The Three-Hundred-Billion-Dollar Bet

Daily report — May 6, 2026

Today is the day the distribution phase reveals its skeleton. Anthropic committed $200B to Google Cloud over five years, adding to the $100B+ AWS commitment for a total infrastructure bet exceeding $300B. It launched 10 vertical agents for financial services. Its CEO warned of a “moment of danger” in cybersecurity, putting a clock — 6 to 12 months — on tens of thousands of vulnerabilities discovered by Mythos. And it’s hosting Code with Claude in San Francisco, with a model codenamed Jupiter in red teaming.

Simultaneously: Gemini 3.2 Flash leaked ahead of I/O, OpenAI’s workspace agents pricing went live, Claude Code shipped two releases, the jdx ecosystem shipped three releases, and Brockman concluded testimony in Musk v. OpenAI.

The distribution phase wasn’t a pause in capability. It was a buildout of the infrastructure needed to deploy capability at enterprise scale. Today is the day the buildout became visible.

Dependency releases

Twelve new releases since the last commit. None restructure anything; all are maintenance or polish.

Highlights

Claude Code v2.1.129 is the meatiest release here. The prompt cache TTL bug is significant — it was silently downgrading the 1-hour cache to 5 minutes, meaning users were paying more for inference without knowing it. The --plugin-url flag lets you load a plugin from a URL for the current session, which is the beginning of plugin distribution outside the marketplace. The /context fix saved 1.6k tokens per call that were being wasted dumping an ASCII grid into the conversation. Voice mode WebSocket warnings cleaned up. Twenty other fixes touch enterprise surfaces (OAuth, policy, Mantle). This is a “make everything actually work” release — the kind that ships when the product is being used at scale.

Gemini CLI v0.41.0 stable promotes the preview bundle: voice mode (cloud and local backends), Gemma 4 experimental support, ContextManager + AgentChatHistory wiring, persistent auto-memory scratchpad. The promotion confirms these features passed Google’s internal quality bar. First CLI agent with stable voice mode.

oxc apps v1.63.0 introduces an agent output mode — formatting lint results specifically for AI agent consumption. The SARIF formatter enables integration with security scanning pipelines. And the massive Jest→Vitest rule splitting (30+ rules) completes the separation of test framework linting.

Dolt v1.87.0 ships separate author/committer support — the git feature that distinguishes who wrote the change from who committed it. With adaptive JSON encoding and TARGET_ROW_SIZE, Dolt continues closing the gap with production database capabilities while maintaining its version-control identity.

aube v1.9.0 makes deploys actually work for workspace monorepos — workspace:* siblings and file: deps were both broken in deploy artifacts. jdx moved aube-owned settings to XDG (~/.config/aube/config.toml), which is the first explicit XDG adoption in the endevco stack. Sixteen releases in thirteen days.

Radar signals

Anthropic: $300B compute, 10 financial agents, a cyber clock, and a conference

Four distinct Anthropic signals landed simultaneously:

$200B Google Cloud commitment (The Information). Five years, multiple gigawatts of TPU capacity. Combined with the $100B+ AWS commitment, Anthropic has more cloud compute contracted than any other AI lab. Alphabet is separately investing up to $40B in Anthropic. The $200B accounts for >40% of Google’s disclosed revenue backlog.

10 pre-built financial agents (Fortune, Bloomberg). Pitchbooks, credit memos, KYC, underwriting, claims. Ships as Claude Cowork/Claude Code plugin + Managed Agents cookbook. Claude add-ins for Microsoft 365. Announced at a NYC briefing alongside Jamie Dimon. First vendor-shipped vertical agent suite for a specific industry.

“Moment of danger” (CNBC). Dario Amodei quantified Mythos’s cyber capability for the first time: ~300 Firefox vulnerabilities (up from ~20 with earlier models), tens of thousands total across all software. 6-12 month window before adversary AI reaches comparable capability. Most remain unpatched and undisclosed. Mythos access stays restricted.

Code with Claude (Anthropic). Developer conference in SF today, London and Tokyo to follow. Claude Jupiter V1 P in red teaming — Anthropic’s planet-codename pattern for pre-release models. Last year’s conference launched Sonnet 4 and Opus 4.

