The numbers landed with a thud. Ramp's enterprise spend data dropped a quiet bomb: Anthropic's flagship Fable 5 accounts for just 11.4% of API expenditure across its corporate client base. Token usage paints an even starker picture — a mere 6% of total enterprise token volume. Compare that to OpenAI's GPT-5.6 Sol, which captures a commanding 25% of enterprise token usage. The spread is not a blip. It is a signal. Code does not lie. Check the contract.
For months, the narrative in AI circles has been a simple arms race. Frontier model, frontier price, frontier adoption. Anthropic priced Fable 5 at double the cost of Opus 5 — $10 per million input tokens, $50 per million output tokens. They positioned it as the pinnacle. The market has responded with a collective shrug. Follow the smart money, not the tweets. The smart money in enterprise AI procurement is not buying the most expensive product on the shelf. They are buying the one that delivers adequate performance at half the cost. This is not a failure of Fable 5's underlying capability. It is a failure of pricing theory. The market has moved past the notion that absolute performance wins. Value is the new benchmark.
The story here is not that Fable 5 is a bad model. The story is that Anthropic built a "price anchor" — a decoy. And the market called their bluff. Enterprise procurement teams have become sophisticated. They are not chasing MMLU leaderboard scores. They are calculating unit economics per query. And when you do the math, Opus 5 wins every time.
The Context: A Two-Tier Strategy in a Commoditizing Market
Anthropic designed a two-tier offering. Opus 5 sits as the "high-performance, reasonable-price" middle tier. Fable 5 is the über-tier, priced at precisely double the Opus 5 rate for both input and output. The official positioning narrative is that Opus 5 is "closer to Fable 5's frontier intelligence at half the price."
This is a classic decoy effect. By setting an expensive anchor, you make the mid-tier option look more attractive. It is a retail tactic. It works in supermarkets when you sell a $50 bottle of wine next to a $20 bottle. But enterprises are not impulse buyers. They run annual cost projections. They see through the trap. My analysis of the data suggests this decoy strategy has backfired. Instead of driving customers to Opus 5 by comparison, it has simply highlighted that the performance delta between the two models is not worth the price delta.
Enterprises have voted with their wallets. The 11.4% spend share versus 6% token share for Fable 5 tells a specific story. This is not a model being used for high-frequency, general-purpose tasks. This is a niche tool. The clients who do use Fable 5 are likely running specific, low-frequency, high-value tasks — complex reasoning, deep code generation, novel contract analysis. The token volume is low because the queries are sparse and expensive. The expenditure share is higher than the token share because the unit price is high.
The reverse is true for GPT-5.6 Sol. A 25% token share indicates broad deployment. It is being used for a wide array of everyday tasks. OpenAI has built a model that enterprises feel comfortable running at scale. Anthropic has built a product that is used sparingly, as a scalpel, not as a workhorse. This is a strategic difference in market positioning.
The Core Evidence: Wallet Votes and Model Routing
The on-chain analogy here is liquidity. Liquidity leaves before the crash hits. In AI, enterprise wallet data is the liquidity. The Ramp data confirms that the liquidity of the AI API market is flowing toward mid-tier models. This is not a one-off data point. Accel partner Miles Clements recently noted that "most people don't need to continuously use frontier models." That aligns perfectly with the data. The era of paying for SOTA simply because it is SOTA is over.
Let's get into the cost breakdown. Consider a standard enterprise customer support query: approximately 500 input tokens and 200 output tokens. On Fable 5, that is $0.015 per interaction. On Opus 5, it is $0.0075. Half the cost. For a company running one million of these queries per day, that is a difference of $7,500 daily. That is over $2.7 million annually. That is a headcount. That is a data center bill. That is the difference between a project getting approved and killed.
The user is not paying for "slightly better reasoning." They are paying for the capability to run a business profitably. This is where the "value-first" procurement framework takes over. A decision that was previously based on the "best model available" is now a calculation of "good enough" vs. "expensive." The token usage data confirms that enterprises are opting for "good enough" at scale.
I see this pattern replicated in the crypto world. We don't see massive wallets paying exorbitant gas fees for simple transfers; they route through Layer-2s. The infrastructure has to match the use case. The same is happening in AI. We are witnessing the rise of the "model router." These are the middleware solutions that automatically triage a prompt: simple task to Opus 5, complex logic to Fable 5, and standard retrieval to a smaller model. This is the layer that will capture value in the next phase of AI infrastructure.
