Anthropic Passes OpenAI as Most Valuable AI Startup at $965B
Anthropic closed a $65 billion Series H at a $965 billion valuation — surpassing OpenAI’s $852B for the first time. Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital led the round, with Google and Amazon contributing billions each. Even chipmakers Samsung, SK Hynix, and Micron joined. Anthropic now projects a $50B+ annualised run rate by end of June.
The raise coincided with KPMG announcing Claude deployment to all 276,000 employees, joining Deloitte’s 470K users. Combined, Anthropic has direct distribution to over a million professional services workers. An IPO is expected this fall.
Why it matters: The narrative has shifted. Anthropic hasn’t just caught OpenAI — it’s passed it on valuation, revenue trajectory, and enterprise distribution simultaneously. The Big Four deployment strategy is a structural moat that OpenAI’s $4B DeployCo subsidiary is now racing to match.
Claude Code Source Leak Exposes 512K Lines — Including Tamagotchi Pet and Always-On Agent
Anthropic accidentally exposed Claude Code’s entire TypeScript codebase through a source map file bundled in the 2.1.88 release on npm. The leak, containing 512,000+ lines of code, revealed:
- A Tamagotchi-style pet: An AI creature that “sits beside your input box and reacts to your coding” — complete with 18 species and a rarity system.
- “KAIROS”: An always-on background agent feature that runs permanently alongside the user.
- “Undercover Mode” and “Ultraplan” — hidden features for stealth operation and parallel subagent orchestration.
- Internal developer comments: “the memoization here increases complexity by a lot, and I’m not sure it really improves performance.”
The code was quickly forked to GitHub (50,000+ forks). Anthropic fixed the leak within hours, calling it “a release packaging issue caused by human error, not a security breach.” Gartner analyst Arun Chandrasekaran says the leak’s long-term impact is limited but serves as “a call for action for Anthropic to invest more in operational maturity.”
Why it matters: The leak gives the most detailed look yet at how a frontier AI company builds agentic tools internally. The Tamagotchi feature is oddly charming, but KAIROS (always-on agents) and Ultraplan (parallel subagent orchestration) are genuinely significant — they reveal where Anthropic is heading with autonomous AI, and it’s farther than the public product shows.
Meta’s $145B AI Bet Spawns a Cloud Business — Zuckerberg Signals War with AWS
Meta CEO Mark Zuckerberg signalled to shareholders that building a cloud computing business is “definitely on the table” if the company’s data centre buildout exceeds internal demand. Meta has committed $145 billion to AI infrastructure — and the company already operates one of the world’s largest private compute fleets.
The $145B figure reflects both capital expenditure and operating costs across Meta’s massive GPU clusters. With 8,000 workers cut in recent layoffs, the cloud play would allow Meta to monetise its overcapacity the same way Amazon turned infrastructure into AWS. Zuckerberg is positioning the move as a natural evolution: “If we build more capacity than we need for our own AI workloads, we should make it available to others.”
Why it matters: The AWS comparison is the right one — Amazon built cloud infrastructure for itself, then sold it. Meta’s $145B bet means it will have surplus compute. If Meta Cloud launches, it would compete directly with AWS, Azure, and GCP, shifting the cloud market’s centre of gravity. The question is execution: Meta has zero enterprise cloud experience.
Waymo Opens Ojai Robotaxi — 42% Fewer Sensors, $75K Cheaper per Vehicle
Waymo has opened its sixth-generation “Ojai” robotaxi to select riders in San Francisco, Los Angeles, and Phoenix. Built by Geely’s Zeekr in China, the Ojai cuts the sensor count by 42% — removing lidars and reducing camera arrays — while maintaining Level 4 autonomous capability. Each vehicle costs roughly $75,000 less to manufacture.
The cabin is roomier with a removable steering wheel, designed explicitly for driverless operation from the ground up rather than retrofitting existing cars. Waymo plans to expand the fleet through 2026, targeting a lower per-mile cost that makes robotaxis economically viable at scale.
Why it matters: The 42% sensor reduction is the key number. Autonomous vehicle companies have long argued they need expensive sensor suites to achieve safety. Waymo just proved they can cut nearly half the sensors and still operate safely in three major cities. That’s the cost curve breaking open — and it changes the economics of robotaxi deployment completely.
GPT-5.6 “Iris” Surfaces in OpenAI Codex Logs — 1.5 Million Token Context
OpenAI’s unreleased GPT-5.6 has appeared in Codex logs under the codename “iris-alpha,” according to leak reports. The model features a 1.5-million-token context window — roughly triple GPT-5.5’s capacity — and sources close to the leaks suggest an early June release target.
