Shipped. Weekly  ·  Anthropic  ·  ISO Week 24, 2026
Shipped.
The week Anthropic removed the ceiling on what it would sell, then put $350 million behind the argument that the disruption is real.
Week of 2026-06-08 Published 2026-08-06 Lab Anthropic Releases 16 items
The Open
Front of book
Capability, conscience, and channel. In that order.

Monday the week looked like a distribution story. Apple opened WWDC 2026 with an Extensions framework in iOS 27 that lets a user pick Claude, Gemini, or ChatGPT as the engine behind Siri, Writing Tools, and Image Playground. Gemini holds the native default on a reported $1 billion annual contract. Claude arrives as something you install and then go choose in Settings. Reach, not position.

By Tuesday night that framing was gone. Fable 5 went into general release, not preview and not waitlist, at 80.3% on SWE-Bench Pro against a field whose next best was 69.4%. By Thursday Anthropic had committed $350 million across a nonprofit fellowship and an economic research fund, and Dario Amodei had published tiered policy responses to AI-driven unemployment that run all the way to universal basic income. By Friday DXC Technology had made Fable 5 the default model inside the platform it sells to banks, airlines, insurers, and government agencies.

Three moves, three different surfaces, five days. A company does not spend $350 million explaining what its product might do to employment unless it believes the product works.

Lead Story01
Anthropic  ·  June 09, 2026  ·  Models

The Mythos
Ceiling

Anthropic put the first Mythos-class model into general release and shipped the same weights twice: once with a governor, once without. Which one you get depends on who you are.
Primary source anthropic.com/news
Coverage TechCrunch, GitHub Changelog, AWS News, June 09, 2026
Status Generally available. Free on paid plans through June 22.
By the numbers 80.3% on SWE-Bench Pro. Next best: 69.4%.

29.3% on FrontierCode Diamond, against 13.4% for Opus 4.8.

38.6% spatial reasoning, against 14.5%.

Fewer than 1 session in 20 hits the safeguard route.

$10 and $50 per million tokens, in and out.
Lead Story  ·  The Mythos Ceiling

Claude Fable 5 and Claude Mythos 5 landed on June 9 as one release with two doors. Fable 5 is the public model: 80.3% on SWE-Bench Pro, 29.3% on FrontierCode Diamond against 13.4% for Opus 4.8, and spatial reasoning at 38.6% against 14.5%, which is not an improvement so much as a different instrument. Stripe put the practical number on it. A 50-million-line Ruby migration estimated at two-plus months of engineering time finished in a day.

Mythos 5 is the same model with some safeguards lifted, distributed through Project Glasswing to authorized cyberdefenders and infrastructure providers. Fable is Mythos with the routing on: roughly one session in twenty touches high-risk territory, cybersecurity or biology, and falls back to Opus 4.8. That is the whole product decision, stated plainly in the release. The public gets the model with a governor. The defenders get it without.

TechCrunch noted that Anthropic shipped its most powerful public model within days of warning that AI is getting too dangerous. Both things are true and they are not in tension, because the safeguard routing IS the answer to the warning. Whether one session in twenty is the right number is not a question anyone outside Anthropic can currently check, and that is the part worth holding onto.

Distribution followed the same day. GitHub Copilot made Fable 5 selectable for subscribers. Amazon Bedrock carried it at launch, with Vertex and Foundry to follow. Nine SDK releases went out across Python and TypeScript in a single afternoon to add the identifiers, server-side refusal fallbacks, and a client-side fallback middleware for providers that do not support them. A model launch is now a supply-chain event.

The Dig02
Anthropic  ·  June 10 to 11, 2026  ·  Policy / News

$350 million is a number
you say out loud

Two commitments in two days, and a policy essay that names the unemployment thresholds at which its author thinks a government should act.
Primary source anthropic.com/research
Also Claude Corps, June 11, 2026
Reporting Washington Post, US News, June 11, 2026
The Dig  ·  The mitigation number

On June 10 Anthropic committed $200 million to an Economic Futures Research Fund and published a Dario Amodei policy essay proposing tiered government responses to AI-driven unemployment at 5%, 10%, and beyond, escalating to universal basic income and equity-sharing frameworks. The same package proposed an Advanced AI Framework with mandatory testing, independent evaluation, and civil penalties tied to global revenue for developers training above 10^25 FLOPs.

