Issue #06  ·  Volume I  ·  The Toolmaker
Anthropic bought the factory that builds SDKs for OpenAI, Google, and Meta, then announced it would close it.
2026 — 05 — 22
Period 2026-05-16 → 2026-05-22
18 Releases · 6 Days
06
Shipped.  ·  Volume I  ·  The Toolmaker  ·  06 Shipped.

The Toolmaker

Anthropic bought the factory that builds SDKs for OpenAI, Google, and Meta, then announced it would close it.

p.02 Shipped. · 06 · Front of book
Front of book
06
In this issue,
01
The Open
02
The Lead Story
03
By the Numbers
04
Also Shipped
05
Quiet on the Wire
06
Term of the Issue
07
The Close
p.03 Shipped. 00 Apr 2026
The Open

Monday morning, Anthropic bought a small company most builders had never heard of and every builder had already used. Stainless generated the official SDK for Anthropic's own API. It also generated the Python, TypeScript, Go, Java, Kotlin, and Ruby libraries that OpenAI, Google, Cloudflare, and Meta ship to their own developers. One startup quietly built the scaffolding under the whole frontier, and almost nobody outside the toolchain knew its name.

Then Anthropic said it would shut the service down.

No model shipped this week. No flagship. The headline was a $300-million-plus acquisition of a tool nobody talked about, and the decision to switch it off. The rest of the week was Claude Code patching itself in public, six versions in five days, twice in one Friday. The center of gravity moved from what the model can do to who owns the machinery that builds with it.

The Lead Story 01
Report by Eddie Belaval

The factory that built everyone's tools.

The frontier race had a quiet move this week, and it wasn't a benchmark. It was an acquisition most people couldn't place.

On Monday, May 18, Anthropic announced it had acquired Stainless, the startup that generates production-ready SDKs, CLIs, and MCP servers directly from API specifications (Anthropic). Founded in 2022 by former Stripe engineer Alex Rattray and backed by Sequoia and Andreessen Horowitz, Stainless sat in a strange place in the toolchain. It generated every official Anthropic SDK since the API's earliest days. It also generated the Python, TypeScript, Go, Java, Kotlin, and Ruby libraries OpenAI, Google, Cloudflare, and Meta hand their own developers (TechCrunch). Terms weren't disclosed. The Information had reported, before the announcement, that talks valued the company at over $300 million, with part of the consideration potentially in Anthropic equity.

Read the headline and you'd file it as a routine dev-tools tuck-in. Read the second sentence of the press release and the story changes. Anthropic will wind down all hosted Stainless products, the SDK generator included. Existing customers keep the SDKs they've already built, with full rights to modify them. The factory closes. The parts stay where they landed.

Sit with the mechanism, because the mechanism is the whole story. A spec-to-SDK generator is not a convenience. It is the thing that turns one OpenAPI document into seven idiomatic client libraries, keeps them in sync release after release, and absorbs the unglamorous labor of pagination helpers, retry logic, streaming parsers, and typed errors in six languages at once. A lab ships an API change on Tuesday; the generator re-emits libraries that match it by Wednesday. Stainless was the part of the supply chain nobody photographed because it sat below the layer anyone looks at. It was the loom under the cloth.

So name the blast radius precisely. OpenAI and Google do not lose the SDKs they have today. They keep them, with full rights to modify. What they lose is the supplier. The next time one of them ships an API change and reaches to regenerate, the hosted machine that did it for them is dark. They can fork what Stainless left behind, hire a team to maintain a generator in-house, or build their own from scratch, and any of those is a real engineering cost on a workstream they had been quietly outsourcing to a competitor. The library on disk still works. The conveyor belt that kept it current is gone. That is the difference between buying a feature and buying a chokepoint.

It rhymes with the spring's other Anthropic story. Capacity was about owning the compute everyone needs to run. This is about owning the tooling everyone needs to build. Same instinct, much smaller check: when the constraint is a dependency you don't control, buy the dependency. A $30-billion-run-rate lab paid roughly the price of a few weeks of its Colossus compute bill to take a load-bearing brick out of its rivals' foundations. The Memphis data center cost an estimated $1.25 billion a month (TechCrunch); Stainless reportedly cost north of $300 million once (TechCrunch). One of those is a recurring tax. The other is a permanent subtraction from someone else's stack.

