Week 28  ·  ISO 2026-28  ·  Anthropic
Weekly review, compiled from daily sweeps
2026-07-06
to 2026-07-10
Published 2026-07-11
28
Shipped.

On the same Friday OpenAI asked you to hand over the work, Anthropic asked if you should.

Five versions of Claude Code, a government scan of 466 million lines, a question submitted by 52,000 people, and an undocumented SDK flag called dreaming. That is week 28.

Confirmed Releases
11
Code Versions
5
Lines Scanned
466M
Countries Surveyed
159
Open The thread that ran through the week
Friday, noon. OpenAI asked you to hand over the next six hours. Anthropic asked if you wanted to.

ChatGPT Work launched on July 10: give the model a task, step away, get back a finished document. The pitch is agency transfer at scale. OpenAI has shipped three agent products in twelve months to reach this one, and the thesis underneath all of them is the same: the assistant becomes the assistant when it is good enough to be left alone.

Anthropic's answer on the same Friday was Reflect, a beta feature that shows you your usage patterns and asks what you want to keep doing yourself. Not a product positioned against ChatGPT Work. A product positioned against the direction of travel the whole industry is running in, ChatGPT Work very much included. The campaign that launched with it, Inviting Hard Questions, put the 52,000-person survey on the table. The concerns documented: job loss, devaluing creative work, reduced human agency, diminished ability to think independently. Anthropic committed to reporting back on its responses, including failures.

Underneath all of that, the week was infrastructure and maintenance. Five Claude Code versions, Monday through Thursday night. The Alberta government case study on Monday: 466 million lines scanned in 20 hours using a two-agent security review system. Cowork expanded to web and mobile, with Microsoft 365 write operations and FedRAMP High government authorization arriving on the same day Fable 5's complimentary subscription window closed. By Friday evening, a TypeScript SDK update shipped an undocumented flag called "dreaming." Nobody explained it. The documentation will follow. It always does.

This week
Jul 06 to Jul 10, 2026

Releases
11 confirmed items

Lead
Reflect, Hard Questions

Also shipped
Alberta 466M lines
Cowork web, mobile
Fable 5 pricing
Claude Code x5

Signal
Dreaming flag: watch API notes
Lead Story
Apps, News  ·  July 10, 2026

The Hard Question

Anthropic shipped a feature that shows you your patterns and a campaign asking what you want to keep doing yourself, on the same day OpenAI launched a product designed to take those things off your plate.

Campaign by the numbers 52,000 Americans surveyed

81,000 Claude users

159 countries, 70 languages

4 documented concern areas

Film via agency Mother

Public: claude.com/hard-questions
Lead Story

The Reflect feature is simple to describe. Enable Memory in Claude settings. Open Settings. Navigate to Reflect. The feature surfaces your usage history as a set of patterns: what topics you bring to Claude most, what hours you reach for it, how the distribution shifts across the month. It adds a reflective prompt above the data. The published example: "What's one thing you want to keep doing yourself, even if Claude could do it faster?" Available in beta on Free, Pro, and Max plans. Incognito chats and health-integration conversations are excluded by design.

That is the product. The positioning is a different story.

The Inviting Hard Questions campaign shipped alongside it. A film commissioned through agency Mother. A public question submission site at claude.com/hard-questions. Research from 52,000 Americans surveyed and 81,000 Claude users across 159 countries and 70 languages. The documented concerns: job loss, devaluing creative work, reduced human agency, and diminished ability to think independently. Anthropic committed publicly to tracking and reporting specific actions it takes in response, including areas where it falls short.

The accountability framing is unusual. Companies that build AI announce capabilities and announce safeguards. They rarely announce that they surveyed 81,000 of their own users about whether the product is doing harm, found four categories of documented concern, and promised to answer for them. That framing is either sincere or a very expensive marketing strategy. The two are not mutually exclusive, and the market will form its own opinion on which it is.