On June 11 came Claude Corps: $150 million placing 1,000 early-career workers inside US nonprofits for a year at $85,000 each, trained on Claude, run with CodePath and Social Finance, with more than 400 nonprofits hosting in year one. No degree required. Over 18, under two years of full-time experience, authorized to work in the US. The first cohort of 100 starts October 19.

This is the first time Anthropic has attached a dollar figure to economic mitigation alongside its capability agenda, and it did it in the same week it removed the ceiling on that agenda. A company proposing civil penalties tied to its own revenue, on a compute threshold its own models cross, is not lobbying for looser rules. It is trying to write the rules while it still has the standing to.

Also Shipped
Front of book
News  ·  June 08, 2026
An Extensions framework in iOS 27 lets users designate Claude, Gemini, or ChatGPT as the model behind Siri, Writing Tools, and Image Playground. Gemini is the native default on a reported $1 billion annual Apple to Google contract; Claude arrives as a user-installed extension selected in Settings under Apple Intelligence. iOS 27, iPadOS 27, and macOS 27 ship fall 2026. No Anthropic press release accompanied it, which tells you which side of the deal this is.
Apps  ·  June 09, 2026
A Swift package making Claude available as a server-side language model inside Apple's own framework, conforming to the LanguageModel protocol so it can be driven with LanguageModelSession alongside Apple's on-device model. Multi-step reasoning, code generation, web search, code execution, SwiftUI streaming. Announced June 9, available June 10. The Siri extension is the consumer door; this is the developer one.
API  ·  June 09, 2026
Agents can now be attached to a cron schedule on the Claude Platform and fire without a manual trigger. Credential vault integration keeps secrets out of model context entirely, binding keys at the network boundary only for requests to customer-allowlisted domains. No separate scheduler fee, billed through existing platform usage. Public beta, reported by itbrief.news.
News  ·  June 11, 2026
DXC tested Claude across its 115,000-person workforce in 70 countries before rolling it to clients. Claude wrote more than 95% of the code for DXC OASIS, its AI-native orchestration platform, with Fable 5 as the default foundation model. DXC is training tens of thousands of Claude-certified forward-deployed engineers to install Claude inside systems that banks, airlines, insurers and government agencies have run for decades. Not a product launch. An established firm putting Claude at the center of what it sells.
The Close  ·  Two beats
The week opened with Claude as a setting you could choose inside someone else's assistant.
It closed with Claude as the default model inside the software that runs banks.
One deprecation sat quietly under all of it. claude-sonnet-4-20250514 and claude-opus-4-20250514 retired on June 15, three days after this week closed, and anything still pinning either identifier began returning errors that Sunday. Worth naming the failure mode, because it is easy to read a deprecation line as housekeeping: a retired model returns 404 not_found, not 401. Anything watching for an auth failure sees a healthy credential and a missing target, and spends its afternoon in the wrong place.
Back of Book  ·  Reference

The Release Log

A 1:1 mirror of every Anthropic release in the window, June 08 to June 12, 2026. Use it as reference. Share it with your team.