Read against the rest of the frontier, the move has teeth the press release declined to show. Every lab now sells agents, and agents are only as good as the SDKs and MCP servers that let them touch the world. Anthropic just bought the standard tool for generating exactly those surfaces, and then it removed that tool from the market it shares with the labs it competes against. The contrast is the read: while the field races to ship the better model, Anthropic spent the week making sure the field has a harder time shipping the better client. Nobody loses a model over this. Everybody who relied on the generator inherits a maintenance problem they didn't have on Sunday.

For a builder, the move tonight is narrow and worth making. If your pipeline pulls a Stainless-generated SDK for any provider, vendor the current output now and read the wind-down terms while the docs are still up. You keep what you have and the right to modify it; you do not keep the button that regenerated it. Pin the version, check it into a place you control, and stop assuming the upstream loom will be there next quarter.

A toolmaker that closes its doors after you've moved in. That's the deal.

✦   ✦   ✦
Shipped. · Also shipped
Also shipped

Anthropic names 2026 the breakaway year.

The week opened on a Saturday with a single loud sentence. Anthropic's research team published "2028: Two scenarios for global AI leadership," a paper that puts semiconductors, not capabilities or safety, at the center of the US-China race (Anthropic Research). In scenario one, democracies tighten chip export controls, block adversarial-state access to American frontier models, and hold a 12-to-24-month compute lead through 2028. In scenario two, policymakers defer, China closes the gap, and the norms governing AI globally get written by authoritarian states. "There is a high likelihood that we will look back on 2026 as the breakaway opportunity for American AI," the paper argues, and the rest of it is the case for that sentence. It published on anthropic.com while Trump was in Beijing for the first time in nearly a decade (coverage). A lab that spent the week buying a tool out from under its rivals also spent a Saturday arguing that the real contest is over chips, not capabilities, and that the clock is the policy lever. The timing was not an accident.

The API quietly gained three betas.

Three API items dated May 18 surfaced through the week, none with an announcement, none with a stage. Cache diagnostics, in public beta, reports a cache_miss_reason field when a prompt-cache prefix diverges from the prior turn, so you can debug multi-turn cache behavior instead of guessing where the prefix broke; it rides behind the cache-diagnosis-2026-04-07 header (API release notes). Fast mode reached Opus 4.7 in research preview, extending the speed tier up from 4.6 at matching pricing. And the built-in web search tool began returning richer SEC filing data, surfacing primary-source citations from filings for financial-research, earnings-analysis, and due-diligence agents. None of it had a launch. All of it is the kind of plumbing that decides whether an agent built on the platform is cheap to run and easy to trust. Read against the Stainless move, it is the same week's other half: own the tooling, then quietly improve the parts of it you already own.

The SDKs got self-hosted sandboxes.

The builder layer moved in lockstep all week. On May 19, Python v0.103.0 and TypeScript v0.97.0 added self-hosted sandbox support in the Compute Management API, with helper utilities for standing your own sandboxes up (GitHub). That is the surface that lets you run agent code in isolation you control rather than renting it, which matters for anyone who can't ship their code to a managed runtime. Point releases the same day taught SessionToolRunner to skip tool calls it doesn't own, closing a dispatch conflict that would otherwise bite multi-agent sessions where two runners fight over the same call. The TypeScript build also picked up Node 26 compatibility. The features Code with Claude London put on stage on May 20, self-hosted sandboxes and MCP tunnels for Managed Agents, had already shipped here the day before the keynote opened.

Both SDKs learned to count thinking tokens.

On May 21, Python v0.104.0 (releases) and TypeScript v0.98.0 (releases) added thinking-token-count beta support on the same day, a parity release surfacing estimated token counts in thinking-block deltas during streaming. The point is budget. Agentic loops that lean on extended thinking have been flying blind on how much context the reasoning is eating until the bill or the context window said otherwise; now the estimate arrives as metadata mid-stream, while there is still time to act on it. A day later, Python v0.104.1 cleaned up the corner the feature exposed: encrypted_content wasn't being carried through the beta compaction accumulator during streaming, which bit anyone running the extended-thinking beta with context compaction on (GitHub). Ship the visibility, then patch the leak it surfaced. The whole week ran on that rhythm.

The bedrock and vertex SDKs got a CI fix.

The least glamorous entry in the log, and worth one line for the people it unblocked. On May 19, bedrock-sdk v0.29.2 (releases) and vertex-sdk v0.16.1 (releases) aligned @types/node across sub-packages to fix CI build failures, with no feature changes. Anyone deploying Claude through Bedrock or Vertex on a pinned TypeScript toolchain had a red pipeline until this landed. No mechanism worth a paragraph, but the blast radius was real, and the move is the obvious one: bump and move on.