The timing is what makes the read precise. When OpenAI launched ChatGPT, Anthropic published Constitutional AI papers. When GPT-4 launched, Anthropic launched Claude 2 and a character card. When ChatGPT Work launched (on this same Friday, in the same news cycle), Anthropic launched Reflect and a hard-questions campaign. You can read that sequence as Anthropic being reactive, or as Anthropic being consistent. Either read lands in the same place: Anthropic has decided that the distinctive thing it can offer is not a faster agent or a cheaper output tier. It is a company willing to run a 52,000-person survey about whether its product is good for you, publish what it found, and then answer for the parts where it wasn't.

Whether that bet pays off commercially is a separate question. The bet is placed. Week 28 made it visible.

ChatGPT Work
agency transfer

Reflect
agency inspection

One tells you the direction of travel. The other asks if you've thought about where you're going.
Also Shipped Four items, each with consequences
Quiet on the Wire
What is dreaming?

TypeScript SDK v0.111.0 shipped Friday afternoon with an API flag called "dreaming." The implementation: gates session tool calls on evaluated_permission, bounds idle behavior by server stop_reason. No public documentation shipped with it. The naming and permission-gating suggest this is an opt-in capability requiring explicit grant rather than a flag available to all callers by default.

This is the same sequencing Anthropic used for the advisor-tool beta and the agent-memory beta: SDK ships the implementation, documentation follows within days to weeks. "Dreaming" is the third undocumented capability in recent months to arrive this way. Watch the API release notes at docs.anthropic.com/en/release-notes/api.

The Close
Five code versions. A government scan. A reflective question on a Friday.
Anthropic spent this week fixing the floor and asking if the ceiling matters.
The dreaming flag is in the SDK. Nobody has explained it yet.
·
Back of Book  ·  Week 28

The Release Log

All confirmed Anthropic releases in the July 6 to July 10, 2026 window, grouped by category.