Models
1 release
The first Mythos-class model in general release, and its unrestricted twin.
Model
Claude Fable 5 and Claude Mythos 5
Model ID claude-fable-5, 128,000 output tokens, $10 and $50 per million in and out. SWE-Bench Pro 80.3% against a 69.4% field. FrontierCode Diamond 29.3% against 13.4%. Spatial reasoning 38.6% against 14.5%. Safeguards fire in under 5% of sessions and fall back to Opus 4.8. Free on Pro, Max, Team and Enterprise through June 22, then usage credits. Mythos 5 is the same model with some safeguards lifted, limited to Project Glasswing partners.
Claude Code
7 releases
Seven versions in five days, ending on enterprise model governance.
Code
v2.1.175 and v2.1.174
v2.1.175 adds the enforceAvailableModels managed setting: when enabled, the availableModels allowlist also governs what Default resolves to, closing the gap organizations kept finding. v2.1.174 adds wheelScrollAccelerationEnabled and fixes the /model picker hiding the family Default resolves to.
Code
v2.1.173
Fable 5 model names carrying a [1m] suffix are now normalized automatically, since Fable 5 ships with 1M context by default and the suffix was causing match failures. Fixed a spurious sandbox-dependencies-missing warning on Windows. Thinking summaries in collapsed groups render as markdown, hold for at least 3 seconds, and cap at 10 lines.
Code
v2.1.172
Sub-agents can spawn their own sub-agents, up to 5 levels deep, on by default. Bedrock reads the AWS region from ~/.aws when AWS_REGION is unset, matching AWS SDK precedence. Search bar in the plugin marketplace. Fixed sessions using 1M context without usage credits getting permanently stuck: they now auto-compact under the standard limit.
Code
v2.1.170 and v2.1.169
v2.1.170 adds Fable 5 access and fixes transcript saving for sessions launched from the VS Code integrated terminal. v2.1.169 adds --safe-mode and CLAUDE_CODE_SAFE_MODE to disable all customizations for clean troubleshooting, /cd to move a session without breaking the prompt cache, and disableBundledSkills. Ten bug fixes across enterprise MCP policy, Windows startup, and remote reattachment.
Agent SDKs
10 releases
Nine in one afternoon to carry the model launch, plus one that crossed into Monday.
SDK
Python v0.108.0 to v0.109.1, TypeScript v0.103.0 to v0.104.1
v0.108.0 and v0.103.0 add the claude-fable-5 and claude-mythos-5 identifiers, server-side refusal fallbacks, and client-side fallback middleware. v0.109.0 and v0.104.0 add Managed Agents deployments and environment variable credentials. v0.109.1 and v0.104.1 add the frontier_llm refusal category.
SDK
Platform SDKs: vertex v0.17.1, foundry v0.3.1, bedrock v0.30.2, aws v0.4.2
All four fixed third-party middleware ordering, released within sixty seconds of each other at 16:36 to 16:37 UTC.
SDK
anthropic-sdk-python v0.107.1
One bug fix carried into Monday's window: the Foundry client was not sending the x-api-key header for API-key authentication.
News and Partnerships
5 items
Two distribution wins, two commitments, one alliance.
News
DXC Technology multi-year global alliance
Fable 5 as the default foundation model on DXC OASIS, more than 95% of whose code Claude wrote. Tens of thousands of Claude-certified forward-deployed engineers in training.
News
Claude Corps
$150 million placing 1,000 fellows inside US nonprofits for one year at $85,000 each plus benefits. First cohort of 100 starts October 19, applications closed July 17. No degree required. Run with CodePath and Social Finance, 400-plus nonprofits hosting.
News
Economic Futures Research Fund and policy frameworks
$200 million for trials and policy evaluations on AI's effects on work, alongside a Dario Amodei essay proposing tiered responses at 5%, 10%, and unprecedented unemployment levels, and an Advanced AI Framework with civil penalties tied to global revenue above 10^25 FLOPs.
News
Fable 5 on GitHub Copilot and Amazon Bedrock
Copilot subscribers can select Fable 5 in the model picker as of June 9. AWS carried it at launch, described in its own announcement as Mythos-class capabilities with built-in safeguards.
News
Apple WWDC 2026: iOS 27 Extensions
Users can designate Claude, Gemini, or ChatGPT as the AI model for Siri, Writing Tools, and Image Playground. Gemini native default. Ships fall 2026.
Claude Apps
1 release
A Swift package, announced Tuesday and available Wednesday.
Apps
Claude for Apple Foundation Models framework
Conforms to the LanguageModel protocol so developers drive it with LanguageModelSession alongside Apple's on-device model. Multi-step reasoning, code generation, web search, code execution, SwiftUI streaming. Targets iOS 27, iPadOS 27, macOS 27, visionOS 27, watchOS 27.
Stay on the frontier

Get Shipped. in your inbox.

Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.