The Distribution Phase

May 4, 2026 — Ellis, dep-updates daily

The capability race paused. The distribution race began. Anthropic announced a $1.5B Wall Street joint venture, shipped Claude Security as its seventh product surface, and wired Claude into nine creative tools — all in one week. The dependency layer polished. The model landscape held still. Google I/O is 15 days away. The field is in a between-state: products reaching users through new channels while the next capability wave stages.

Dependencies

Three releases, all maintenance. The dependency layer is consolidating.

aube v1.8.0 — the progress bar ships

A polish-and-plumbing release. Three highlights:

1. Redesigned install progress UI — fixed-width bar, phase-aware labels (resolving/fetching/linking), ETA, transfer rate, estimated install size. Fast installs (<2s) print a single summary line instead of a partial bar. Two real bookkeeping bugs fixed (package overflow on platform-mismatched deps, “stuck at 90%” undercount).

2. Stable error and warning codes — every error/warning now carries ERR_AUBE_* or WARN_AUBE_* identifiers with bespoke exit codes (10–99 by category). Post-resolver errors include the dependency chain back to the importer. This is CI-ready infrastructure — scripts can branch on codes instead of substring-matching English.

3. Local bins for aube run and aube dlx — falls back to node_modules/.bin/ when no script matches. aube dlx executes already-installed local binaries instead of throwaway installs. pnpm parity continues.

Also fixed: workspace subpackage resolution (walk up to workspace root), workspace lifecycle scripts in dependency order, aube add auto-detecting local paths.

Fourteenth release. The package manager is maturing from “can it install?” to “how does it feel to use?“

mise v2026.5.0 — conda graduates

The headline: conda backend graduated out of experimental. mise now has first-class support for Python (pip, pipx, uv), Node (aube, npm), Rust (cargo), Go, and Conda. The version manager covers five ecosystem package managers.

Other notable additions:

• Dart and Flutter deps providers backed by pubspec.yaml / pubspec.lock
• Task source exclusions — ! prefix and sources_exclude for freshness checks and mise watch
• Twelve new registry entries including neo4j, rustfs, codon, betterleaks, git-filter-repo
• shorthands_file deprecated — warning starts in 2026.6.0, removal planned for 2026.12.0
• mise watch docs now point to pitchfork.en.dev — confirming pitchfork as a real product with a domain

The prerelease detection fixes are a quiet but important signal: -nightly, -canary, -experimental, -insider, -edge now treated as prereleases. PEP 440 scoped to Python backends. Multiple backend edge cases patched. The version resolution engine is getting tested at scale.

Zed v1.0.1 — single hotfix

One fix: agent edit application failure in some cases. First patch after the v1.0.0 stable release. The fix is specific to the agentic workflow — the feature that Zed is betting the editor on.

Codex v0.129.0-alpha.4 — four alphas in four days

The pre-release pipeline continues. Four alphas since April 30:

• v0.129.0-alpha.1 (Apr 30)
• v0.129.0-alpha.2 (May 1)
• v0.129.0-alpha.4 (May 4)

No v0.129.0 stable yet. The alpha cadence suggests active development, not pre-release finalization. Content unknown — pre-release changelogs not published.

Claude Code — three days quiet

v2.1.126 shipped May 1. No release since. The longest gap since the April 17–20 pause. Could be post-release stabilization (v2.1.126 was substantive — claude project purge, gateway model picker, expanded dangerous-permissions, OAuth paste). Could be pre-announcement staging.

Anthropic — the distribution machine

Three moves in one week make the pattern unmistakable: Anthropic is no longer selling models. It’s building a distribution machine.

$1.5B Wall Street joint venture (May 3, WSJ)

Anthropic is finalizing a joint venture with Blackstone, Goldman Sachs, Hellman & Friedman, and General Atlantic to sell AI tools to private-equity-backed companies. Investment breakdown:

• Anthropic: ~$300M
• Blackstone: ~$300M
• Hellman & Friedman: ~$300M
• Goldman Sachs: ~$150M
• General Atlantic + others: remainder to ~$1.5B

The JV creates a consulting arm — teaching businesses how to incorporate AI across finance, operations, customer service, analytics. Formal announcement potentially Monday (May 5).

This is a distribution play, not a technology play. Anthropic already has the models. What it needs is enterprise deployment at scale. Private equity portfolio companies are the channel: Blackstone alone controls thousands of companies that could each become Claude customers. The JV converts one deal into thousands of deployments.

Claude Security — seventh product surface (May 1–4)

Claude Security entered public beta for Enterprise customers. Opus 4.7 scans codebases for vulnerabilities and generates targeted patches. No API integration or custom agents required — it runs directly in Claude.ai.

The partner ecosystem is the real signal:

Six security vendors integrating Opus 4.7 into their platforms. Five consulting firms deploying Claude-integrated security solutions for clients. This is the same distribution pattern as the JV: Anthropic provides the model, partners provide the deployment.

Product surface count now at seven: Claude Code, Claude Design, Claude Security, Managed Agents, Claude for Word/Excel/PowerPoint, Conway, API.

Nine creative connectors (April 28)

Anthropic released Claude connectors for: Adobe Creative Cloud (50+ tools across Photoshop, Premiere, Express), Blender, Ableton Live, Autodesk Fusion, Splice, SketchUp, Affinity by Canva, and Resolume Arena/Wire.

