Anthropic Overtakes OpenAI for the First Time — The May 2026 "Changing of the Guard" in Enterprise AI, Explained for Non-Experts

"Wasn't ChatGPT supposed to be #1?" "I keep seeing Claude in the news lately — what changed?"—— This article walks through the Ramp AI Index May 2026 report (published 2026-05-13), Anthropic's $30B Series G ($380B valuation, 2026-02-12), the Claude Code $2.5B revenue milestone, and the Uber AI budget runaway, all from primary sources. The goal: help non-expert readers understand "what the changing of the guard in enterprise AI actually means, and how to apply it to your own work and AI tool choices." 2026 is shaping up to be the year the "ChatGPT-only" era ended.

Primary sources: Ramp AI Index May 2026 (2026-05-13), Anthropic Series G announcement ($30B / $380B valuation, 2026-02-12), TechCrunch coverage, VentureBeat analysis, Yahoo Finance (Uber reporting), OpenAI's enterprise next-phase post, and PwC 2026 AI Predictions. For the governance side of enterprise AI, see our Claude Compliance API + 28 integrations explainer. For the coding-tool layer, see Claude Code v2.1.149 + Gemini CLI v0.43 same-day release. For Google's counter-attack, see our Google I/O 2026 roundup.

1. May 2026: The "Changing of the Guard" in Enterprise AI

Ramp is a U.S. corporate-card and expense management company. Each month it publishes the Ramp AI Index, a public look at which AI vendors are getting paid by which companies. [Official] Per the May 13, 2026 issue, April 2026 enterprise AI adoption was 50.6% overall (+0.2pt). By vendor:

• Anthropic: 34.4% (+3.8pt MoM, ~4x over the past year)
• OpenAI: 32.3% (−2.9pt MoM, +0.3pt over the past year — essentially flat)
• Others (Google / Microsoft / Meta / inference services): remaining share

OpenAI had held the enterprise AI top spot continuously since ChatGPT launched in November 2022. Anthropic, founded in 2021 by former OpenAI researchers, had spent years in a "well respected but behind on share" position. This flip is the first time the order has reversed — and that is what the May 2026 Ramp AI Index records.

Ramp economist Ara Kharazian told TechCrunch that Anthropic took a deliberate strategy: "start with a very technical customer base, succeed in execution, and broaden from there". That is the opposite of OpenAI's consumer-marketing-first approach. The result is that "people who actually use AI every day" shifted preference, and the change shows up in share.

2. Why Did Anthropic Grow? — Three Concrete Reasons

Reason 1: Claude Code became a "hit product"

[Official] At Anthropic's Series G announcement (2026-02-12, $30B raised at $380B valuation), the company revealed that Claude Code reached a $2.5B annualized run rate — about ¥375B at current rates. The product went GA in May 2025 and hit that figure in nine months. Anthropic positions it as "the fastest-growing product in company history."

Claude Code is a terminal-based AI tool for developers. It reads, writes, edits, and tests code on your behalf. Per Anthropic's internal analysis (Feb 2026), Claude Code authored approximately 4% of all public GitHub commits worldwide.

Reason 2: Model quality on hard tasks

Anthropic's flagship Claude Opus 4.7 and mid-tier Claude Sonnet 4.6have, since late 2025 into early 2026, matched or beaten OpenAI's GPT-5.5 family on hard, agentic tasks:

• SWE-bench (real software engineering tasks): Claude Opus 4.7 hit 90.2% — #1 as of January 2026
• Long-document reasoning: 200K context window vs OpenAI's standard 128K
• Agentic workflows (multi-step plan + act loops): Claude consistently ahead in reported benchmarks

Developers, data scientists, and consultants weight this category heavily — they care about "does the model get it right when I'm actually stuck?" — so Claude getting picked as the "reliable teammate" snowballed.

Reason 3: Enterprise-readiness caught up

Through late 2025 and 2026, Anthropic shipped the boring-but-decisive infrastructure that IT departments demand. [Official] On 2026-05-21, Anthropic launched the Claude Compliance API plus 28 security partner integrations (Cloudflare, Wiz, Palo Alto, Microsoft Purview, Okta, Datadog, ...), wiring Claude usage into the same pipes IT teams already use for every other SaaS (details: Claude Compliance API explainer).

[Author's view] "Boring enterprise plumbing" is not glamorous — but it is the layer that actually decides share fights in the long run. Anthropic's Compliance API ("put Claude on the same management pipe as every other SaaS") was effectively the catch-up move on OpenAI's 2024-era ChatGPT Enterprise lead — and the catch-up is now done.

