explainx / blog
Ramp's May 2026 AI Index shows Anthropic at 34.4% of U.S. businesses paying for AI vs OpenAI's 32.3% — the first crossover. What the data measures, why Claude Code drove it, and what loop engineering has to do with Amodei's "engineers don't write code" quote.
Jun 11, 2026
At Code with Claude Tokyo, Anthropic shipped scheduled deployments and environment-variable vaults in public beta for Claude Managed Agents, plus general availability for dynamic workflows in Claude Code—agents that run on a schedule, use tools securely, and tackle bigger parallel jobs.
Jun 3, 2026
In a stunning reversal, Sam Altman admits he was 'pretty wrong' about AI wiping out entry-level jobs. Dario Amodei quietly shifts from '50% of white-collar jobs eliminated' to focusing on augmentation. The timing? Both companies eyeing IPOs as enterprise AI costs spiral out of control.
Jun 29, 2026
Most teams conflate prompt writing with context design, loop orchestration, and harness code. They are four layers of the same stack. Here is how they nest, what breaks when you skip one, and which layer to fix when agents fail.
For years the enterprise AI conversation assumed OpenAI was the default vendor. ChatGPT Enterprise, API keys, Copilot-era habits — OpenAI was the name on the corporate card.
That assumption broke in April 2026, according to Ramp's AI Index: Anthropic passed OpenAI in U.S. business adoption for the first time — 34.4% of businesses on Ramp paying for Anthropic tools versus 32.3% for OpenAI. Overall paid AI adoption in the dataset hit 50.6%, meaning half of Ramp-tracked companies now pay for at least one AI product.
Within days, the data collided with a separate viral thread: Dario Amodei describing Anthropic engineers who barely write code anymore, Boris Cherny-adjacent claims that Claude authors 80%+ of production code at Anthropic, and a flood of loop engineering posts arguing the shift from prompting to autonomous verify-and-retry cycles is the real story.
This guide separates what Ramp actually measured from what X inferred, why Claude Code likely drove the crossover, and what engineering and finance leaders should do with the signal.
| Question | Answer |
|---|---|
| What flipped? | Anthropic 34.4% vs OpenAI 32.3% business adoption share (April 2026, Ramp) |
| Adoption or spend? | Adoption — % of Ramp businesses paying each vendor, not total ARR |
| Geography? | U.S.-skewed Ramp customer base (~50k+ companies) |
| Anthropic 1-year change? | ~4× adoption share; OpenAI ~+0.3% |
| Why now? | Claude Code + workflow/agent tools vs chat-only seats |
| Amodei quote about? | Review-and-guide loops inside Anthropic, not "no humans" |
| 80% code claim? | Boris Cherny / Claude Code team — production code at Anthropic with agent loops |
| Link to loop engineering? | Enterprises buy Claude Code when they adopt verify-and-retry dev workflows |
| Should I switch vendors? | No — run your own evals; many teams use both |
| Related Ramp data? | Token spend up 13× since Jan 2025 — see token governance guide |
Before interpreting the crossover, understand the metric — most hot takes skip this.
Ramp is a corporate spend platform. The AI Index counts corporate card and invoice payments to AI providers across businesses that use Ramp. Each month Ramp reports:
From Ramp's May 2026 release:
Adoption of Anthropic rose 3.8% in April to 34.4% of businesses. OpenAI adoption fell 2.9% to 32.3%. Overall AI adoption rose 0.2 percentage points to 50.6%.
Over the last year, Anthropic has quadrupled business adoption while OpenAI grew business adoption by only 0.3%.
What that means: More distinct companies on Ramp started paying Anthropic bills — not necessarily that Anthropic suddenly owns most AI dollars worldwide. A company adding a $20/month Claude Pro seat counts the same as one adding a six-figure API contract for adoption share.
Why the data still matters: Payment adoption is a leading indicator for vendor default status. Finance teams approve vendors before engineers scale usage. When Anthropic crosses OpenAI on "who has a corporate relationship," procurement and security reviews follow.
For the spend intensity side — token bills, 13× growth, lumpy months — see explainx.ai's Ramp token spend and governance guide.
| Metric | Anthropic | OpenAI | Notes |
|---|---|---|---|
| Business adoption (Apr 2026) | 34.4% | 32.3% | First time Anthropic leads in Ramp index |
| YoY adoption change | ~4× | ~+0.3% | Ramp stated figures |
| Overall AI-paying businesses | — | — | 50.6% of Ramp panel |
| Primary driver (analyst read) | Claude Code, long-context workflows | Chat/API incumbency | BI, Decoder, LinkedIn summaries align |
OpenAI is not collapsing — one in three Ramp businesses still pays OpenAI. The story is share shift, not extinction. Anthropic went from challenger to plurality leader in this specific dataset.
Downstream reporting in May–June 2026 also cited internal urgency at OpenAI (enterprise leadership changes, "code red" narratives in trade press) — treat those as unverified culture signals, not data. The defensible public fact is Ramp's panel.
Ramp's own economists and coverage in Business Insider converge on a simple mechanism:
Enterprises stopped buying only chat. They started buying tools embedded in work.
| Chat-era purchase | Workflow-era purchase |
|---|---|
| ChatGPT Enterprise seats | Claude Code + API for agentic dev |
| Generic copilot add-ons | Long-document analysis for legal/finance |
| One-shot drafting | Integrated loops with internal systems |
Claude Code is the clearest product bridge: a paid, terminal-native coding agent with MCP, hooks, subagents, and /loop — exactly the kind of line item that shows up on a Ramp corporate card when eng leadership standardizes on Anthropic.
That aligns with explainx.ai's Anthropic Economic Index cadences data: Claude usage follows work rhythms (evening meal planning, tax-season spikes) — not random chat. Businesses pay for tools that slot into schedules and pipelines.