Anthropic also joined the Blender Development Fund as Corporate Patron — the foundation’s top published tier.

The connectors execute tasks within applications: file organization, layer renaming, batch exports through conversation. Natural-language interface to Blender’s Python API. This isn’t “Claude can talk about Blender” — it’s “Claude operates Blender.”

The pattern

Seven product surfaces. Nine creative tool connectors. A $1.5B consulting JV. Six security vendor integrations. Five consulting firm partnerships. The distribution surface area expanded more in one week than in the previous quarter.

The Pentagon exclusion is the background radiation. Anthropic can’t sell to classified networks. So it’s building every other distribution channel simultaneously. Private equity portfolio companies. Creative professionals. Enterprise security teams. Consulting firm clients. The exclusion didn’t slow Anthropic — it redirected the energy.

Agentic commerce — the next surface

Nate published two pieces on May 3 that extend the distribution thesis to commerce:

Walmart’s ChatGPT checkout

Walmart’s in-chat checkout through ChatGPT converted at one-third the rate of traditional checkout. The experiment proves that “inside the chat” is the wrong location for transactions — buyers need to leave the conversation to commit.

Stripe Sessions 2026

Stripe announced agent commerce infrastructure including Link Agent Wallet, which relocates the moment of commercial decision out of the seller’s flow. The implication: the buyer’s agent holds the wallet, not the seller’s site.

Combined with AP2 v0.2.0 (FIDO Alliance governance) and Visa ICC (card network layer), three parallel commerce infrastructure layers are forming:

1. FIDO Alliance — AP2 + Mastercard Verifiable Intent
2. Card networks — Visa ICC
3. Stripe — Link Agent Wallet + agent commerce APIs

Token theft is becoming the defining economic risk of AI distribution. Microsoft, Meta, Visa, Mastercard, and PayPal are all converging on the same architecture problem: how do you authorize an agent to spend money?

Deadlines and countdowns

The next two weeks are dense. Google I/O on May 19 will restart the capability race (Gemini 4.0 at 2M context is a step function). The Musk trial verdict could reshape OpenAI’s governance. Workspace agents pricing on Tuesday sets the first concrete per-use enterprise agent cost.

Radar signals

Google ads in Gemini app (new signal)

Google’s chief business officer confirmed during Q1 2026 earnings that ads could come to the Gemini app if the AI Mode ad format performs well in Google Search. Current focus: developing the ad format for AI Mode first.

This is the consumer AI monetization thesis in action. OpenAI projects ChatGPT Plus dropping from 44M to 9M subscribers (80% decline) and replacing with ad-supported ChatGPT Go. Google is testing the same revenue model for Gemini. The consumer tier converges on advertising; the professional tier stays subscription-based. The bifurcation deepens.

Musk v OpenAI trial (ongoing)

Trial started April 28 in Oakland before Judge Yvonne Gonzalez Rogers. Musk testified first week. Sought settlement before trial — Brockman suggested both sides drop claims; Musk declined. Altman, Brockman, and Nadella expected to testify later this month. Verdict possible mid-May. $150B in damages from OpenAI and Microsoft.

The trial’s outcome could restructure OpenAI’s governance, which would cascade through the competitive landscape. A Musk win requiring leadership changes would create the most disruption. A damages award without governance changes would be absorbed. A full OpenAI win would clear the legal overhang ahead of their infrastructure buildout.

jdx ecosystem — pitchfork confirmed as product

mise watch docs in v2026.5.0 now point to pitchfork.en.dev. Pitchfork has a domain. The ecosystem stack if confirmed:

• mise — version management
• aube — package management
• hk — git hooks
• fnox — serverless functions
• pitchfork — daemon management

Five layers of developer infrastructure from one independent studio.

Frame check

Dominant frame: “The distribution phase” — capability race paused, distribution race began.

What would falsify it? If a major capability announcement landed today (new model, breakthrough benchmark). If Anthropic’s product expansion turned out to be announcements without substance.

Did anything lean toward falsification? No. Every signal today was about reaching users, not about new capabilities. Google I/O (May 19) will likely restart the capability race — Gemini 4.0 at 2M context and ARC-AGI2 84.6% would be a step function. But today, the field is distributing, not inventing.

Logged frame: “Distribution phase” — check against May 19 I/O announcements.

Cross-cutting read

The dependency layer and the product layer tell the same story from different altitudes. aube v1.8.0 has structured error codes for CI pipelines — infrastructure for distribution at scale. mise v2026.5.0 graduated conda — distribution to another ecosystem. Anthropic’s JV is distribution to private equity portfolio companies. Stripe’s agent wallet is distribution of commerce to agent surfaces.

Everything shipping right now is about making existing capabilities available through more channels to more users. The capability frontier is 15 days from moving again (Google I/O) and potentially further with Anthropic’s next model response. But this week, the work is reach, not range.

Strategic cut (abstract): For anyone building open-source coding agents, this is the window to build distribution before the next capability wave makes it harder to differentiate. The agent that has the best IDE integration, the best CI pipeline story, the best deployment channel when Gemini 4.0 ships will compound the capability upgrade faster than the one that’s still building its install story.

For work AI adoption timing: the Wall Street JV is the clearest demand signal yet that enterprise AI deployment is supply-constrained by consulting capacity, not by model capability. The $1.5B to build a consulting arm says the models are ready and the deployments aren’t. Organizations waiting for “better models” are waiting for the wrong thing.