80% of Companies Buying AI Models Now Pay Anthropic: What It Means for Finance Teams

Anthropic now reaches 80% of businesses that buy foundational AI models, up from 29% in October 2024, according to Ramp Rate's Anthropic vendor data. OpenAI sits at 78%, down 17 percentage points from a year earlier, per Ramp Rate's OpenAI page.

Those two numbers add up to more than 100%. That means at least 58% of companies buying AI models paid both vendors in the past 12 months. For finance teams, AI has stopped being one software line item. It is now two or more vendors, billed through a mix of seat licenses, usage-based API invoices, and employee card charges. Each one needs its own budget logic, approval path, and month-end treatment.

What happened

Anthropic is now the most adopted AI model vendor among US businesses, and most of its growth happened in five months. Adoption among businesses that buy in the foundational LLM category went from 43% in December 2025 to 75% in May 2026, then reached 80% by September.

The Ramp Rate foundational LLMs category is built from aggregated, anonymized card and bill pay data across more than 70,000 US businesses. The adoption rate is the share of businesses buying anything in the category that bought from a given vendor in the trailing 12 months.

The September 2026 snapshot:

Two details stand out. First, 61% of businesses making their first purchase in the category chose Anthropic. Second, 54% of Anthropic's current customers bought from another model vendor first. Companies are adding a second vendor, not just picking one.

What a foundational model vendor actually bills you for

AI model spend reaches the books through three separate channels, and most finance teams only see one of them clearly. A company can look like it spends $2,000 a month on AI when the real number, across seats, API usage, and personal cards, is three or four times that.

The three channels:

• Seat licenses. Chat products sold per user. Anthropic's Claude Team plan costs $20 per standard seat per month billed annually, or $25 billed monthly. OpenAI's ChatGPT Business plan is also $20 per user per month billed annually and $25 billed monthly.
• API usage. Metered by tokens, billed after the fact. Anthropic's API pricing lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. OpenAI's API pricing lists gpt-6.1-sol at the same $2 and $10.
• Hybrid enterprise contracts. Anthropic's Enterprise plan is "$20 per seat per month plus usage billed at API rates," so one contract carries a fixed and a variable component.

The categories blur in practice. An engineer's API key on a personal card is API spend that looks like a software subscription. A team that upgrades from monthly to annual seats cuts that line by 20% without changing usage.

Why this matters for finance teams

The move from one AI vendor to two changes the finance job from approving a subscription to managing a variable cost category. The a16z survey of 100 enterprise CIOs found that 37% of respondents used five or more models in 2025, up from 29% the year before. It also found that innovation budgets had dropped from a quarter of LLM spending to 7%. AI has moved into permanent operating budgets.

Three implications:

• Budgets need a usage forecast, not a renewal date. Seat costs are predictable. API costs move with product usage, model choice, and prompt design. Anthropic's pricing docs note that its newer tokenizer "produces approximately 30% more tokens for the same text," so a model upgrade can raise costs with no change in the rate card.
• Vendor concentration risk now runs both ways. With adoption above 75% for both leading vendors, most companies depend on two suppliers whose pricing and model lineups change several times a year. Renewals need the same review you would give any critical supplier, covered in our procurement best practices guide.
• Shadow AI spend is real spend. Individual seats bought on employee cards rarely pass through procurement. A $25 monthly charge is small. Forty of them across six departments is $12,000 a year that nobody budgeted.

The most common failure we see: a company signs an annual contract with one vendor, then finds at year end that half the teams had switched to the other one on expense reports. It paid for both, and the contract usage never justified the commitment.

What changes for budgeting, close, and controls

Finance teams that handle AI spend well treat it like cloud infrastructure: inventory first, then allocation, then controls. The work is the same as broader SaaS spend management, with one difference. Usage bills arrive after the month they cover.

Inventory. Pull every charge from both vendors across card feeds, AP, and expense reports. Card statements catch most of the shadow seats. Treat small, scattered AI subscriptions as tail spend and consolidate them onto a company plan where it makes sense.

