For the first time, the company that built Claude is taking more enterprise dollars than the company that built ChatGPT. According to the Ramp AI Index, Anthropic now reaches 34.4 percent of companies paying through Ramp, edging past OpenAI at 32.3 percent. That’s a coin-flip difference on paper. In practice, it’s the loudest signal yet that enterprise AI buying is being decided by coding agents, not chat windows — and that the leaderboard can flip again before the next earnings call.
What the Ramp AI Index Actually Measures
The Ramp AI Index pulls from spending data across companies paying through Ramp by corporate card or invoice, and it tracks what share of those companies pay which AI vendor — not how much they use or how much value they get. Per Ramp, Anthropic quadrupled its penetration over the past year while OpenAI grew just 0.3 percent, with the crossover landing in April 2026. Ramp also flags that its dataset skews heavily toward US companies, so this is not a global market read.
Why it matters: a spending-share metric is a procurement signal, not a usage signal. It tells you which vendor a finance team has approved an invoice for, which is exactly the data point investors and competitors care about when sizing the durable revenue base. If you’re an enterprise buyer benchmarking your own vendor mix against peers, this is the closest public proxy you’ll get without scraping LinkedIn job posts. The take: treat the 34.4 vs 32.3 split as a tie that happens to favor Anthropic — the real story is the slope of the curve, not the current point on it.
Why Coding Agents Are Driving the Shift
Anthropic’s growth has been pulled along by Claude Code and agentic developer tooling, where models don’t just answer questions — they call tools, write files, and run shell commands inside a loop. Ramp economist Ara Kharazian notes that agentic workflows can push token counts higher because the model generates more text, invokes more tools, and writes or runs code as part of a single task. A workflow that looked identical a year ago can burn through far more tokens today, even with flat user demand.
Developers feel this first. If you’re a platform team that swapped a templated code-review bot for a Claude-Code-style agent that reads the repo, runs tests, and proposes patches, your monthly invoice is going to look unrecognizable — even if your headcount and ticket volume haven’t moved. That’s also why the tradeoff between agents and traditional automation has become a real budgeting decision rather than a marketing diagram. The prediction: by the end of 2026, every serious engineering org will have a per-developer monthly token budget the same way they have a per-developer seat license, because the alternative is the Uber scenario.
The Three Headwinds That Could Reverse This in a Quarter
Kharazian flags three specific risks to Anthropic’s lead, and each one is concrete enough to plan around. First, Anthropic earns more when customers run its expensive frontier models, which creates pressure on buyers to substitute cheaper options — Uber’s CTO has already said the company blew through its 2026 AI budget, per The Information. Second, users have been openly complaining about Claude outages and declining quality, which Anthropic has publicly acknowledged. Third, the Opus 4.7 model triples the cost of image processing compared to its predecessor.
Why it matters: enterprise stickiness in this category is unusually weak. “We have never seen a software industry as dynamic, where newcomers can disrupt market leaders in a matter of months, and where the pace of development overrides the typical forces of vendor stickiness,” Kharazian writes. If you’re running a fintech platform where every basis point of inference cost shows up in unit economics — the kind of constraint familiar to teams building audit-grade financial systems — a 3x bump on image tokens isn’t a line item, it’s a re-architecture trigger. The take: the next twelve months will reward whoever ships a credible router that swaps Claude, GPT-5.5, and an open-source model per request without breaking agent behavior.
Price Hikes Are Distorting the Scoreboard
Because Ramp tracks dollars rather than tokens, the index inflates whenever vendors raise prices. And that’s been happening on both sides. Anthropic’s Opus 4.7 costs more than 4.6 despite flat headline pricing, according to first-token analyses cited in the source. OpenAI hiked GPT-5.5 prices by 49 to 92 percent compared to its predecessor depending on input length. Meanwhile, OpenAI’s Codex handles similar coding tasks at a lower price point, and cheaper inference platforms running open-source models are gaining ground, Kharazian says.
If you’re a CFO trying to read this index as a market share chart, you’re going to misread it. Two vendors raising prices in the same quarter can both look like they’re “winning” even if real consumption is shifting to a third option that doesn’t show up in the dataset at all — typically self-hosted Llama or DeepSeek deployments billed through cloud infrastructure rather than as an AI line item. The prediction: within two Ramp Index releases, we’ll see a visible drag on both Anthropic and OpenAI shares as more teams formalize a custom-versus-SaaS decision and route commodity workloads to open models.
FAQ
Q: Does Anthropic actually have more enterprise users than OpenAI now? A: Not necessarily. The Ramp AI Index measures the share of Ramp-billed companies paying each vendor — 34.4 percent for Anthropic versus 32.3 percent for OpenAI as of April 2026. It does not measure seat count, token volume, or revenue, and Ramp’s data skews toward US companies.
Q: Why did Anthropic grow so much faster than OpenAI in this dataset? A: Anthropic quadrupled its B2B penetration over the past year while OpenAI grew 0.3 percent, per Ramp. The most cited driver is enterprise adoption of Claude Code and agentic coding workflows, which generate higher token volumes per task and pull in entire engineering teams as paid users.
Q: Will OpenAI take the lead back? A: Plausibly. OpenAI’s Codex offers similar coding capability at a lower price point, GPT-5.5 covers a wide cost range, and Anthropic’s Opus 4.7 pricing for images is three times higher than before. Switching costs in agentic tooling are low enough that a single budget overrun can trigger a vendor change.
Key Takeaways
- Treat the Ramp index as a procurement signal, not a usage signal — buyers and analysts who confuse the two will misprice both vendors.
- Set a per-developer monthly token budget before your first Claude Code or Codex rollout, because the Uber overrun scenario is now the base case, not the edge case.
- Build a model-routing layer into any agent system you ship in 2026 — single-vendor lock-in is the most expensive architectural decision available right now.
- Watch Opus 4.7 image pricing and GPT-5.5 input-length pricing as the two specific levers that could flip the index again within a quarter.
- Commodity workloads — summarization, classification, internal search — belong on cheaper inference or open-source models; reserve frontier spend for agents that actually justify the per-task cost through autonomous work, the kind of split that decides whether AI automation programs pay back inside a fiscal year.