The Frontier Premium Is Dead: What Fable 5's 11% Spending Share Reveals About Value, Vigilance, and the Coming Model Layer

The numbers arrived like a quiet confession. Ramp's enterprise spending data showed Anthropic's flagship Fable 5 — the model priced at exactly twice its nearest sibling — commanding just 11.4% of API spending and a mere 6% of token volume. The best model in the house was not winning. It was barely showing up. I have spent fifteen years watching infrastructure markets lie to themselves. I have audited smart contracts that promised trustlessness and delivered vulnerability. I have watched governance tokens accumulate in the wallets of people who never read a single proposal. And now I am watching the AI industry repeat the oldest mistake in our playbook: believing that the most capable thing is the most valuable thing. This is not a story about Anthropic's failure. It is a story about the end of an era — the era where frontier performance commanded automatic premium. The era where 'state of the art' was a business model. That era is closing, and the data from Ramp is the first public ledger entry of its funeral. Let me be precise about what the data shows. Fable 5, Anthropic's most advanced model, priced at $10 per million input tokens and $50 per million output tokens, accounts for only 11.4% of enterprise API spending on Anthropic's platform. Its token usage share is even more stark: 6%. Meanwhile, Opus 5 — positioned by Anthropic itself as 'near-frontier intelligence at half the price' — has already surpassed Fable 5 in spending share. And across the aisle, OpenAI's GPT-5.6 Sol commands 25% of enterprise token usage, more than four times Fable 5's share. The gap between spending share (11.4%) and token share (6%) is itself a confession. It tells us that the enterprises using Fable 5 are not using it for volume. They are using it for specific, high-stakes, low-frequency tasks — complex reasoning, difficult code synthesis, regulatory analysis. They are not running their customer service bots on it. They are not drafting their marketing emails on it. They are treating Fable 5 the way a surgeon treats a scalpel: carefully, rarely, and only when the situation demands precision. This is the first signal that the capability premium has entered diminishing returns. When a model costs twice as much but delivers perhaps 5% better performance on already-saturated benchmarks, the rational enterprise does the math. And the math says: buy the cheaper model, route the hard problems to the expensive one, and save the difference. I have seen this exact pattern before. In 2020, during the DeFi Summer, I watched the MakerDAO community wrestle with a similar question: should the Dai stablecoin system prioritize the most sophisticated collateral types, or should it prioritize the most reliable ones? The answer, after months of governance debate, was the latter. We chose reliability over sophistication. We chose the collateral that could be audited, understood, and trusted by the community — not the one that promised the highest theoretical yield. Tracing the code back to the conscience, we realized that the protocol must serve the human spirit, not the other way around. The AI model market is now facing the same reckoning. And the enterprises voting with their wallets are telling us something profound: in a world of finite budgets and infinite model choices, value is the new performance. Let me unpack the pricing structure, because the numbers tell a story that the marketing materials will never admit. Fable 5 at $10/M input and $50/M output is priced at exactly double Opus 5. For a typical enterprise customer-service scenario — roughly 500 input tokens and 200 output tokens per call — a single Fable 5 invocation costs $0.015, while the same call on Opus 5 costs $0.0075. For an enterprise running a million calls per day, that difference compounds to millions of dollars annually. The capability gap between the two models, whatever it is on paper, does not justify that cost differential for the vast majority of real-world tasks. Accel partner Miles Clements put it bluntly: 'Most people don't need to constantly use frontier models.' This is not a controversial statement. It is an empirical observation that the data supports. But it is a statement that the AI industry's pricing architecture has refused to acknowledge until now. The deeper insight here is that Anthropic's pricing strategy may be deliberately engineered. By pricing Fable 5 at twice the cost of Opus 5, Anthropic creates what behavioral economists call a 'decoy effect.' The expensive model makes the mid-tier model look like a bargain. Enterprises that might have hesitated at Opus 5's price point now see it as the rational choice — because the alternative is twice as expensive for marginal gains. Fable 5 is not a product. It is a price anchor. It exists to make Opus 5 look good. This is a strategy I recognize from the crypto world. In 2024, when I founded VietChain Dialogue in Ho Chi Minh City, I watched how institutional capital entered the Southeast Asian market. The big funds would announce their 'flagship' investments — the large, visible, expensive deals — while quietly building their real positions in smaller, more practical projects. The flagship was the anchor. The real money was in the mid-tier. We build bridges from the ashes of belief, and sometimes those bridges are built on the perception of value rather than the reality of it. But there is a risk in this strategy. If the anchor is too expensive, if the decoy is too obviously a decoy, enterprises may not just