The X thread that amplified Ramp's index was not about finance — it was about how Anthropic builds software.
Circulated clips from late June 2026 quote Amodei roughly as:
I have engineers within Anthropic who don't write any code — they let Claude write the code and they edit it and look it over.
At Anthropic, writing code means designing the next version of Claude itself — so Claude is already helping design the next Claude.
explainx.ai read: This is role reframing, not headcount elimination. The engineer's job shifts from typing to specifying, reviewing, and designing verification — the same shift loop engineering describes at industry scale.
Cherny has stated publicly that Claude authors 80%+ of production code at Anthropic, but only after the team moved from reviewing every response to building loops with programmatic checks — tests, lint, eval gates. Humans design the harness; the loop executes.
That claim is the operational backbone of loop engineering going mainstream in June 2026 and the Claude Code implementation guide.
Posts also cited an Anthropic research lead describing hundreds of agents collaborating with "close the loop — let the AI check its own work." Treat specific agent counts as illustrative unless published in a paper or official talk transcript. The pattern — multi-agent loops with self-verification — matches Anthropic's public agentic architecture messaging and multi-agent orchestration production patterns.
June 2026 X was full of:
That is not coincidence. Loop engineering names what enterprises pay for when they buy Claude Code at scale:
Goal → [Plan → Act → Observe → Verify] → ... → Done
| One-shot prompting | Loop engineering |
|---|---|
| Human reviews every diff | Harness checks tests/lint/eval |
| Chat session ends at reply | Session ends at verifiable success |
| Low token burn per task | Higher burn, higher completion rate |
AI companies' incentives align: agents and loops consume far more tokens than chat — see token economics. Enterprises still buy because ** shipped software** beats drafted paragraphs.
Ramp's adoption crossover says: CFOs now approve Anthropic relationships at higher rates than before. Loop engineering explains why eng teams asked for those approvals.
| Claim on X | Safer interpretation |
|---|---|
| "Anthropic won enterprise AI" | Won adoption share in Ramp U.S. panel |
| "OpenAI is losing" | Flat YoY adoption (+0.3%); still 32%+ of businesses |
| "Everyone will use Claude Code" | Strong in software verticals; legal/finance/chat still mixed |
| "Engineers obsolete" | Role change — review, harness design, accountability remain |
| "80% code = no bugs" | Volume metric, not quality; loops need good gates |
Global market: Ramp does not represent China, EU sovereign cloud, or Google/Gemini-heavy shops.
Dual-vendor reality: Many companies pay both OpenAI and Anthropic — adoption shares are not mutually exclusive per company in a multi-tool stack.
If half your peers pay for AI tools (50.6% in Ramp's panel), the question is no longer if but which workflows. Chat pilots → agent pilots with explicit success criteria.
See AI for business leaders for decision framing.
Amodei and Cherny's message is not "fire developers." It is invest in checks — CI, evals, human gates on irreversible actions. See human-in-the-loop AI.
Ramp's other 2026 headline: token spend 13× in a year. Adoption without cost attribution is how teams burn 2026 budgets by Q2 — Uber's CTO cited exhausting a year AI budget in four months in downstream commentary on the same index.
| Use case | Often strong on | Evaluate with |
|---|---|---|
| Agentic coding | Claude Code, Codex | Your repo, tests, security rules |
| General chat + plugins | ChatGPT ecosystem | Your non-dev workflows |
| Long-document review | Claude context | Your contract/RFP samples |
| Multimodal product features | Gemini, GPT | Your UX requirements |
Compare harnesses: Codex vs Claude Code.
Not every employee needs Claude Code. Every software team shipping agent-assisted code needs loop design literacy — Building AI Agents pathway, Loop Engineering pathway.
Neither directly. Ramp counts businesses with a payment relationship to each vendor. A small Claude Team plan and a large API invoice both count once.
OpenAI Codex usage limit resets made headlines the same week (June 2026) — separate from Ramp's April adoption snapshot. Possible sentiment overlap for developers, but do not conflate a May/June product incident with April billing adoption without evidence.
Anthropic's export-control and government friction (Mythos/Fable restrictions, supply-chain rhetoric) ran parallel to commercial growth in some May summaries — correlation, not proven causation. Enterprise buyers weigh compliance separately; see Fable 5 availability for status.
| Ramp release | Measures |
|---|---|
| AI Index (this story) | Who pays which vendor (% of businesses) |
| Token spend posts | How much inference costs grew (13× avg monthly) |
You can adopt Anthropic and blow your budget. Adoption ≠ affordability.
If your company standardizes on Claude Code, loop and harness skills beat prompt trivia. If not, vendor-agnostic agent architecture (four-layer stack) future-proofs you.
Ramp's May 2026 AI Index reported a historic first: Anthropic at 34.4% vs OpenAI at 32.3% of U.S. businesses on Ramp paying for AI tools in April 2026, with Anthropic quadrupling adoption in a year while OpenAI grew 0.3%.
That is an adoption-share signal from payment data, not a global revenue coronation — but it is a real shift in default vendor relationships, driven heavily by workflow and coding agents (Claude Code) rather than chat alone.
The viral Amodei and loop engineering thread explains how Anthropic's own team builds — review-heavy, loop-driven, Claude-assisted — which matches what paying enterprises are trying to replicate.
Do not switch vendors because of a leaderboard. Do treat agentic workflows, verification gates, and token governance as 2026 baseline infrastructure.
/loop, verification gates, production patternsPrimary source: Ramp AI Index — May 2026
Adoption figures and CEO quotes reflect public reporting and social clips as of June 29, 2026. Ramp methodology may revise; verify primary sources before procurement decisions.