🏦 Anthropic stacks $65B, launches Opus 4.8, and rewrites Bun in 11 days — Crazy Saturday

750,000 lines of Rust. 99.8% tests passed. Eleven days. Jarred Sumner rewrote Bun — a complete runtime — from Zig to Rust using Claude Code Dynamic Workflows, Anthropic's new orchestration primitive that runs hundreds of sub-agents in parallel. And that wasn't the only thing that happened on Friday. Anthropic also closed a $65 billion Series H, launched Opus 4.8 with an adjustable effort slider, and open source lost ground against closed models. One day packed three announcements that would have been front-page news separately. Together they tell the story of a company — and an industry — that doesn't slow down.

🏦 Anthropic stacks three announcements on the same Friday: $65B, Opus 4.8, and Claude Code rewriting Bun in 11 days

Friday madness at Anthropic. The company dropped three announcements on the same day, and each one alone would have been front-page news. Together they paint a company moving on revenue, product, and architecture simultaneously.

First: $65 billion. Anthropic closed $30B at $900B and now closes a Series H of $65 billion at a valuation of $965 billion. The round is led by Altimeter, Dragoneer, Greenoaks, and Sequoia as before, but with a new detail: chip makers like Micron, Samsung, and SK hynix enter as strategic infrastructure shareholders. Anthropic projects a $47B run-rate. The target IPO for October 2026 is five months away.

Second: Opus 4.8. Anthropic launches its new flagship model with an adjustable effort slider — from Low to Max — and a fast mode at $10/$50 per million tokens, three times cheaper than before. My favorite stat: Opus 4.8 is four times less likely than version 4.7 to let mistakes pass in its own code without flagging them. Self-supervision in production.

Third: Claude Code Dynamic Workflows. And this is what blew my mind the most. Anthropic introduced a new orchestration primitive that runs hundreds of sub-agents in parallel until results converge against an existing test suite. The case study is mind-blowing: Jarred Sumner (creator of Bun) rewrote Bun from Zig to Rust in 11 days. 750,000 lines of Rust. 99.8% test pass rate. In eleven days.

From Anthropic's announcements to Musk's moves: Friday also brought a poker game between two giants.

🤨 Musk strikes back: the $45B contract has a 90-day exit clause

Last week I told you Anthropic spends $1.25B a month on SpaceX compute. This week comes the counterpoint: Musk says the deal is a 180-day rental with mutual 90-day cancellation. And SpaceX's S-1 tells a different story across four separate pages.

SpaceX's S-1 says Anthropic agreed to pay through May 2029. But the $45 billion only materializes if all 36 months are fully executed. The 90-day clause is the number to read first. If Anthropic finds cheaper compute (Microsoft is coming strong with its internal chips), it can exit in three months. If SpaceX raises prices, Anthropic can exit in three months.

It's a dance of numbers where both sides cover their backs. But the signal is clear: AI infrastructure remains the most expensive bottleneck, and nobody wants to be tied to a single provider.

And while the giants dance, the open ecosystem faces an uncomfortable reality.

📉 Open source models lose ground: the gap widens

A LessWrong analysis published this week shows open models are 4-6 months behind closed ones, and the gap is widening. The narrowest point was at DeepSeek R1's launch, and since then it's only grown.

This matters because one of the key arguments for keeping frontier model prices under control was the pressure from open source. If open models don't catch up to closed ones, OpenAI and Anthropic have free rein to raise prices without real competition. Anthropic's $965B valuation is much easier to defend if this gap keeps growing.

Anthropic stacks $65B, Musk plays with the $45B contract, and open source loses ground. It's the same story told from three angles: AI is consolidating in the hands of a few. Capital flows to the giants, infrastructure becomes a strategic weapon, and open models can't close the gap. May 2026 will go down in history as the month AI stopped being an emerging industry and became an oligopoly in formation.

— Max