3. The Claude Code Boom and the Uber Budget Runaway

Claude Code revenue trajectory

[Official] Per Anthropic's Series G announcement (2026-02-12), Claude Code grew like this:

• May 2025: General availability (GA)
• November 2025: $1B annualized run rate (about 6 months from GA)
• February 2026: $2.5B annualized run rate (another 3 months → 2.5x)
• May 2026: Weekly active users doubled vs. January 2026

Enterprise Claude Code subscriptions quadrupled since the start of 2026; named customers include Netflix, Spotify, KPMG, L'Oréal, and Salesforce. [Official] Anthropic's overall run rate climbed from $9B at end-2025 to $30B by April 2026 — 3.3x in four months, an unprecedented curve for enterprise software.

The Uber AI budget runaway — when the story hit the boardroom

But there is a flip side. [Official] In April 2026, Uber CTO Praveen Neppalli Naga confirmed (via The Information and Yahoo Finance) that the company had burned through its entire 2026 AI budget in four months.

• Uber rolled out Claude Code to roughly 5,000 engineers in December 2025
• By February 2026 32% of engineers were active users; by March, 84% were classified as "agentic users"
• Per-engineer API cost: $500–$2,000 per month (about ¥75,000–¥300,000)
• Result: 2026 AI budget depleted by April
• By spring, 95% of engineers used AI tools monthly; 70% of code originated from AI

[Author's view] The Uber story really says "Claude Code is so good it is being adopted faster than the org can absorb the spend." $500–$2,000 per engineer per month, in Japanese terms, is "one engineer's cloud bill = 10–20% of a senior engineer's monthly salary." It is worth paying for if productivity moves — but without per-project visibility and budget guardrails, the budget is guaranteed to evaporate.

4. This Is Not a Victory Declaration — Three Risks Anthropic Still Carries

Risk 1: Cost structure — "the more it gets used, the more it costs Anthropic too"

VentureBeat's 2026-05-14 analysis points out that Anthropic's biggest risk is a structural mismatch between customer token spend and Anthropic's own GPU + power cost. Claude Code's agentic design uses a lot of tokens per task — as customers use more, Anthropic's infrastructure cost rises proportionally.

• Anthropic needs massive compute investment ahead (the $30B Series G is partial cover)
• Customers are hitting cost shocks (Uber), and price changes are inevitable
• Either raise prices (lose customers) or absorb (compress own margin) — a structural squeeze

Risk 2: Reliability — outages and rate-limit incidents

[Official] Anthropic published a postmortem on a major infrastructure incident in May 2026 (see coding AI ops phase roundup), candidly acknowledging that three concurrent bugs affected users for roughly four weeks. That candor is admirable, but it did surface real "Anthropic reliability is not yet at OpenAI's bar" sentiment among heavy users.

Risk 3: Cheaper models — open source and OpenAI Codex push from below

Google's Gemini 3.5 Flash and OpenAI's Codex increasingly deliver comparable performance at roughly half the cost. Open-source families (Meta Llama, DeepSeek) accessed via inference services are also reaching production-grade quality. Anyone operating under "good-enough quality, low cost" pressure may keep moving away from Anthropic.

5. 79% Pay Both: The Multi-Model Strategy

Why pay both?

VentureBeat's "VB Pulse Foundation Models tracker" (May 2026) estimates 79% of enterprises pay both OpenAI and Anthropic. Simple reason: their strengths diverge.

• Coding / hard reasoning: Claude Opus / Sonnet currently lead (SWE-bench 90.2%)
• Multimodal (image / audio / video): GPT-5.5 family and Gemini 3.5 Flash lead
• Very long context (200K–1M): Claude and Gemini lead
• Conversational naturalness: GPT-5.5 and Claude Sonnet are roughly even
• Cost efficiency (high-volume, low-unit-cost): Gemini 3.5 Flash and OSS lead

Common implementation patterns

In real enterprises, three patterns have settled in:

1. Task-based routing: route by inferred request type — "Claude for code, GPT for image gen, Gemini for internal search." Built via LangChain / LlamaIndex etc.
2. Cost-tier routing: "paying users get Claude Opus, free-tier users get Gemini 3.5 Flash" — switch cost / quality based on user tier.
3. Failover: "Claude API slow → auto-fall to GPT-5.5." Spread rapidly after the Anthropic outage.

Can individuals and SMBs implement this?

