Google Opens I/O as Anthropic Surpasses OpenAI in Revenue and Targets a $900 Billion Valuation
Anthropic hit $30B ARR in April — passing OpenAI's $25B for the first time — and is closing a $30B funding round at $900B with a potential October IPO; Google's answer opens at Shoreline Amphitheatre this morning as Meta prepares 8,000 layoffs for tomorrow.
The story
The last week crystallized a reversal that felt theoretical six months ago: Anthropic surpassed OpenAI in annual revenue run rate. Dario Amodei's company hit $30 billion ARR in April — against OpenAI's $25 billion — driven by Claude Opus 4.7's coding gains and Claude Code deployments into enterprises like Stripe's 1,370 engineers. Anthropic is now in early talks to raise $30 billion at a $900 billion valuation, nearly tripling its February mark of $380 billion, with Dragoneer, Greenoaks, Sequoia, and Altimeter as co-leads; a potential October IPO is being discussed internally. Adding noise to the week: Mythos, Anthropic's restricted cybersecurity model capable of 90x Opus 4.6's exploit performance, was accessed by unauthorized users on its launch day — a guessed endpoint URL combined with Mercor data-breach credentials — exposing the model to the open internet for hours before Anthropic contained it. As of this morning, Sundar Pichai is on stage at Google I/O 2026. The question for the next 72 hours is whether Gemini 3.2 Flash and a potential Gemini Omni announcement can reframe the story before the Anthropic revenue-rank narrative cements into a consensus.
Who moved this cycle
| Player | What changed | Consequence |
|---|---|---|
| Anthropic | Hit $30B ARR in April, passing OpenAI · In talks to raise $30B at $900B valuation (Bloomberg, May 12), Dragoneer/Greenoaks/Sequoia/Altimeter as co-leads, October IPO being discussed · Mythos security incident: unauthorized users accessed restricted cybersecurity model on launch day via guessed endpoint + Mercor data-breach credentials | First time Anthropic has led the revenue race; $900B valuation exceeds OpenAI's $852B March mark and would set the pre-IPO benchmark for the entire AI lab sector heading into a potential public listing |
| I/O 2026 keynote opening today (May 19, 10AM PT) at Shoreline Amphitheatre · Gemini 3.2 Flash appeared in iOS app May 5 at rumored $0.25/$2.00 per 1M tokens · Gemini Intelligence OS-level agent announced for Android 17 · Android XR glasses previewed · Gemini Omni (unified text+image+video pipeline) rumored for today's keynote | If Gemini 3.2 Flash and Omni ship as rumored, Google competes simultaneously on cost (cheapest Flash tier) and multimodal capability — a combination no other lab currently offers at scale | |
| Meta | 8,000 layoffs beginning May 20, canceling 6,000 open roles · $115-145B capex guidance for 2026 (up from $72.2B in 2025) · Q1 revenue $56.3B, net income $26.8B · Llama 4 Maverick (MoE architecture, natively multimodal open weights) shipping | Meta is trading headcount for compute at a scale no other open-source provider can match; Llama 4's open weights backed by $145B infrastructure pressures every model vendor on cost |
| xAI | Launched Grok Build coding agent CLI on May 15 (early beta, SuperGrok Heavy $299/month standard, $99/month intro) · Grok 4.3 backend, 2M token context, 8 concurrent agents · Apollo Global Management and Morgan Stanley began testing Grok internally | xAI is 12+ months behind Anthropic and OpenAI on coding agents, but Wall Street firms with direct Musk relationships give Grok a financial-services wedge neither Claude Code nor OpenAI Codex has fully established |
| OpenAI | GPT-5.5 (released April 23) at $5/$30 per 1M tokens — double GPT-5.4's API price · 1.5M enterprise clients · OpenAI Deployment Company raised $4B from 19 investors including Bain & Company and Capgemini (May 11-12) · Revenue chief Katie Dresser: enterprise adoption 'at a tipping point' (CNBC, May 11) | First time OpenAI has lost the revenue-rank to a competitor; the Deployment Company mirrors Anthropic's enterprise JV and signals both labs see systems-integrator leverage as the next enterprise monetization layer |
| DeepSeek | V4-Pro and V4-Flash released April 23-24, MIT license with downloadable weights · V4-Flash at $0.14/1M input tokens (cheapest flagship-class model available) · V4-Pro: 80.6% SWE-bench Verified, 52 on Artificial Analysis Intelligence Index | MIT-licensed weights at sub-$0.15 input pricing puts sustained margin pressure on every proprietary mid-tier model; developers can self-host V4-Flash for near-zero marginal cost on owned hardware |
