comes into question .
P.S. Accel's Miles Clements made a point worth sitting with: breakthrough AI advances still matter—for attracting talent and fulfilling promises like curing diseases—but they may increasingly serve a "showcase" role rather than a commercial one . The market is separating what's impressive from what's profitable.
Frequently Asked Questions
Q: How much did Fable 5 capture in enterprise spending?
A: Fable 5 captured just 11% of enterprise spending on Anthropic's tools two months after launch, according to Ramp data tracking 70,000 companies . The cheaper Opus 5, released in late July, has already surpassed it .
Q: Why are enterprises avoiding Fable 5?
A: Enterprises are shifting to cheaper alternatives. Fable 5 is priced at $50 per million output tokens, while Opus 5 is more affordable, and open models are even cheaper . Most businesses don't need frontier-level performance for everyday tasks .
Q: What is the "frontier premium" problem?
A: For years, enterprise customers defaulted to the most powerful available model, justifying billions in training costs. Fable 5's weak adoption suggests that dynamic is reversing—customers are now weighing cost against performance .
Q: How are Chinese open models performing?
A: According to OpenRouter data, Chinese models now account for roughly 63.5% of global token volume, while U.S. models hold 35.5% . DeepSeek, MiniMax, Z.ai, and Qwen consistently top the token consumption charts .
Q: What happened with Fable 5's launch?
A: The Trump administration forced Anthropic to withdraw Fable 5 in early June over national security concerns. It only returned online July 1 after approval . Government data retention rules also slowed adoption .
Q: What does this mean for Anthropic's IPO?
A: Anthropic is preparing a potential $2 trillion IPO . The weak uptake of Fable 5, combined with $65 billion in annualized revenue missing the $80 billion forecast, raises questions about the sustainability of its high-growth narrative .
Q: Are enterprises switching from frontier models?
A: Yes. Lindy shifted its entire workflow from Anthropic to DeepSeek V4, saving millions and improving performance . Harvey cut inference costs to one-third by combining Claude Opus with Zhipu's GLM .
Q: How does the pricing gap affect adoption?
A: Anthropic's Fable/Mythos 5 costs $50 per million output tokens, while cheaper alternatives can cost far less . This makes model routing—sending simple queries to cheap models and complex ones to expensive models—economically attractive .
Q: What is the "volume game"?
A: Open-weight models, particularly from Chinese developers, are winning on token volume. They're cheaper, freely downloadable, and increasingly competitive on performance . This creates a different business model—one built on ecosystem scale rather than per-token revenue.
Q: What are the broader implications for AI investment?
A: If model-side revenue growth slows, the sustainability of massive upstream capital expenditure—pushed to nearly $750 billion by the four largest cloud vendors—comes into question . The industry may need to rethink the economics of frontier AI.