Yes — multi-model routing is not only for the Fortune 500. Concrete options:

• OpenRouter / LiteLLM: API gateways that present multiple vendors under one interface
• Continue.dev / Aider / Cline: editor extensions where you switch models by config
• Roll your own: even a rough classifier ("light = Haiku, heavy = Opus") buys real savings

6. Risks and Cautions — What Japanese Companies Should Watch Now

Watch 1: The risk of single-vendor lock-in

Because rankings can flip in nine months, "3-year contract, Anthropic only" or "ChatGPT Enterprise for the whole company" are high-risk bets. Through late 2026 and into 2027, Google and OpenAI counter-attacks are likely; no one can guarantee today's leader is next year's leader.

• Insist on 12-month review clauses in every contract
• Abstract the API layer so vendor swaps are cheap, not a rewrite
• Internally communicate "tool choice is fluid," so employees do not over-attach to one vendor

Watch 2: Wire up cost visibility

The Uber lesson is that Claude Code is powerful enough that the budget will melt without visibility. To avoid the same trap:

• Turn on /usage breakdown (Claude Code v2.1.149) on day one
• Set hard limits by team / project / month
• Measure ROI in shipping speed / incident rate / labor cost saved, not "lines of code"

Watch 3: Regulation and domestic SIer moves

[Official] In May 2025 Japan passed the "Act on Promotion of R&D and Application of AI Technologies" (AI Promotion Act), signaling a government posture of "promote use while managing risk."[Speculation] Through 2026, NTT Data / Fujitsu / NEC and other Japanese SIers are likely to announce partnership programs with Anthropic / OpenAI (only limited public detail as of 2026-05-26).

• Re-check AI Promotion Act / PIPA / Specified Commercial Transactions Act updates twice a year
• Track domestic cloud / SIer (AWS / Google Cloud / Azure / NTT-line) announcements continuously
• Sector-specific guidance (FSA / MHLW / METI) — also worth a quarterly read

7. Sales Claw View — Implementing "Multi-Model" at SMB / Indie Scale

The pattern Sales Claw uses

In Sales Claw v0.5, watching enterprise share flip drove a decision to refuse single-vendor design and rebuild around multi-model routing from the start. Concretely:

1. Hard architecture calls / complex refactors → Claude Opus 4.7 (via Claude Code)
2. Bulk form-submission logic and tests → Claude Sonnet 4.6 (Opus quality, half the price)
3. Sales-copy drafts / A/B variants → GPT-5.5 (best conversational naturalness)
4. Log analysis / large-file reads → Gemini 3.5 Flash (1M context = read everything in one shot)

We rewrote this split three times in three months before landing on a stable version that runs at about ¥12,500 / month (mostly Claude Code) + ¥5,000 for the restat indie-developer scale. After 14 days of hands-on use, the lesson stuck: bolting multi-model routing on later costs 10x more than designing it in from day one.

"Policy-controlled autonomy" meets the changing of the guard

Sales Claw's other axis is "autonomy without per-send human approval". That ties into multi-model strategy tightly:

• No per-send approval means pre-send automatic inspection is everything
• Inspection quality depends on the reasoning power of the model you use
• So single-model failure = inspection failure — a real operational risk
• The fix is cross-checking with multiple models — Sales Claw v0.5 doubles sales-NG detection across Claude Opus + GPT-5.5

今すぐ Sales Claw で、営業をもっとスマートに。

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub

8. Conclusion — Turning the "Changing of the Guard" into Daily Decisions

The May 2026 Ramp AI Index recorded the first changing of the guard in enterprise AI. The more important fact, though, is not "Anthropic won" but "the market has left the fixed-leader phase and entered a dynamic-leader phase." The ChatGPT-only era is over; Anthropic, OpenAI, and Google will trade the lead back and forth.

Non-experts, SMBs, and indie developers should take away three things:

1. Refuse single-vendor lock-in: insist on 12-month review clauses, design with an API abstraction layer from day one.
2. Wire cost visibility into the deploy: Uber is not somebody else's problem. Turn on /usage, hard limits by team, and real ROI metrics from the first day.
3. Implement multi-model at your scale: this is not enterprise-only. OpenRouter / LiteLLM / Continue.dev let an individual run it for ¥10,000–¥30,000 a month.

2026 will likely be remembered as the year "single-vendor AI dominance ended". As a Sales Claw maintainer, the design principle I will keep stating publicly is: "Assume the leader will change; design from day one to keep multiple options open."

Japanese-language original: Anthropic が OpenAI を初めて逆転 (2026-05-26)