GPT-5.5 launched April 23, reaching 82.7% on Terminal-Bench with a $5/1M token API price. The 5.6 “Iris” jump would position OpenAI to compete directly with Claude Opus 4.8 on both context length and agentic capability.
Why it matters: Three frontier model releases in two weeks — Claude Opus 4.8, Gemini 3.5 Flash, and now GPT-5.6 Iris — suggest the release cadence is accelerating, not slowing. A 1.5M context window would enable use cases (full codebase analysis, multi-hour conversation memory) that were impractical even a month ago. June is shaping up to be the most competitive month in AI history.
ADL Report: Leading AI Models Show Systemic Anti-Israel and Antisemitic Bias
The Anti-Defamation League released the most comprehensive evaluation yet of bias in major AI models, finding “significant instances of anti-Israel and antisemitic bias” across GPT-5.5, Claude Opus 4.8, Gemini 3.5, and Meta’s Llama 4. When prompted about Israel-related topics, models produced historically inaccurate claims and denied Jewish ties to Jerusalem at disproportionately higher rates than for comparable geopolitical conflicts.
A companion ADL AI Index found that no model could consistently detect antisemitic content across all tested categories — and performance gaps between the best and worst models exceeded 40 points.
Why it matters: The ADL study adds to a growing body of evidence that AI models reflect and amplify the political biases of their training data. The finding that bias is highest on Israel-related topics — and that no major lab has solved it — raises questions about deploying these models in contexts where geopolitical neutrality is essential.
EAGLE 3.1 Solves Attention Drift — Makes LLM Inference Up to 5× Faster
The EAGLE series (speculative decoding) reached version 3.1 through a collaboration between the EAGLE team, vLLM, and TorchSpec. EAGLE 3.1 fixes “attention drift” — a previously undiagnosed problem where draft tokens from speculative decoding cause attention pattern divergence in the target model.
The fix: a lightweight correction mechanism that adjusts the draft model’s attention as sequence length grows. Benchmarks show 3-5× speedups on standard LLM inference workloads, with greater gains on longer generations.
Why it matters: Speculative decoding is one of the few techniques that can speed up inference without sacrificing output quality. EAGLE 3.1 fixing attention drift removes a key barrier to wider adoption. For anyone running LLMs in production, this means faster responses without changing models.
🔍 THE BOTTOM LINE
Three themes dominate this week’s news: the infrastructure race (Meta’s $145B cloud play, Waymo’s cost breakthrough), the capability disclosure dilemma (Claude Code’s leak reveals more than Anthropic intended, GPT-5.6 Iris surfaces before launch), and the bias accountability problem (ADL’s findings show no major lab has solved it yet). June is shaping up to be the most competitive month in AI history, with Anthropic, OpenAI, Google, and xAI all expected to ship significant releases.
❓ Frequently Asked Questions
Q: Is Anthropic really worth more than OpenAI now? On paper, yes — $965B to $852B. But OpenAI’s valuation hasn’t been updated since its last raise, and both are pre-IPO. The real comparison will come when both file to go public, expected this year.
Q: What did the Claude Code leak actually expose? 512,000 lines of TypeScript source code from Claude Code’s CLI tool. It revealed unannounced features (Tamagotchi pet, KAIROS always-on agent, Ultraplan orchestration) and internal developer comments. No customer data was exposed — it was a build/release packaging error.
Q: Is Waymo’s Ojai safe with 42% fewer sensors? Waymo’s own data says yes — the 6th-gen system achieves the same Level 4 safety metrics with fewer, more strategically placed sensors. The company has operated in San Francisco, LA, and Phoenix without major incidents.
SOURCES
- The Verge — Claude Code source code leak
- CNBC — Anthropic tops OpenAI as most valuable AI startup
- Electrek — Waymo Ojai robotaxi 6th-gen Driver
- 247 Wall St / FrontierNews.ai — Meta $145B AI bet
- TechCrunch — Anthropic $65B raise
- Ars Technica — Illinois AI regulation
- ADL — Anti-Israel bias in LLMs
- FrontierNews.ai — EAGLE 3.1 attention drift fix, Meta cloud
- Make Tech Easier — Claude Mythos banking concerns
- Geeky Gadgets — Anthropic model leaks
- AIScroll — GPT-5.6 Iris leak