Ramp Data: Enterprises Adopt Anthropic But Reject the Frontier

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last verified: August 2026 - Ramp AI Index, Claude Fable 5, GPT-5.6 Sol, token-spend allocation across US business customers

Adoption Is Not Spend

The most misunderstood chart in enterprise AI right now comes from Ramp's spending data, and it is one line that flips the whole narrative: Anthropic leads adoption at 43.5% of US business clients, yet only 6% of those clients' token spend goes to its frontier flagship, Claude Fable 5. In the same dataset, OpenAI's GPT-5.6 Sol captures 25% of OpenAI customers' token usage, and open-source and Chinese models have quietly reached 6.1% of AI-spending customers.

Adoption measures who signs a contract. Spend measures what actually runs in production. The two have diverged so sharply that they are now telling opposite stories: Anthropic won the enterprise relationship war while losing the production volume war inside its own accounts. If you are building or buying AI infrastructure, this divergence is not trivia - it is the single best evidence we have for how enterprises actually procure models in 2026. The AI news desk has tracked this data all year, and the pattern has hardened into a rule.

The Ramp Numbers: A Split Personality

Ramp's panel of more than 50,000 US businesses tracks two things separately: the share of companies paying each provider (adoption), and where the actual token dollars flow (spend). The split is stark. Anthropic reached 43.5% adoption while drawing just 6% of customer token spend to Fable 5. That is the signature of an account being won at the relationship layer - security review, procurement approval, an enterprise seat program - while the actual inference load runs on cheaper tiers.

OpenAI tells the mirror-image story. Fewer clients sign the contract, but the clients that do are burning real volume, and GPT-5.6 Sol at 25% of OpenAI customer token usage shows that a mid-tier model can dominate a provider's own token mix. Sol is the workhorse: strong enough for production, cheap enough to run continuously. Meanwhile, open-source and Chinese models reaching 6.1% of AI-spending customers is the first time the "alternative tier" has shown up as a measurable line item rather than a pilot footnote. Put together, the data describes an enterprise that has learned to buy intelligence the way it buys compute: on a price-performance curve, not on brand loyalty.

The Enterprise Price-Performance Floor

What the data really reveals is a price-performance floor that enterprises will not cross. Every task has an implicit ceiling: the most you are willing to pay per unit of output quality. Frontier models like Fable 5 cross that ceiling for most routine work, so they get bought for credibility and reserved for the tasks that genuinely need frontier reasoning - complex agentic coding, long-horizon planning, high-stakes drafting. Everything else is routed to the cheapest model that clears the quality bar.

This is a structural shift, not a temporary cost optimization. Early 2026 was full of stories about teams hitting budget ceilings and pulling back; Ramp's own data showed token spend up 13x since January 2025, and the top 1% of firms now spend about $7,400 per employee per month versus roughly $12 for the median firm - a 600x spread. The firms winning that game are not the ones spending the most; they are the ones whose routing layers push work down to the floor. Teams that want the operational template for this belong in our workflows section, where the routing and governance patterns are documented end to end.

Benchmark: Where the Token Dollar Actually Lands

The lesson of the 6% number is that the frontier tier is a thin premium on top of a much larger workhorse base - not the other way around. Most organizations should be planning for a mix that looks like 5-10% frontier, 30-40% workhorse, 40-50% fast mid-tier, and the remainder open-source and on-device.

Procurement: The 60/30/10 Playbook

The Ramp data points to a procurement model that is already standard practice in mature teams. The 60/30/10 rule: about 60% of volume runs on cheap, reliable mid-tier models that clear the quality bar for 90% of tasks; 30% runs on workhorse models like Sol when the task needs stronger reasoning or longer context; and 10% runs on frontier models for the work that measurably fails on the cheaper tiers. The routing decision is an evaluation question, not a preference question: you cannot know which tier clears the bar without a per-task-class eval harness that measures quality, not vibes.

Two operational truths fall out of the data. First, prompt caching is now a budget line, not an optimization: one controller at AngelList told Ramp that a caching fix recovered about $10,000 a month that had been silently leaking. Second, model switching is fast and cheap, so vendor stickiness has collapsed; Ramp's own economist noted that businesses can move volume between providers in weeks. That means procurement is no longer a contract negotiation; it is a continuously re-run routing decision.

def route(prompt, task_class):
    # Enterprise tier router: push work down to the price-performance floor
    if task_class == "frontier_only":
        return "claude-fable-5"          # <10% of volume
    if task_class == "agentic_coding":
        return "gpt-5.6-sol"             # workhorse tier
    if cheap_quality_eval(prompt)["pass"]:
        return "claude-haiku-class"      # fast mid-tier
    return "open-source-qwen"            # bulk tier

def cost_per_month(tokens, mix):
    per_tier = {"frontier": 0.10, "workhorse": 0.30, "mid": 0.50, "open": 0.10}
    blended = sum(per_tier[k] * PRICE[k] for k in per_tier)
    return tokens / 1_000_000 * blended

ROI: What the 6% Actually Costs

Run the unit economics for a firm spending $100K per month on inference. If they spend like the Ramp average - 6% frontier, the rest across workhorse and mid-tier - their blended rate is around $4-$6 per million tokens, and they get roughly 18-25M tokens for that $100K. If they naively run the same workload entirely on the frontier tier at $15-$60 per million tokens, the same output costs $270K-$1.5M. That is not a 10% optimization; it is a 63% to 93% cost difference on the exact same workload.

The reverse insight is just as important: the 6% frontier spend is not waste. Those dollars buy the highest-value reasoning in the portfolio, and cutting them to zero would degrade agentic and planning quality disproportionately. The efficient allocation is not "spend less" but "spend at the right floor per task." A firm doing 100M tokens a month that moves 20% of its frontier-tier volume down to mid-tier (a ~5x price cut) saves roughly $240K per year at blended rates, with no measured quality change on the routed tasks - provided the eval harness validated the move first. That is the ROI case for the price-performance floor in one number.

Frontier Rejection Is Rational

None of this is a rejection of frontier AI. It is a rational rejection of paying frontier prices for non-frontier work. The models are not failing; the market is finally pricing them correctly. Fable 5 gets bought, benchmarked, and then routed to the slice of work where its cost is justified. Sol gets the volume because it is good enough and dramatically cheaper. Open-source and Chinese models get the long tail because at $0.15 per million input tokens, running them everywhere is cheaper than thinking about whether to run them.

What It Means for Model Procurement

The practical consequences are clear. Evaluate per task-class, never per vendor. Build a router, not a preference. Treat the frontier tier as a reserved capability, not a default. Negotiate with the data, because providers know exactly how this splits and will price accordingly. And build an exit path from every provider, because the Ramp data proves the switching cost is now measured in weeks, not years.

The workflows on this site walk through the routing and governance layer in detail, and the MCP directory is the tool-interoperability half of the same stack. The bottom line is that enterprise AI procurement in late 2026 is a portfolio problem. Adoption decides who you are allowed to buy from. Spend data decides who you actually run. And the firms that separate the two - like the Ramp panel is already doing - are the ones whose budgets survive contact with production reality.