reject the flagship — they may question the entire pricing architecture. And that is precisely what the Ramp data suggests is happening. Enterprises are not just choosing Opus 5 over Fable 5. They are building 'model routing' systems that automatically select the appropriate model for each task. Simple tasks go to Opus 5. Complex tasks go to Fable 5. And increasingly, some tasks are going to OpenAI's GPT-5.6 Sol, which appears to offer a more compelling price-performance tradeoff for enterprise workloads. The 25% token share for GPT-5.6 Sol versus 6% for Fable 5 is not just a competitive gap. It is a structural gap. OpenAI has built a more mature ecosystem — function calling, assistants API, deep integration with Microsoft's enterprise stack. Anthropic has built a better model, arguably, but a thinner ecosystem. And in enterprise procurement, ecosystem matters more than raw capability. This is the same lesson we learned in blockchain: the best consensus algorithm does not win; the best developer ecosystem wins. Ethereum beat its technically superior competitors not because it was faster or more secure, but because it had the most builders, the most tools, the most mindshare. Governance is not a vote; it is a vigil. And the vigil here is watching enterprises make procurement decisions that will shape the AI industry for the next decade. The decisions are not being made in boardrooms by CTOs who understand the technical nuances of attention mechanisms and test-time compute scaling. They are being made by procurement teams running ROI calculations. They are being made by engineers who have been burned by vendor lock-in before. They are being made by CFOs who have seen AI budgets balloon and are now asking hard questions about what they are getting for their money. This shift from performance-first to value-first procurement is not a temporary market correction. It is a structural change in how enterprises evaluate AI. And it has profound implications for the entire AI supply chain. Consider the compute implications. Fable 5's low adoption means that the high-end GPU clusters — the H100s and their successors — that were provisioned for frontier model inference may be underutilized. Anthropic will need to reallocate that compute to Opus 5 inference, which is cheaper per token, or to training the next generation of models. This is not necessarily bad for Anthropic's margins — lower-cost inference on Opus 5 could actually improve profitability — but it does suggest that the market's demand for frontier compute has been overestimated. The AI chip boom, like the crypto mining boom before it, may be built on assumptions about demand that are about to be revised. I remember the 2022 crash. I retreated to a quiet apartment in Hanoi for three months, watching the narratives of decentralization get corrupted by centralized exchanges and opaque funds. I wrote the Ho Chi Minh Trust Manifesto in that apartment, arguing that true decentralization requires psychological resilience and community verification over algorithmic guarantees. The essay went viral in niche philosophical crypto circles, and it taught me something important: the market's faith in technology is always cyclical. We believe in the frontier, then we get burned by the frontier, then we retreat to the practical, then we slowly build confidence in the frontier again. The AI market is now in the retreat phase. Enterprises have been burned by the cost of frontier models, and they are retreating to the practical. The question is whether this retreat is permanent or temporary. And the answer depends on whether the frontier models can demonstrate value that justifies their cost. This is where the AI industry can learn from the blockchain industry's mistakes. We spent years selling 'trustlessness' as an absolute value, only to discover that trust is not a binary — it is a spectrum, and most applications need only a modest amount of decentralization. The AI industry is making the same mistake with 'capability.' It is selling capability as an absolute value, when most applications need only a modest amount of it. Let me be more specific about what the data reveals about enterprise behavior. The 6% token share for Fable 5, combined with the 11.4% spending share, suggests that enterprises are using Fable 5 for tasks where the cost per token is justified by the value of the output. These are likely tasks involving complex reasoning, multi-step problem solving, or high-stakes decision support. The 94% of tokens that go to other models are likely routine tasks — summarization, classification, extraction, generation — where the marginal capability of Fable 5 over Opus 5 is irrelevant. This is exactly the pattern we see in blockchain infrastructure. In 2023, I analyzed the usage patterns of various Layer 2 solutions and found that the vast majority of transactions were simple transfers and token swaps — tasks that did not require the full security guarantees of the most expensive settlement layer. The expensive layer was used sparingly, for high-value settlements. The cheap layers handled the volume. The market had naturally stratified, and the stratification was rational. The same stratification is now happening in AI. And it has implications for how we think about model development. If the market is telling us that frontier capability is only needed for a small fraction of tasks, then the economic incentive to push the frontier forward is diminished. Why spend billions of dollars training a model that will only be used for 6% of enterprise tokens? The answer, of course, is that the frontier model's