Anthropic passed OpenAI in US business AI adoption for the first time in April, claiming 34.4%1 of Ramp’s 50,000-plus1 customer base against OpenAI’s 32.3%1, per the May 2026 AI Index1 published May 13. The 3.8 percentage-point swing in a single month is less a gradual shift than a step change, driven by developer tooling that converts individual engineer enthusiasm into enterprise spend faster than traditional procurement cycles can govern it.
[Updated June 2026] The lead has held and widened. On June 9 Ramp reissued the index with methodology changes meant to better capture enterprise invoiced spend, the exact blind spot OpenAI complained about, and restated the levels to 41%6 for Anthropic against 39.5%6 for OpenAI, with OpenAI down 0.1 point. In other words, Ramp adjusted for the objection OpenAI raised and Anthropic still came out ahead.
The Crossover: What Ramp’s May Index Actually Shows
The headline numbers from Ramp1 are unambiguous: Anthropic’s share rose 3.8 points to 34.4%1 while OpenAI’s fell 2.9 points to 32.3%1. Overall paid AI adoption across Ramp’s sample reached 50.6%1, up just 0.2 percentage points month-over-month, which means nearly all the movement happened within the installed base rather than from new entrants. Anthropic quadrupled its business adoption year-over-year1; OpenAI grew only 0.3%1 over the same period.
That is a breadth metric, not a depth metric. Ramp measures which vendors show up on corporate cards and invoices across its customer base. It does not measure contract value, deployment scale, or API call volume. OpenAI’s spokesperson pushed back on this exact point, telling Axios2 that the company is “driving enterprise transformation through large invoiced contracts” that Ramp’s methodology misses. The objection is fair: a single Fortune 500 ChatGPT Enterprise deal could dwarf a hundred mid-market Claude Code subscriptions in revenue while registering as a single data point, or none at all, in Ramp’s sample. OpenAI has been pushing its own depth-oriented yardsticks for exactly this reason, like the internal-GPT deployment counts it now cites as an adoption metric, which capture seat-level usage that a card-spend index cannot see.
[Updated June 2026] There is a second-order wrinkle in the June restatement that deserves a skeptical read. When a data provider reworks its own methodology and the revision happens to push both vendors higher (Anthropic from 34.4% to 41%, OpenAI from 32.3% to 39.5%)6 while preserving the same ordering, the cleanest month-over-month comparison is gone. The April-to-June numbers now sit on two different rulers, so anyone trending the gap is splicing a pre-revision series onto a post-revision one. Ramp says the changes were meant to better capture enterprise spend it had been undercounting6, which is the correct direction to fix and also conveniently the objection OpenAI raised in May. The honest takeaway is narrow: on Ramp’s current methodology Anthropic leads by about 1.5 points6, and the slope of that lead is no longer something the index can prove on its own.
How Claude Code Drove Anthropic’s Quadruple-YoY Growth
The mechanism behind the surge is not abstract brand preference. It is specific tooling. Anthropic’s growth correlates with the adoption of Claude Code, the agentic coding assistant, and its integration into workflows inside Cursor and other IDEs. Developers adopt first; finance discovers the spend later.
The anecdote that best illustrates this dynamic came from Uber’s CTO, cited in Ramp’s commentary1, who said the company spent its entire 2026 AI budget in just four months, largely on Claude Code and Cursor, with engineers reporting monthly API costs between $5001 and $2,0001 per person. If accurate, that is not a pilot program. It is an uncontrolled burn rate that procurement only notices after the fact. The pattern is reproducible across any organization where developers hold corporate cards or can expense SaaS without IT review.
The Per-Token Procurement Trap: Why Lower Seat Fees Raise TCO
Anthropic’s enterprise pricing shifted in 2026 to a per-token model with mandatory monthly consumption commitments3, eliminating prior 10, 15%3 API volume discounts. The sticker price for a seat may look lower than competitors, but the total cost of ownership moves in the opposite direction for most organizations because the bill includes a mandatory commit floor plus metered overage.
[Updated June 2026] The shape of that restructuring is now concrete. Anthropic retired its old Premium ($2008 per user per month) and Standard ($408 per user per month) enterprise tiers and replaced them with two role-based seats: Claude Code at $208 per technical user per month and Claude.ai at $108 per business user per month. Those seat fees buy platform access only. Usage across Claude, Claude Code, and the new Cowork product is metered separately at standard API rates, and every account carries a monthly spending commitment that Anthropic sets from its own estimate of the customer’s token consumption; the buyer pays that floor whether or not usage reaches it8. The change applies on contract renewal, so the seat-price drop lands first and the metered bill arrives a month later. Groundy covered the mechanics of the cutover in Anthropic’s move to force per-token billing on enterprise Claude above 150 seats.
The structural problem is incentive misalignment. Ramp economist Ara Kharazian noted1 that Anthropic makes more money when businesses purchase more tokens, which creates pressure to steer customers toward more expensive models or higher consumption tiers. In a flat-fee subscription, the vendor absorbs compute volatility. In a token model, the buyer does. When an engineering team ships a feature that unexpectedly doubles API calls, the CFO sees it in next month’s invoice, not in a capacity planning meeting.
OpenAI’s Bundled Alternative and the Flat-Fee Advantage
OpenAI’s counteroffer is predictable cost. ChatGPT Business runs $204 per seat per month on annual contracts as of early April 2026, with bundled access included. ChatGPT Enterprise uses custom negotiated flat rates, typically between $404 and $100-plus4 per seat per month4. The unit economics are straightforward: the vendor carries the risk of overuse, and the buyer budgets a fixed line item.
