The Week Anthropic Eclipsed OpenAI

Weekly number 4 of the series, and we just watched a historic flip. Anthropic closed a $65 billion round at a $965 billion valuation — more than OpenAI, at the exact moment OpenAI filed its confidential S-1 with the SEC. Meanwhile, Claude Opus 4.8 shipped and retook the lead on SWE-bench. Anthropic also opened its sixth European office in under a year (Milan), and Meta got caught in a storm after an audio leak revealed how it surveils employees to train internal models.

I've sorted through it. No hype on startups that raised $5M each, no idle speculation on the OpenAI IPO. Here are the announcements that actually change something for your tradeoffs this week.

The Major News This Week

Anthropic closes $65B at a $965B valuation — May 28: Anthropic just became the most valuable AI startup in the world. Sequoia, Dragoneer, Altimeter and Greenoaks co-lead the round, with Capital Group, Coatue, GIC, ICONIQ, and even Micron, Samsung and SK hynix following on. The symbolic bar: Anthropic is now worth more than OpenAI at the exact moment OpenAI is filing to go public. Numbers-wise, Anthropic has doubled quarterly revenue — $10.9B projected for Q2 vs $4.8B in Q1 — and is telling investors ARR will exceed $50B by end of June. The take: the frontier model race isn't a 3-4 player game anymore, it's a 2 player game. Anthropic and OpenAI together capture over $1.8 trillion in private valuation. Everyone else is in the second tier.

OpenAI files its confidential S-1 with the SEC — May 22: Goldman Sachs and Morgan Stanley lead the deal, JPMorgan as co-manager. Targeted valuation range: $852 billion to $1 trillion. Listing target: Q4 2026, possibly as early as September. The filing will finally lift the veil on numbers OpenAI had kept private: $1.22 burned for every $1 of revenue, and across all of 2025, $13.1B revenue against $22B in spend — a $9B net loss. The market's bet: API growth will offset the burn by 2027. For CTOs, the timing gets interesting: your main vendor will soon have to publish its margins, and contracts will start to reflect a profitability pressure you didn't see before.

Claude Opus 4.8 — the first time Claude retakes SWE-bench in months — May 28: Anthropic answers OpenAI's filing on the same day it closes its raise. Most loaded communications calendar of the year. Opus 4.8 jumps to 88.6% on SWE-bench Verified (vs 87.6% for 4.7 and 88.7% for GPT-5.5) and 96.7% on USAMO 2026 (vs 69.3% for 4.7). More importantly, the model is described as 4× less likely to let a bug in its own code pass unflagged. Pricing identical to Opus 4.7, available immediately on Anthropic API, AWS Bedrock, Vertex AI and Microsoft Foundry. Fast mode is 2.5× quicker than 4.7. If you run Opus in production, the migration is worth an afternoon of bench testing on your existing prompts — no more, no less.

Anthropic opens its Milan office with Generali, Pirelli and Enel as anchor customers — May 27: Sixth European office in under twelve months, after London, Dublin, Paris, Zurich and Munich. The Italian team is led by Thomas Remy, Head of Southern Europe. The anchor customer mix is telling: Generali and Unipol in insurance, Pirelli in automotive, Enel in energy, Bracco and Angelini in pharma, plus Bending Spoons and Satispay in tech. Anthropic announces EMEA ARR up 9× year-on-year and is tripling its international headcount. For European mid-caps, the signal: you no longer have to arbitrage between AWS Bedrock-Claude and Anthropic direct — pre-sales is now one TGV ride away. Euro pricing probably lands by Q3.

Meta in the storm: the Zuckerberg leak on the MCI program — May 25: An internal audio clip, leaked by More Perfect Union, shows Mark Zuckerberg defending the Model Capability Initiative — the collection of Meta employees' keystrokes, clicks and screenshots to train internal models. Same day: 8,000 layoffs. A departing Meta engineer's video openly criticizing the AI strategy goes around the internal network. Staff circulate a petition, the program shifts from HR matter to governance crisis. For companies watching how competitors build their internal datasets, the European legal precedent will get set in the next 6 to 12 months. If you're working on fine-tuning from employee data, get your DPO to validate the approach before going further.

New Models to Know

Claude Opus 4.8 — May 28: Beyond benchmarks, two concrete usage changes. First, the model now handles dynamic workflows where it can spawn hundreds of sub-agents in parallel (useful for massive tasks like multi-document extraction or full-codebase analysis). Second, Opus 4.8 is explicitly trained to flag when it's stuck instead of hallucinating a plausible answer. For agentic workflows where reliability matters more than throughput, that's a real step forward. Test it first on your code review pipelines and your technical support agents before widening rollout.

OpenAI Codex Goal Mode hits GA — May 21: Goal Mode is now generally available in the Codex CLI, app, and IDE extension. You define an outcome and success criteria, and the agent loops on plan-act-test for hours or even days, auditing its own progress at every turn. File changes, tests, benchmarks are all tracked. The Codex CLI v0.134.0 of May 26 adds conversation history search, OAuth for streamable HTTP MCP servers, and per-server MCP environment targeting. For those who've tested Claude Code and Codex side by side these past months, this is OpenAI's most serious catch-up on agentic coding.