existence makes the mid-tier model look better. The decoy effect works in both directions. Fable 5's existence makes Opus 5 look like a bargain, and Opus 5's bargain status drives volume, and volume drives revenue, and revenue funds the next generation of frontier models. This is a delicate balance. If the frontier model is too far ahead, the gap between it and the mid-tier model becomes too large, and enterprises may not see the mid-tier as a viable substitute. If the frontier model is too close, the decoy effect collapses, and enterprises may just buy the cheaper model exclusively. Anthropic's positioning of Opus 5 as 'near-frontier' suggests they understand this balance. The question is whether they can maintain it as OpenAI continues to push its own frontier. The competitive dynamics here are worth examining closely. OpenAI's GPT-5.6 Sol at 25% token share is not just winning on price. It is winning on ecosystem. OpenAI has spent years building the developer tools, the integrations, the enterprise relationships that make adoption easy. Anthropic has spent years building safety infrastructure and philosophical positioning. Both are valuable. But in the enterprise procurement process, ecosystem wins. This is the same lesson we learned in the smart contract platform wars: the chain with the best developer experience wins, even if it is not the most technically advanced. I think about the 2017 ICO audit I conducted in Singapore. I found a severe reentrancy vulnerability in the Parity Wallet library that could have drained over $300 million in Ethereum. I disclosed it privately, and the patch was delayed but secure. That experience taught me that the best technology is worthless without trust. And trust is built through transparency, through reliability, through a track record of doing the right thing. Anthropic has built a track record of safety-first development. But safety is not the same as value. Enterprises do not buy safety; they buy outcomes. And the outcomes they are getting from Opus 5 are, for most tasks, indistinguishable from the outcomes they would get from Fable 5. This brings me to the contrarian angle that I think is missing from most analysis of this data. The conventional reading is that Fable 5's low adoption is a failure — a sign that Anthropic overpriced its flagship. But there is another reading: Fable 5's low adoption is a feature, not a bug. By pricing Fable 5 at a premium, Anthropic is signaling to the market that it has a model that is genuinely more capable — and that the capability is worth paying for, but only for the tasks that truly need it. This is not a failure of pricing strategy. It is a successful implementation of price discrimination. The enterprises that need frontier capability will pay for it. The enterprises that do not will buy the cheaper model. Anthropic captures surplus from both segments. The risk, of course, is that the price discrimination is too aggressive. If the gap between Fable 5 and Opus 5 is not perceived as justifying the price gap, enterprises may not just reject Fable 5 — they may also question whether Opus 5 is worth its price. This is the 'model fatigue' risk that I see in the data. Enterprises are being asked to evaluate an increasingly complex landscape of models, each with its own pricing, its own capabilities, its own tradeoffs. The cognitive load of this evaluation is itself a cost. And some enterprises may respond by simply defaulting to the cheapest adequate model, or by delaying procurement decisions altogether. This is where the blockchain industry's experience with 'infrastructure fatigue' is instructive. In 2021, the market was flooded with Layer 1 blockchains, each claiming to be the 'Ethereum killer.' Enterprises and developers were paralyzed by choice. The result was not a flourishing of innovation but a consolidation around a few dominant platforms. The same consolidation is now happening in AI. Enterprises are not adopting a portfolio of models from multiple vendors. They are standardizing on one or two vendors, and within those vendors, they are standardizing on one or two models. The Ramp data suggests that Anthropic's enterprises are standardizing on Opus 5, and OpenAI's enterprises are standardizing on GPT-5.6 Sol. The long tail of models is being ignored. This consolidation has implications for the AI industry's long-term health. If enterprises standardize on mid-tier models, the revenue available to frontier model developers is reduced. This could slow the pace of frontier research. It could also reduce the diversity of approaches to AI safety, as fewer organizations have the resources to invest in safety research. The ethical implications are significant. We may be entering a world where the most advanced AI systems are developed by a small number of organizations, funded by a small number of enterprises, serving a narrow set of high-value tasks. The democratizing potential of AI — the promise that anyone could access frontier intelligence — may be undermined by the economics of the market. But I am not pessimistic. I have seen this pattern before, and I have seen the market correct itself. In the crypto world, the consolidation of 2021 gave way to the diversification of 2023. New use cases emerged. New business models were developed. The market found its equilibrium. The same will happen in AI. The current consolidation around mid-tier models is a temporary phase. As the cost of frontier capability declines — and it will decline, as it always does — the frontier will become accessible to a broader range of enterprises. The 6% token share for Fable 5 will