That predictability carries its own tradeoffs. Flat-fee models can encourage over-provisioning or under-utilization, and they rarely offer the granular model selection that power users want. But for procurement teams writing 2027 budgets, a known quantity beats a variable one, especially after a year of token-price volatility and capacity shortages.
Three Threats to Anthropic’s Lead: Misalignment, Reliability, and Cost Creep
The first threat is the incentive structure Kharazian identified. A vendor that profits from every additional token has no natural incentive to help customers optimize usage, and every incentive to steer them toward more expensive models.
The second is operational. Token-based billing forces FinOps teams to monitor consumption in real time, set quotas, and negotiate commit tiers, none of which are standard competencies outside of cloud infrastructure teams. Most mid-market companies lack the tooling to attribute API spend to specific projects or engineers, which means the first sign of trouble is an oversized invoice.
The third is competitive response. OpenAI’s 0.3%1 business adoption growth over the past year suggests saturation at the high end, not weakness. If OpenAI chooses to compete on procurement convenience rather than model capability, it can undercut Anthropic’s enterprise TCO without touching its model weights.
The June Reset: Reliability and an Image-Token Surcharge
[Updated June 2026] Two developments since the May crossover sharpen the risk. The first is reliability. VentureBeat’s review of the index flagged recent Claude outages and rate-limit throttling7 as a live threat to the lead, because a per-token customer who gets throttled mid-sprint still owes the monthly commit. Availability is now a billing problem, not just an ops problem: the commit floor does not flex down when the service does. The second is a pricing change that triples the token cost of image-inclusive prompts7. For teams that pipe screenshots, diagrams, or design files through Claude, that is a step increase in the metered portion of the bill with no corresponding line on the seat invoice, which is precisely the kind of move the incentive critique predicts.
There is a counter-current worth naming, because it complicates the tidy “vendor profits from every token” story. Anthropic’s per-token rates have been falling, not rising. The current flagship, Claude Opus 4.8, lists at $5.009 per million input tokens and $25.009 per million output tokens, down from the $15/$759 that Opus 4.1 charged a year earlier, a roughly 67% cut on the rate card. So the squeeze on buyers is not the headline rate; it is the structure around it. Unit prices drop while consumption climbs faster, and the mandatory commit converts that volume into a fixed obligation. A CFO who benchmarks only the per-token sticker will read the trend as deflation and miss that the commit floor, the eliminated volume discount, and the image surcharge are where the money actually moves.
What Multi-Vendor AI Contracts Look Like After the Flip
The Ramp crossover changes the negotiation math for multi-vendor AI contracts. A year ago, most procurement teams treated OpenAI as the primary vendor and Anthropic as a secondary or experimental line item. Reversing that relationship means rethinking how consumption commitments, commit tiers, and model routing are structured.
Organizations running both vendors now face a harder problem than single-sourcing: they must forecast token consumption across two pricing models with different commit structures, overage penalties, and model upgrade cadences. The vendor that makes that forecasting easier, not the one that wins the benchmark, may determine where the next procurement cycle lands.
Anthropic’s $30 billion5 annualized revenue run rate as of April 2026, up from roughly $9 billion5 at the end of 2025, confirms the scale of the shift. But revenue run rate is a trailing indicator of adoption, not a leading indicator of retention. If per-token billing continues to produce sticker shock at renewal, the same procurement teams that approved Claude Code experiments may impose stricter controls, or migrate predictable workloads back to flat-fee alternatives.
Frequently Asked Questions
Can non-US organizations extrapolate from Ramp’s adoption percentages?
No. The index tracks US payment flows exclusively. EU and APAC markets have different vendor penetration, data-residency constraints, and procurement norms, so a company headquartered in Frankfurt or Singapore should treat the 34.4%/32.3% split as a directional signal rather than a planning assumption.
How does Anthropic’s token commitment compare to cloud reserved-instance models?
Both require volume commitments for lower unit rates, but Anthropic’s monthly consumption floor renews far faster than the 1–3 year lock-in typical of AWS or Azure reserved instances. The key difference: Anthropic eliminated its prior 10–15% volume discounts entirely, so the commit tier is now the only pricing lever, there is no discount to stack on top of it.
What should a CFO verify before renewing an Anthropic enterprise agreement?
Isolate Claude Code spend from API usage on other Anthropic models, then compare per-engineer monthly cost against a flat-fee ChatGPT Enterprise seat at $40–100/month. Teams where average API spend exceeds roughly $200 per engineer per month, well within Uber’s reported $500–$2,000 range, should negotiate a commit tier pegged to actual trailing consumption rather than projected growth.
Where does the flat-fee versus per-token framing break down?
It assumes single-vendor selection. Teams already running Claude Code for development and ChatGPT for general knowledge work operate hybrid spend, token-metered for one, flat-fee for the other, which means the real forecasting challenge is managing two incompatible billing structures simultaneously, not choosing between them.
What could reshape the token-vs-flat-fee calculus in the next 12 months?
If Anthropic’s revenue trajectory continues toward analyst estimates of roughly $44 billion by mid-2026, the aggregate volume of enterprise token commitments could be large enough to support secondary markets for commit reselling or token-hedging instruments, similar to how cloud marketplace economics evolved. Until then, buyers absorb 100% of consumption volatility with no hedge.