Term of the Issue  ·  p.19  ·  First observable 2026-05-18, the week Anthropic acquired the SDK generator behind OpenAI, Google, Cloudflare, and Meta, then chose to wind it down.
Toolmaker/ˈtuːlˌmeɪ.kər/  ·  noun

The party that owns the machinery others use to build, as distinct from the party with the best product. In a frontier race fought over models, the toolmaker move is to acquire the shared dependency, the SDK generator, the compute, the spec-to-library factory, and decide who keeps renting it.

First observable instance: 2026-05-18, the week Anthropic acquired the SDK generator behind OpenAI, Google, Cloudflare, and Meta, then chose to wind it down.

Usage"They didn't ship the better model. They bought the toolmaker."
p.20  ·  Notes

Quiet on
the wire.

No models. No flagship. anthropic.com/research stayed dark after the May 15 leadership paper (coverage), and anthropic.com/news hadn't moved since the May 14 Gates Foundation partnership, $200 million over four years across global health, life sciences, education, and economic mobility, which sat just outside the window (Anthropic). Code with Claude London wrapped its two-day run on May 20, with MIT Technology Review and Fortune both covering it (MIT Technology Review), but the features on stage had already shipped in the SDKs the day before. The docs.anthropic.com release-notes pages returned 403 on every sweep this week, so anything that moved there silently isn't captured here; the three API betas above were confirmed by web search, not by the source. Tokyo is the next stop, June 10.

The Close

No new model. A factory bought and closed. Six patches to make the agent safe to hold. The toolmaker was the story.

Don't miss the next issue

Get Shipped. in your inbox every Friday at 9 AM ET.

One read on what Anthropic shipped this week. Editorial up front, comprehensive Release Log in back. No spam, unsubscribe one click.

Next issue, May 29, 2026
Back of Book  ·  p.21 →

The Release Log

A 1:1 mirror of every Anthropic release in the window. Use it as reference. Share it with your team.

A
Research and Policy
1 Model in window

1 entry in window.

Research

2028: Two scenarios for global AI leadership

Anthropic's research team frames 2028 as the inflection year for US-China AI dominance, placing semiconductors at the center rather than capabilities or safety. Scenario one: democracies defend their compute advantage, tighten chip export controls, and hold a 12-to-24-month frontier lead. Scenario two: policymakers defer, China closes the gap, and AI governance is authored by authoritarian regimes. The paper recommends restricting adversarial-state access to American models and tightening semiconductor supply-chain controls. Published May 15 while Trump made his first trip to Beijing in roughly a decade (coverage).

B
API and Platform
3 Release notes

3 entries in window. All dated May 18 on the API release notes; the docs domain was unreachable during the sweeps and items were confirmed via web search.

API

Cache diagnostics, public beta

Pass diagnostics.previous_message_id on a Messages request and the API returns a cache_miss_reason field explaining where the prompt-cache prefix diverged from the previous turn. Removes the guesswork from debugging multi-turn cache behavior.

How to useInclude beta header cache-diagnosis-2026-04-07. Add diagnostics: { previous_message_id: "<id>" } to the request body.
API

Fast mode for Claude Opus 4.7, research preview

Fast mode now supports Opus 4.7, extending the speed tier up from Opus 4.6. Pricing and rate limits match Opus 4.6 fast mode. Waitlist access via the Console.

How to useSet speed: "fast" with model: "claude-opus-4-7" and the fast-mode-2026-02-01 beta header.
API

SEC filing data in web search tool

The built-in web search tool now returns richer SEC filing data, surfacing primary-source citations from filings for financial-research, earnings-analysis, and due-diligence workflows.

C
Claude Code
5 versions  ·  v2.1.144 → v2.1.149

5 entries in window.

Code

Claude Code v2.1.149

/usage now shows per-category cost breakdown (skills, subagents, plugins, per-MCP-server) instead of a single total. /diff gains keyboard navigation (arrows, j/k, PgUp/PgDn, Space, Home/End). Markdown renders GFM task-list checkboxes. Enterprise gets the allowAllClaudeAiMcps managed setting for claude.ai cloud MCP connectors. Four security fixes: PowerShell permission bypass via built-in cd functions, git-worktree sandbox write allowlist covering the entire repo root instead of scoped paths, PowerShell prefix/wildcard rules pre-approving native executables, and a permission-analysis gap from stale PWD/OLDPWD/DIRSTACK tracking. Also fixes find exhausting the macOS file/vnode table on large trees, managed-settings dialog freezes, and /ultraplan remote-session creation errors.