Anthropic is trying to be the oil company in a world that has moved to solar panels. They are pushing the high-margin product that is too expensive for the grid. The market is shifting to a renewable, cheaper model. The contrast is stark when looking at OpenAI. They have not only captured 25% token share but they have built the ecosystem that allows for this routing. They have the API integration, the Function Calling, the simple-to-use Assistants API. They have the "suite" approach, making it easier for enterprises to build and deploy. Anthropic has the pure model, but the ecosystem is the lock-in. If you are an enterprise, why do you buy Fable 5 when the ecosystem around OpenAI is more mature and the price is lower? It is a network effect problem. And in network effects, the winner is the one who owns the developer workflow. It isn't the one who owns the most complex algorithms.
The data also suggests that the high-end compute is being underutilized. If Fable 5 is not being used at scale, then the H100 clusters dedicated to serving it are running at lower utilization rates. In contrast, the mid-tier compute for Opus 5 is likely running hot. This is a resource allocation problem. Anthropic is spending millions to maintain a cluster for a product that is not a top seller. It is an inefficient capital allocation. In a private market where cash is king, this is a risk. They could potentially be utilizing that compute to train the next model, or they could be serving Opus 5. The market is telling them to shift the compute to the mid-tier.
The Contrarian Angle: Correlation is Not Causation — and the Data May Be Distorted
The narrative is seductive. It screams "Anthropic's flagship is a failure." But let's stop and question the data source. Ramp is an expense management company. Their data skews heavily toward mid-market and smaller enterprises. They are not the Fortune 100. They are the Agile startup buying tokens with a credit card.
This is a critical blind spot. The largest enterprises — the JPMorgans, the Microsofts, the defense contractors — they do not buy API credits from a dashboard. They sign direct contracts. They create private deployments. They likely negotiate volume discounts that make the raw pricing irrelevant. These are the deals that move the needle for Anthropic's actual revenue, but they are invisible in Ramp's dataset.
So, the data might be telling us about the SMB market, not the enterprise market. The SMBs are price-sensitive because they have finite burn. The Fortune 500 is buying for a specific, high-value use case like deep code synthesis or proprietary data analysis. For them, Fable 5 may be the only model that can do the job, and the price is a rounding error on their balance sheet.
Additionally, Anthropic's pricing might be a strategic "haircut" to the value of the model. By setting Fable 5 so high, they're telling the market "this is the platinum, this is for the very best." They are not aiming for volume. They are aiming for margin. It is a classic product differentiation strategy. They might be okay with only 11% of spend if that 11% has a 90% gross margin. The strategy could be working exactly as intended — capture the high-end, and let the volume of Opus 5 bring in the revenue. The "Struggles" narrative is a media interpretation of a strategy they don't fully understand. The fact that Opus 5 is exceeding Fable 5 in spend share may be the plan.
But this strategy has a shelf life. The "security" angle is another layer. Anthropic has built its brand on AI safety. They are the "safe" option. Yet, if the market is price-sensitive and they are pushing clients to Opus 5, are they pushing them to a less safe model? Is there a "safety discount"? If an enterprise has to choose between a safe-but-expensive model and a cheaper-but-slightly-less-safe model, will they choose safety? In my experience auditing the crypto markets, I have seen that security is the first thing cut when budgets are tight. I suspect the same is happening in AI.
The Takeaway: Watch the Price, Not the Benchmarks
Over the next quarter, I will be watching for a specific set of signals. The first is pricing. If Anthropic is serious about competing in the enterprise market, they will cut Fable 5's price. If they cut the price by 20-30%, adoption will jump. The demand elasticity is high. They have already learned that lesson. If they do not cut the price, they have decided to keep the halo product and will rely on Opus 5 for revenue.
The second is the "model router" market. If we see a surge in startups building this middleware, we know the market is moving to the "blended model" approach. That is a big opportunity.
The third is OpenAI. They have already won the token share. If they lower the price of GPT-5.6 Sol, they will cut the legs out from under Opus 5. This is the bigger threat to Anthropic's actual revenue. The battle is no longer at the frontier. It is at the mid-tier. That is where the war for the enterprise will be won or lost.
Follow the smart money, not the tweets. The smart money is buying the mid-tier. Are you listening, Anthropic?