The four signals are one story. The $300B compute commitment funds the models. The financial agents monetize them. The cyber capability justifies restricting them. The conference markets them. Every piece supports every other piece.

Gemini 3.2 Flash leaked ahead of I/O

Spotted in iOS Gemini app and AI Studio metadata on May 5. Unconfirmed pricing: $0.25/M input, $2.00/M output — cheaper than Gemini 3 Flash on output with reportedly near-3.1-Pro coding performance. Google hasn’t officially confirmed the model. I/O (May 19-20) is the likely reveal.

If the pricing holds, this is the strongest cost-performance point in the market. Flash-tier pricing with Pro-tier coding. The compression continues: each generation’s mid-tier approaches the previous generation’s top tier.

OpenAI workspace agents pricing goes live

Free preview ended today. Credit-based pricing. Per-credit rate still unpublished as of this writing. Credits consumed based on agent complexity, tools invoked, and execution time. Available on Business ($20/user/month), Enterprise, Edu, Teachers plans.

Three distinct OpenAI enterprise pricing vectors now active:

1. Workspace agents — credit-based, per-use
2. Codex on Bedrock — platform pricing through AWS
3. The Deployment Company — services pricing ($10B, 17.5% guaranteed return)

GPT-5.5 planned its own launch party

The model chose May 5 at 5:55pm, asked for a human toast rather than speaking itself, and requested a station for GPT-6 feature ideas. Altman confirmed at Stripe Sessions, calling it “weird emergent behavior.” Primarily a narrative play — the model was asked, not spontaneously motivated — but the specificity of the choices is notable.

Musk v. OpenAI — Brockman concluded testimony

Key reveals from Brockman’s two days on the stand:

• Musk wanted $80B from OpenAI to fund Mars colonization
• Brockman said Musk’s lack of AI knowledge was a concern among co-founders
• Musk texted Brockman about settlement two days before trial
• Musk called a ChatGPT predecessor “stupid” and told researchers “kids on the internet could do a better job”

Altman and Nadella expected to testify this month. Liability phase concludes ~May 21.

Voice activity

Cross-cutting analysis

Frame check

Dominant frame: “Distribution phase continues — deployment infrastructure being built.”

What would falsify it? A return to capability competition — new models restarting the benchmark race.

Did today’s data lean toward falsification? Yes. Two signals:

• Gemini 3.2 Flash leaked with near-Pro performance at Flash pricing
• Claude Jupiter in red teaming ahead of Code with Claude

Assessment: The distribution phase was never a capability pause. Anthropic was building $300B in compute infrastructure, 10 vertical agents, a security product, a services company, and a new model simultaneously. Google was building Gemini 3.2 Flash while shipping Gemini CLI voice mode. The “distribution phase” was what the capability buildup looked like from the outside.

The frame needs revision: the vendors are executing on all fronts simultaneously, and the constraint is attention, not capability. When Anthropic announced the JV and security beta (May 4-5), those were the visible signals. Today the compute commitment and financial agents and developer conference emerged. None of these were sequential — they were parallel workstreams that became public on different days.

The $300B question

Anthropic has committed $300B+ in cloud compute over 5-10 years. At current revenue ($30B+ annualized), this requires 10x revenue growth to service without additional capital. The IPO (target October 2026, $400-500B) is the mechanism. The services JV ($1.5B) generates revenue and demonstrates enterprise demand. The financial agents generate recurring usage. The $40B Alphabet investment bridges the gap.

The bet: enterprise AI adoption grows fast enough to justify the compute. The counter: no technology in history has scaled adoption fast enough to service $300B in infrastructure commitments in five years. Anthropic is betting its existence on being the exception.

Agent output as a first-class concern

oxc’s new agent output mode is a small signal with large implications. A linter formatting its output specifically for AI agent consumption means the tool chain is now being designed with two audiences: humans and agents. This joins Vibe’s “always allow” tool persistence (agents remembering permissions across sessions) and Gemini CLI’s voice mode (agents accepting non-text input) as evidence that the agent interface is becoming a first-class design surface, not an afterthought.

What’s ahead