News
2items
Government partnership and accountability campaign.
NEWS
Government of Alberta: 466 million lines in 20 hours
Anthropic case study from Alberta's Ministry of Technology and Innovation. Two-agent security review system using Claude Code with Opus and Sonnet: red team probing exploit paths, blue team writing remediation plans to exact files. 466 million lines scanned, vulnerabilities remediated, new tooling built. 20 hours total. 20x speed improvement, up to 95% reduction in modernization timelines reported. Industry day in Edmonton shares the playbook; province-wide scale-out starts fall 2026.
Why it mattersFirst public government-scale case study with Claude Code as the primary tool, and Alberta is sharing the playbook openly.
NEWS
Inviting Hard Questions campaign
Film commissioned via agency Mother. Public submission site at claude.com/hard-questions. Research from 52,000 Americans surveyed and 81,000 Claude users across 159 countries and 70 languages. Documented concerns: job loss, devaluing creative work, reduced human agency, diminished ability to think independently. Anthropic commits to tracking and reporting specific actions in response, including failures.
Claude Apps
3releases
Cowork, Government Desktop, and Reflect all shipped in a single week.
APPS
Cowork expanded to web and mobile
Sessions now run in the cloud by default. Work continues when laptop closes or device switches. Chat and Cowork share one home tab in the Claude interface. Microsoft 365 integration deepened: draft and send email from inside Claude, manage calendar events, create and update OneDrive and SharePoint files. Entra admin consent required for write operations. Doubled usage limits extended through 2026-08-05.
How to useAvailable on Claude web and mobile. For Microsoft 365 write operations, workspace admins must grant Entra consent. Usage limits doubled through August 5.
APPS
Claude for Government Desktop, public beta
FedRAMP High authorized. Local conversation history stored on agency-managed devices. Department-level administration with spending and model limits. Tamper-evident audit logs.
How to useContact Anthropic for Government access. Requires agency-managed device configuration. Audit logs available for compliance review.
Why it mattersFedRAMP High is the authorization level required for US federal agencies handling controlled unclassified information. This is not a pilot.
APPS
Reflect, beta
Usage pattern analytics inside Claude settings. Requires Memory enabled. Shows monthly topic distribution, most active day, peak usage hour, and patterns across sessions. Includes reflective prompts. Available on Free, Pro, and Max plans. Incognito and health-integration conversations excluded by design.
How to useEnable Memory in Claude settings, then open Settings and navigate to Reflect. Available immediately after Memory is enabled.
Models
1item
MODEL
Fable 5: end of complimentary subscription access
Fable 5 access on Pro, Max, Team, and select Enterprise plans ended effective 2026-07-08. Previously included at no extra cost. Now requires usage credits at $10 per million input tokens and $50 per million output tokens. No formal announcement preceded the change.
How to useTeams running Fable 5 workflows on subscription plans should audit current usage and add usage credits in the billing dashboard before the next billing cycle.
Claude Code
5releases
v2.1.202 through v2.1.206. Fifty-plus fixes. The floor, repeatedly.
CODE
Claude Code v2.1.202
Dynamic workflow size config (workflow.size: small, medium, large), two new OpenTelemetry attributes (workflow.run_id and workflow.name), 14 bug fixes. Fixes include: Ctrl+R history search crash on accept or cancel, /rename reversion on job restart, transient mTLS handshake failures during certificate rotation, Remote Control "Unknown command" errors in interactive sessions, unicode quote corruption in workflow scripts, unbounded voice dictation retry loop on audio failures, and MCP error clarity for servers with url but no type.
How to useRun claude update to get the release. New /config workflow.size option takes small, medium, or large. Advisory, not a hard agent cap. OpenTelemetry traces now emit workflow identity for better activity reconstruction.
CODE
Claude Code v2.1.203
30-plus reliability fixes. Critical: macOS background session stall regression introduced in v2.1.196, causing 15 to 20 second freezes in background sessions (the kind users diagnose as their own network, not the tool). Second critical fix: stale daemon session tokens producing permanently unresponsive background sessions. Memory savings of approximately 7 MB each on binary size and startup.
How to useRun claude update immediately. If you experienced background session stalls or macOS slowdowns in the prior week, this release is the fix.
Why it mattersA regression that makes users blame themselves is worse than a regression that makes the tool fail loudly. v2.1.203 fixed both categories.
CODE
Claude Code v2.1.205
20-plus stability fixes. Critical: Windows worktree deletion bug that could remove files outside the worktree via NTFS junctions. Silent message loss when a message arrived at the max-turns limit. Peak memory on auto-updates cut by approximately 400 MB via streaming binary downloads.
How to useRun claude update or reinstall. Windows users running multi-worktree setups should update before their next session. The NTFS junction bug was data loss risk.
CODE
Claude Code v2.1.206
New: /doctor now proposes trimming oversized CLAUDE.md files. /commit-push-pr auto-allows push to the repo's configured push remote, not only to origin. Background agents auto-upgrade after Claude Code version updates, resolving version drift across long-running sessions. EnterWorktree asks for confirmation before entering worktrees outside .claude/worktrees/. /cd shows directory path suggestions. Fixes: expired login handling, MCP configuration errors, model picker display, desktop sessions stuck on "running" after slash commands.
How to useRun claude update or reinstall. If you have oversized CLAUDE.md files, run /doctor to get trim suggestions. Teams using non-origin git remotes: the /commit-push-pr fix resolves the prior need to manually whitelist them.
Agent SDKs
1release
SDK-TS
anthropic-sdk-typescript v0.111.0
Adds API support for "dreaming" capability flag. Gates session tool calls on evaluated_permission, bounds idle behavior by server stop_reason. Field description updates and model examples. No public documentation for "dreaming" published at time of release.
How to useInstall: npm install @anthropic-ai/sdk@0.111.0. "Dreaming" documentation expected within days to weeks. Watch docs.anthropic.com/en/release-notes/api.
Why it mattersAnthropic has used this same sequencing for the advisor-tool beta and agent-memory beta: SDK first, documentation follows. "Dreaming" is the third such capability in recent months. Watch this one.
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.