For the first time since the AI race started in earnest, more American businesses are paying for Anthropic than for OpenAI. The May 2026 Ramp AI Index, compiled from transaction data across more than 50,000 US businesses and $100 billion in annual spend, shows Claude at 34.4% adoption — up 3.8% in April — while ChatGPT fell to 32.3%, down 2.9% in the same period.
This is a single data point from a single measurement methodology. It is also the most credible third-party spend-tracking dataset in the AI market, and the crossover it records is unambiguous.
How This Happened#
The trajectory is not subtle. Anthropic climbed from 0.03% of businesses in June 2023 to 7.94% by April 2025 — then rocketed to 34.4% by April 2026. That is a quadrupling in a single year. OpenAI’s business adoption grew 0.3% over the same period.
The driver is not Claude’s chat product. It is Claude Code.
Ramp’s analysis identifies Claude Code as the fastest-growing product in Anthropic’s history and the primary mechanism behind the adoption surge. That tracks with external signals: a separate analysis of public GitHub data published this month estimated that 4% of all global public commits are now being authored by Claude Code — double the percentage from just one month prior. For context, GitHub processed roughly 4 billion commits in 2025. Four percent of that is 160 million commits per year. One tool, one year, that kind of scale.
The growth compounds because Claude Code is a workflow tool, not a chatbot. When a team adopts Claude Code, usage accumulates through API calls rather than seat licenses. One engineer doing serious agentic work can consume between $500 and $2,000 per month in API costs. Multiply that across engineering orgs, and Anthropic’s revenue per customer is significantly higher than a traditional SaaS model where everyone pays the same monthly fee regardless of usage.
Anthropic’s annualized revenue run rate hit approximately $30 billion in early 2026, up from $9 billion at the end of 2025. OpenAI is tracking at $24-25 billion over the same period — a reversal from a lead that had seemed structural just eighteen months ago.
Three Threats the Data Surfaces#
Ramp doesn’t just track the crossover. It flags three structural risks that could unwind Anthropic’s position. Each deserves honest analysis.
Threat 1: The token incentive trap. Anthropic makes money when customers use more tokens. That creates a structural incentive to push customers toward expensive models and high-context workflows even when cheaper alternatives would suffice. Ramp frames this bluntly: “Anthropic profits from increased token consumption, creating pressure to push customers toward expensive models even when cheaper ones are sufficient.”
This is the underlying economics behind what Uber’s CTO described publicly: the company burned through its entire 2026 AI budget in four months, largely on Claude Code and Cursor. Individual engineers are reporting $500 to $2,000 per month in personal API costs for serious agentic workflows. At those numbers, CFOs start paying attention. And when CFOs start paying attention, they look for alternatives.
Threat 2: Reliability and cost shifts. Ramp’s data captured a period of user frustration — “frequent outages, rate limits, and increasing dissatisfaction with results.” Anthropic responded by resetting usage limits in April and securing additional compute capacity through the SpaceX Colossus deal (300MW, 220,000+ NVIDIA GPUs in Memphis). The rate limit reset helped. But the underlying compute constraint that created the reliability problems is a consequence of the 80x growth in Q1 2026 that Anthropic had only planned as 10x. That kind of demand mismatch doesn’t resolve cleanly.
A separate cost issue: recent model changes tripled token costs for image-inclusive prompts. That’s a significant jump in a category where usage is growing. Claude Code’s computer use features and the visual analysis capabilities of Opus 4.7 both involve image tokens. Developers building on those capabilities took an unexpected cost hit.
Threat 3: OpenAI Codex as cost-effective alternative. OpenAI’s Codex — the async agentic coding agent, not the legacy model — now covers substantial overlap with Claude Code’s core workflow at a lower per-task cost and with minimal switching friction. Ramp identifies inference platforms offering cheap, open-source alternatives as the fastest-growing competing category in their dataset. Codex isn’t open-source, but its pricing structure and the ease of migration via standard API patterns means that cost-sensitive teams have a credible exit path.
The switching cost from Claude Code to Codex is lower than it looks from the outside. Both tools operate via terminal, both support CLAUDE.md-style configuration, both integrate with GitHub. The moat is model quality, CLAUDE.md ecosystem depth, and the Managed Agents platform. If OpenAI closes the model quality gap on SWE-bench Pro (currently 58.6% GPT-5.5 vs 64.3% Opus 4.7), the Codex cost argument gets harder to dismiss.
What Actually Changes#
The Ramp crossover is symbolically significant and operationally real. “Most businesses paying for Anthropic vs OpenAI” means enterprise IT procurement conversations are now tilted differently than they were six months ago. When Anthropic walks into a 10,000-seat enterprise negotiation, it no longer needs to defend itself against “but everyone uses ChatGPT.” The data now says the opposite.
But the threats Ramp surfaces are also real, not hypothetical. Uber’s budget story will be repeated in CFO conversations at every large enterprise that has given engineers open-ended Claude Code access. The response from those CFOs won’t necessarily be “switch to a competitor” — it may be “governance and spend controls.” That’s exactly what Claude Cowork GA addresses (group spend limits, per-user caps, Analytics API for cost attribution). Anthropic has built the enterprise controls. The question is whether adoption of those controls keeps pace with the cost concerns they’re meant to address.
The deeper question is whether Claude Code’s architectural advantages are durable. The terminal-native, agent-owned model — where Claude Code has full environment access and owns the full development lifecycle from spec to deployment — is qualitatively different from IDE-embedded tools. But “qualitatively different” only maintains a price premium if users feel the difference in their outcomes, not just in their benchmarks.
The 4% of GitHub commits metric is the most direct signal available. At 160 million commits per year, something about the outcomes is working. The business adoption crossover confirms the enterprise is noticing.
The threats are real. The lead is real. The next quarter of Ramp data will be informative.
Sources:
- Ramp AI Index — May 2026 — Ramp
- Anthropic now has more business customers than OpenAI, according to Ramp data — TechCrunch
- Anthropic finally beat OpenAI in business AI adoption — but 3 big threats could erase its lead — VentureBeat
- Anthropic Passes OpenAI in Business Adoption: Ramp AI Index — Let’s Data Science
- Anthropic 34.4% Just Passed OpenAI — Ramp Flip May 2026 — ThePlanetTools.ai
- Anthropic vs OpenAI Business Adoption: What the Data Says About Enterprise AI — MindStudio