| Alibaba / Qwen | Qwen consumer app passed 300M monthly active users · Cloud AI revenue: triple-digit growth for 11th consecutive quarter (company-announced) · Merging Qwen into Taobao/Tmall for conversational shopping across 4B SKUs (announced May 11) · FAW Group Hongqi HS6 integrating Qwen in-car system | 300M MAU makes Qwen the largest consumer AI assistant outside the US; Taobao integration converts an AI chat product into a commerce transaction layer — a monetization model none of the US labs have replicated at scale |
New models on the board
| Player | Model & what changed | Pricing |
|---|---|---|
| Anthropic | Claude Opus 4.7 10.9-point SWE-bench Pro gain over Opus 4.6 (53.4% → 64.3%); 13% better on Cursor's internal 93-task benchmark per CEO Michael Truell; new tokenizer consumes ~35% more tokens for the same text — effectively a ~35% cost increase despite nominally unchanged pricing | $5/$25 per 1M tokens (same list price as Opus 4.6; effectively higher due to tokenizer) |
| OpenAI | GPT-5.5 Released April 23; strongest agentic coding and computer-use performance in OpenAI lineup; 1M token context window; 2x GPT-5.4's API price — Simon Willison flagged as the largest single-model price jump in his tracking | $5/$30 per 1M tokens (standard); $30/$180 per 1M tokens (Pro tier) |
| DeepSeek | DeepSeek V4-Flash / V4-Pro MIT license, downloadable weights; V4-Pro: 80.6% SWE-bench Verified, 52 on Artificial Analysis Intelligence Index; V4-Flash is the cheapest flagship-class model currently available | V4-Flash: $0.14/1M input; V4-Pro pricing not yet formally published |
| Gemini 3.2 Flash (rumored at I/O today) Appeared in iOS Gemini app and Google AI Studio on May 5 ahead of I/O; native multimodal across text, image, audio, video; near-Gemini 3.1 Pro performance on coding tasks at Flash-tier pricing; I/O announcement expected today | $0.25/$2.00 per 1M tokens (rumored; unconfirmed pre-keynote) |
Market signals
| Player | Type | Value | Context |
|---|---|---|---|
| Anthropic | Revenue | $30B ARR (April 2026) | Up from $9B at end of 2025, $14B in February, $19B in March; passed OpenAI's $25B ARR; Dario Amodei called growth 8x ahead of internal forecasts (VentureBeat) |
| Anthropic | Funding | $30B round at $900B valuation (in talks as of May 12) | Dragoneer, Greenoaks, Sequoia, Altimeter as co-leads; would nearly triple February's $380B valuation; potential October IPO being explored (Bloomberg, The Information) |
| OpenAI | Partnership | OpenAI Deployment Company raising $4B from 19 investors | Bain & Company and Capgemini invested May 11-12; $10B entity valuation; targets large-scale enterprise AI deployment via systems integrators globally (company-announced) |
| Meta | Hiring | 8,000 layoffs beginning May 20; 6,000 open roles canceled | $145B 2026 AI capex up from $72.2B in 2025; cuts are structural, not performance-based — reorganizing into AI-focused pods; AI researchers receiving pay increases while median total comp fell to $388K (Axios, The Next Web) |
| Anthropic | Customer | Stripe deployed Claude Code to 1,370 engineers via enterprise MDM binary | Zero-configuration deployment; limited data retention, no model-training feedback; signals enterprise-grade security posture (company-announced at Code with Claude 2026, May 6) |
| Alibaba / Qwen | Revenue | Triple-digit cloud AI revenue growth, 11th consecutive quarter (company-announced) | 300M monthly active users on consumer Qwen app; Taobao/Tmall conversational shopping integration across 4B SKUs announced May 11 |
| OpenAI | Revenue | $25B ARR (Q1 2026) | Enterprise now 40%+ of total revenue; 1.5M business clients globally; 340% YoY enterprise growth by OpenAI's own accounting (company-announced) |
What builders are saying
Simon Willison (simonwillison.net). Flagged two back-to-back pricing shocks: Opus 4.7's new tokenizer effectively raises costs ~35% for the same text compared to Opus 4.6, and GPT-5.5 is priced at double GPT-5.4 over the API — 'the most significant price hikes since I started tracking model pricing.'
Michael Truell, CEO of Cursor. Opus 4.7 lifted resolution by 13% over Opus 4.6 on Cursor's internal 93-task benchmark, solving four tasks that neither Opus 4.6 nor Sonnet 4.6 could touch — corroborating Anthropic's SWE-bench claims in a real production coding environment.