Allocation. Assign each API key and each seat to a cost center. The FinOps Foundation's AI guidance recommends a showback model "to provide visibility into the costs incurred by different teams." Product usage of the API often belongs in cost of revenue, while internal chat seats usually sit in SG&A.

Month-end. API invoices for September usage often arrive in October. Under accrual accounting, estimate the month's usage from the vendor's usage dashboard and book the accrual before close, rather than expensing the invoice in the month it arrives.

Controls. Issue a dedicated card per vendor with a monthly cap, so a runaway script hits a limit instead of the P&L. A dedicated virtual corporate card for each API account also makes reconciliation a one-line match. Some spend platforms now track token usage by provider, model, and user. Ramp's AI token spend management is one example.

Accounting treatment. API access is generally a service cost, expensed as consumed. Implementation work on a cloud computing arrangement is evaluated under internal-use software guidance. PwC's summary of ASU 2025-06 confirms that scope, and Deloitte notes the update takes effect for annual periods beginning after December 15, 2027. Confirm capitalization calls with your auditor.

How this scales by headcount

A worked example

A 200-person company runs 60 Claude Team seats and 40 ChatGPT Business seats, all billed monthly at $25. Its product team uses about 500 million input tokens and 100 million output tokens a month on a $2/$10 model.

Two changes move that number. Switching seats to annual billing drops them to $2,000 a month, a $6,000 annual saving. Upgrading to a model with a tokenizer that produces 30% more tokens raises the API line from $2,000 to about $2,600 at the same rates. The net effect is close to a wash, which is why the forecast needs both levers in it.

The international angle

Companies with global teams or customers face two extra costs. Both Anthropic and OpenAI charge a 10% premium for regional or data residency endpoints. Teams that need EU data processing pay that on every token.

Both vendors bill in US dollars. Teams paying on cards issued outside the US take an FX markup on every charge, so a business card with no foreign transaction fees matters more as usage grows. VAT and GST treatment of digital services also varies by country. Check whether the vendor charges local tax or whether your entity self-assesses it.

How this fits the broader software picture

AI models are now the largest growth area in business software. Ramp Rate ranks foundational LLMs as the #1 software category by adoption, with 66% of software-buying businesses purchasing in it. Menlo Ventures estimates companies spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024.

The pattern looks like early cloud adoption. Teams bought on cards first, finance found the spend later, and the companies that built allocation and forecasting early had far fewer surprises at renewal. The same playbook works here, along with the usage review habits covered in our SaaS spend reduction playbook.

What to watch next

• Pricing changes. Both vendors adjust model lineups and rates several times a year. Review the rate card at every renewal and before every model upgrade.
• The OpenAI and Anthropic split. Anthropic's lead in new adopters (61% vs 51%) suggests the gap may widen. Companies with committed spend on one vendor should check whether teams are actually using it.
• Smaller vendors. xAI is at 10% adoption and growing about 0.6 points a month. A third vendor in the stack means a third invoice and a third set of keys.
• The ASU 2025-06 effective date. Annual periods beginning after December 15, 2027. Teams that capitalize AI implementation work should plan the transition now.

FAQs

Do most companies use both Anthropic and OpenAI?

Yes, based on Ramp Rate data. Anthropic's adoption is 80% and OpenAI's is 78% among businesses that buy AI models, so at least 58% paid both in the past 12 months. Some of that overlap is companies that switched vendors during the year.

How should finance budget for usage-based AI pricing?

Forecast tokens, not dollars. Multiply expected input and output tokens by each model's rate, add a buffer of 20 to 30% for model upgrades and new use cases, and re-forecast monthly against the vendor's usage dashboard.

Are AI model subscriptions capitalized or expensed?

API access and seat subscriptions are generally expensed as a service. Costs to implement a cloud computing arrangement, such as integration work, are evaluated under internal-use software guidance and may be capitalized. Confirm the treatment with your auditor.

How do you find shadow AI spend?

Search card and expense data for both vendors' merchant names, and include monthly charges under $50. Most shadow AI spend is individual seats bought on employee cards, which rarely pass through procurement or show up in AP.