How to useclaude update. The /usage breakdown and /diff navigation are available immediately. Enterprise admins set allowAllClaudeAiMcps in managed settings.
Code

Claude Code v2.1.148

Critical hotfix: the Bash tool was returning exit code 127 (command not found) on every invocation, a regression introduced in v2.1.147. Any workflow depending on shell command execution was broken until this shipped.

How to useclaude update. If you're on v2.1.147 and Bash commands are failing, this is why.
Code

Claude Code v2.1.147

/simplify renamed to /code-review, with effort-level reporting (/code-review high) and a --comment flag that posts findings as inline GitHub PR review comments. Pinned background sessions (Ctrl+T in claude agents) now survive idle and restart in place when an update lands. 28 additional fixes spanning auto-updater retries, diff-rendering performance, enterprise login enforcement, PowerShell on Windows with winget/Microsoft Store installs, MCP server pagination (resources, templates, and prompts past page 1 were being dropped), shell snapshot user-function handling, plugin-agent multi-type declarations, and hook if condition matching.

How to useclaude update. Run /code-review, or /code-review high for a thorough pass; add --comment to post results to a GitHub PR.
Code

Claude Code v2.1.145

Scripting, observability, plugin discoverability, and a security fix. claude agents --json surfaces live sessions as JSON for external tooling. OTEL gains agent_id and parent_agent_id on claude_code.tool spans with corrected trace parenting for background subagents. Status-line JSON now includes GitHub repo and PR context. /plugin Discover and Browse show a plugin's full manifest before install. Security fix: bare variable assignments to non-allowlisted env vars in Bash were auto-approved past the permission prompt, now blocked. Also fixes MCP error messaging for missing required arguments, a spinner freeze after terminal resize, the Windows PowerShell ; separator, task-list ordering on concurrent creation, a /review GraphQL regression on Classic Projects repos, and a skill context: fork infinite loop.

How to useclaude update. Script sessions: claude agents --json | jq '.[] | .id'
Code

Claude Code v2.1.144

Released Monday evening. Resume support for background sessions, elapsed-duration notifications when agents complete, and model selection scoped to individual sessions. Fixes startup hangs when the API is unreachable, terminal display corruption, and macOS background-session crashes in protected folders.

How to useclaude update or npm i -g @anthropic-ai/claude-code.
E
Agent SDKs
8 SDK releases

8 entries in window.

SDK · PY

anthropic-sdk-python v0.104.1

Bug fix: encrypted_content was not being carried through the beta compaction accumulator during streaming. Affects users of the extended-thinking beta with context compaction enabled.

How to usepip install anthropic==0.104.1
SDK · PY

anthropic-sdk-python v0.104.0

Adds thinking-token-count beta support: when streaming extended thinking, the SDK surfaces estimated token counts in thinking-block deltas. Useful for cost estimation and context-budget management in agentic loops that lean on extended thinking.

How to useInclude the thinking-token-count beta header. Token estimates arrive as metadata on thinking-block delta events during streaming.
SDK · TS

anthropic-sdk-typescript v0.98.0

Parity release: the same thinking-token-count beta as Python v0.104.0 (releases). Both SDKs shipped the capability the same day.

How to usenpm install @anthropic-ai/[email protected]. Include the thinking-token-count beta header.
SDK · PY

anthropic-sdk-python v0.103.0

Adds support for self-hosted sandboxes in the Compute Management API (CMA) with new sandbox helper utilities.

How to usepip install anthropic==0.103.0
SDK · PY

anthropic-sdk-python v0.103.1

Bug fix: SessionToolRunner skips tool calls it does not own, preventing dispatch conflicts in multi-agent sessions.

How to usepip install anthropic==0.103.1
G
News and Partnerships
1 Items

1 entry in window.

News

Anthropic acquires Stainless

Anthropic acquired Stainless (founded 2022, backed by Sequoia and a16z), which generates SDKs, CLIs, and MCP servers from API specifications. Stainless generated every official Anthropic SDK to date and also generates the Python, TypeScript, Go, Java, Kotlin, and Ruby libraries used by OpenAI, Google, Cloudflare, and Meta. Post-acquisition, all hosted Stainless products, including the SDK generator, wind down; existing customers retain their generated SDKs with full modification rights. Terms undisclosed; The Information reported talks valuing the company at over $300 million, with part potentially in Anthropic equity (TechCrunch).