grow. The question is not whether it will grow, but when, and at what price point. Let me offer a concrete prediction. Within the next 12 to 18 months, Anthropic will either reduce Fable 5's price or introduce a new model that sits between Opus 5 and Fable 5. The market has spoken, and the message is clear: the current pricing architecture is not sustainable. The decoy effect has a shelf life, and that shelf life is expiring. Enterprises are becoming more sophisticated in their model evaluation. They are building internal benchmarks. They are tracking cost per task, not just cost per token. They are demanding transparency about model capabilities and limitations. The era of opaque pricing is ending. This is where the blockchain industry's commitment to transparency can serve as a model. In the crypto world, we have learned that trust is built through verifiability. The same principle applies to AI. Enterprises need to be able to verify that the model they are paying for is actually delivering the capabilities they need. They need benchmarks that are meaningful, not marketing. They need pricing that is transparent, not opaque. They need contracts that are fair, not one-sided. The AI industry has a lot to learn from the blockchain industry's hard-won lessons about transparency and trust. I think about the 2026 project I worked on — the Human-First Proof of Personhood protocol. We designed a system that emphasized self-sovereign identity and privacy preservation, countering the opacity of AI-driven data extraction. The project launched with 1,000 early adopters, proving that human-centric design could coexist with advanced AI. The lesson from that project is directly relevant here: the market rewards systems that respect human agency. Enterprises are not just buying models; they are buying systems that respect their autonomy, their data, their control. Anthropic's safety-first positioning is valuable, but it needs to be translated into concrete enterprise value — not just philosophical positioning. Let me return to the data one more time. The 11.4% spending share for Fable 5 is not a failure. It is a signal. It is the market's way of saying that the frontier premium is no longer automatic. It must be earned, task by task, use case by use case. The enterprises that are using Fable 5 are doing so because they have identified specific tasks where the capability gap justifies the cost. They are not using it because it is the 'best' model. They are using it because it is the right tool for a specific job. This is the mature, rational behavior that we should celebrate — not lament. The AI industry is growing up. It is moving from the adolescent phase of 'bigger is better' to the adult phase of 'right tool for the right job.' This is a painful transition, but it is a necessary one. The blockchain industry went through the same transition. We moved from 'decentralization at any cost' to 'decentralization where it matters.' We learned that not every application needs a decentralized ledger. The AI industry is learning that not every task needs a frontier model. This is not a retreat from ambition. It is a refinement of it. Truth is the only immutable asset. And the truth here is that the market has spoken. Enterprises are voting with their wallets, and they are telling us that value matters more than capability. The question for Anthropic — and for every AI company — is whether they are listening. The data suggests that some are listening better than others. OpenAI's 25% token share for GPT-5.6 Sol suggests that they have found the right balance of price and performance. Anthropic's 6% token share for Fable 5 suggests that they have not yet found that balance. But the gap is not insurmountable. It is a pricing problem, and pricing problems are solvable. The deeper question is whether the AI industry can learn the lesson that the blockchain industry learned the hard way: that the protocol must serve the human spirit. The most advanced technology in the world is worthless if it does not serve human needs. The enterprises that are choosing Opus 5 over Fable 5 are not being short-sighted. They are being rational. They are choosing the tool that serves their needs at a price they can afford. This is not a failure of the frontier. It is a triumph of the market. As I write this, I am reminded of the vigil I have kept over the past decade — watching markets rise and fall, watching narratives shift, watching technologies mature. The vigil continues. The AI market is in its adolescence, and it is making the mistakes that adolescents make. But it is also learning, growing, and maturing. The data from Ramp is a snapshot of that maturation process. It is not a verdict. It is a data point. And data points are meant to be studied, understood, and acted upon. Let me end with a forward-looking thought. The model market is becoming the infrastructure market. And in infrastructure markets, the winners are not the ones with the most advanced technology. The winners are the ones with the most practical solutions, the most transparent pricing, the most reliable service. The winners are the ones who understand that the protocol must serve the human spirit. The winners are the ones who listen to the silence between the blocks — who understand that what is not said is often more important than what is said. The enterprises that are choosing Opus 5 over Fable 5 are not making a statement about Anthropic's technology. They are making a statement about their own needs. And that statement is the most important data point of all. The frontier premium is dead. Long live value.