Katie Dresser, OpenAI revenue chief (CNBC, May 11). Enterprise AI adoption is 'at a tipping point' — OpenAI now counts 1.5 million business clients globally, more than 1 million of them outside the US, with 340% year-over-year enterprise growth.
CIO Dive. Grok Build enters the coding-agent market 12+ months after Anthropic's Claude Code and OpenAI's Codex, but notes xAI's direct Wall Street relationships — Apollo and Morgan Stanley testing internally — may give it a financial-services wedge neither rival has fully established.
Threads worth pulling
Anthropic $30B ARR → $900B valuation round → October IPO → first AI lab public listing sets sector P/S multiple An Anthropic IPO in October would be the first major AI lab to trade publicly, giving the market a real-money P/S benchmark. If Anthropic lists at $900B on $30-45B ARR, that implies 20-30x revenue for the sector — a multiple that either makes OpenAI's $852B look cheap or triggers a broad re-rating down for every lab with a private valuation built on AI hype.
Claude Mythos unauthorized access → AI cybersecurity model governance gap → EU AI Act high-risk deadline in 10 weeks Mythos is the first known instance of a state-of-the-art autonomous exploit model being accessible outside controlled channels. The EU AI Act's GPAI obligations for frontier models are already in force; the high-risk compliance deadline is August 2, 2026. Regulators now have a concrete incident to cite when deciding whether restricted-access frontier models require mandatory incident reporting — a precedent Anthropic, OpenAI, and Google all have an interest in shaping before August.
Meta $145B AI capex + Nvidia Rubin ramping → CPU supply crisis emerging → rack deployment delays GPU supply constraints are loosening as Nvidia Rubin enters production, but CPU delivery lead times have stretched to six months with prices up 10%+. Meta, Microsoft, Google, and Amazon are committing nearly $700B combined in 2026. The bottleneck is shifting from GPU allocation to the CPU host systems that pair with them — hyperscalers who locked GPU supply early may find racks sitting idle waiting on CPU availability.
DeepSeek V4-Flash MIT license at $0.14/1M → self-hostable flagship inference → margin pressure on managed-inference players Cohere, Mistral, and mid-tier API providers whose product is essentially managed inference at a markup now compete with a model developers can self-host for near-zero marginal cost. The durable revenue in the API layer moves toward fine-tuning tooling, deployment compliance, and support contracts — not raw token sales. Labs with deep model differentiation (Anthropic, OpenAI on reasoning) hold; commodity-inference plays compress.
Grok Build coding agent → xAI Wall Street network (Apollo, Morgan Stanley) → financial services as the next contested enterprise AI territory Financial services firms are the highest-margin, most security-constrained enterprise AI customers. xAI is entering with Musk relationships and a terminal-native coding agent; OpenAI's Deployment Company is moving through Bain and Capgemini into the same segment; Anthropic has Stripe and Goldman Sachs in its enterprise JV. All three major US labs are now competing for the same financial-services accounts simultaneously — and the winner likely locks multiyear contracts worth hundreds of millions.
What we’ll be watching
- Google I/O keynote today (May 19, 10AM PT) — watch for Gemini 3.2 Flash pricing confirmation, Gemini Omni multimodal demo, and whether a Gemini 4.0 flagship is announced ahead of schedule
- Meta layoffs begin May 20 (tomorrow) — 8,000 initial cuts plus 6,000 open roles canceled; watch for Llama 5 / Avocado model timeline and whether Mark Zuckerberg announces an accelerated open-weights roadmap to compensate for headcount reduction
- Anthropic's $900B funding round expected to close by end of May — watch for formal announcement, IPO filing signal, and whether the round price settles at $900B or edges toward the $950B figure some term sheets reportedly show
- Claude Mythos security investigation ongoing — Anthropic said no system-level compromise, but watch for a formal incident report under EU GPAI transparency obligations and any White House briefing following Mythos's autonomous identification of a 27-year-old OpenBSD vulnerability
- OpenAI Deployment Company funding close — 19 investors committed, $4B target, Bain and Capgemini already announced; watch for additional named partners and first disclosed enterprise deployment contracts
- EU AI Act high-risk compliance deadline is August 2, 2026 — with 10 weeks left, watch for major enterprise AI vendors publishing formal compliance roadmaps and any Commission guidance on whether restricted-access frontier models like Mythos fall under incident-reporting rules
Sources: anthropic.com, bloomberg.com, techcrunch.com, cnbc.com, openai.com, simonwillison.net, venturebeat.com, fortune.com, thenextweb.com, theinformation.com, androidcentral.com, engadget.com, x.ai, pymnts.com, axios.com, buildfastwithai.com.