AI Tools to Activate This Week

Microsoft Copilot Studio — computer-using agents now GA — May 2026: Microsoft moves its computer-using agents to general availability in Copilot Studio. Your agents can now click, type and navigate directly inside web or desktop applications without an API. The new workflows experience lets you integrate these UI actions into multi-step automations (approvals, business logic, API calls) in the same interface. And Microsoft announces GA for Microsoft 365 E7 — the Copilot + Agent 365 + E5 bundle. For companies already on Microsoft 365, this is the moment to look at whether E7 fits your next license renewal. The 3-year CSP pricing aligned with E3 and E5 signals Microsoft is pushing commercialization hard.

Anthropic Milan — direct access for southern European mid-caps — May 27: Why put an office in the tools section? Because for the first time, if you're a mid-cap in southern France, Italy or Spain, you can have an Anthropic Solutions Engineer on-site within 24 hours. That's a commercial argument that changes the game against OpenAI, which is still EMEA-managed from London and San Francisco. If you're POC-stage on Claude and you need support on a complex deployment, mention Milan in your next exchange — Anthropic prioritizes geographically close customers during ramp-up.

The Numbers That Matter

$965 billion — Anthropic's valuation after this round. That's more than Meta before 2023, more than Tesla today, and $215 billion more than the high end of OpenAI's IPO range. Anthropic becomes the first startup with the trillion in clear line of sight, despite being under 5 years old with fewer than 2,000 employees.

+127% quarter-over-quarter — Anthropic's revenue growth, from $4.8B in Q1 to a projected $10.9B in Q2. For comparison, that's AWS's pace in its first 2 years, but on absolute numbers 5× higher. Projected ARR exceeds $50B by end of June.

$1.22 burned for every $1 of revenue — the line in OpenAI's S-1 that will feed investor debate for 6 months. Across 2025: $13.1B revenue against $22B in spend. Goldman and Morgan Stanley's bet: API gross margin reaches operating breakeven by 2027. If that calendar slips, the IPO becomes a massive dilution.

88.6% on SWE-bench Verified and 96.7% on USAMO 2026 — Claude Opus 4.8's scores. Beyond the coding number, this is the first time a model breaks 95% on USAMO (the US math olympiads), a benchmark where Opus 4.7 plateaued at 69%. For quantitative workflows (finance, research, engineering), reasoning reliability just crossed a useful threshold.

9× EMEA ARR growth at Anthropic year-on-year. That's the number that explains the aggressive European expansion: sixth office in twelve months, tripling of international headcount. Europe is now Anthropic's second market after the US, ahead of Asia.

8,000 jobs cut at Meta on the same day as the Zuckerberg leak on the MCI program. Add that to the 92,000+ tech jobs eliminated in 2026 so far. The official "reorienting toward AI" narrative gets harder to defend publicly when AI itself becomes the subject of the internal scandal.

My Take — What I'm Actually Doing With This

Lots of money and lots of signal moving this week. Here's how I'm prioritizing tradeoffs after testing Opus 4.8 and reading the S-1 coverage closely:

On Claude Opus 4.8 in your code pipelines → Immediate migration if you're already running Opus 4.7. Fast mode 2.5× quicker at the same pricing changes the cost/quality ratio, and the "flag the block instead of inventing" behavior is exactly what was missing on long workflows. At Skello, that's exactly the kind of update we ship to prod the day after release, because the gain is measurable on code reviews and generation pipelines without requiring massive retesting.

On multi-model strategy facing the Anthropic and OpenAI consolidation → The $965B Anthropic plus OpenAI S-1 combo says the same thing: there will be only two frontier vendors that matter in 18 months. Architect your stack to switch from one to the other in under an hour (a homemade router or OpenRouter, 2 days of dev). And negotiate your annual contracts by explicitly mentioning the competitor — the favorable pricing window exists because both are fighting for enterprise share pre-IPO.

On OpenAI Codex Goal Mode → If you've tested Claude Code and you stay on the OpenAI side for ecosystem reasons (Azure, ChatGPT Enterprise agents, existing contracts), this is the right moment to seriously test Goal Mode. The quality jump from Codex CLI v0.130 is tangible on multi-file tasks. For B&Inside, that's exactly what I'm benchmarking right now against Claude Code on massive data extraction pipelines.

On Microsoft 365 E7 and Agent 365 → If your company is on E5 and renewal lands within 12 months, ask for E7 pricing now. The feature that justifies the upgrade: computer-using agents in GA let you automate the workflows that have no API (legacy ERP, custom internal tools, old back-offices). If you're in consulting or public sector, this is probably your best short-term ROI before looking elsewhere.

On Meta MCI → Not a tool to adopt, a risk to monitor. If your company is starting to explore fine-tuning on employee data (Slack logs, meeting transcripts, screenshots, IDE telemetry), get your DPO in the loop before touching the dataset. The European legal precedent will get set in the next 12 months, and nobody wants to be the test case.

See You Next Week

The weekly cadence is holding. Same next week: no hype, no recap of startups that each raised $5M, just the announcements that move your stack or your decisions.

If one of these stories resonates — the Opus 4.8 vs GPT-5.5 tradeoff, multi-model strategy facing consolidation, or caution on employee data fine-tuning — let's talk. See you next week.