Week 27  ·  ISO 2026-27  ·  Anthropic
Weekly review, compiled from daily sweeps
2026-06-29
to 2026-07-03
Published 2026-07-04
27
Shipped.

Three clouds closed. One model promoted.

Claude Sonnet 5 became the default for every Free and Pro user. Fable 5 returned after 19 days of export control. Claude Code shipped six releases and flipped to Manual mode. Agent memory is three weeks out.

New Default
Sonnet 5
Days Exiled
19
Code Releases
6
Clouds Covered
3
Open The thread that ran through the week
The enterprise rails closed. The default model shifted. And Claude Code asked to stop and ask.

Monday morning, June 29. Claude landed in Microsoft Azure Foundry as generally available: Opus 4.8 and Haiku 4.5 in the Messages API, NVIDIA GB300 Blackwell Ultra GPUs, Azure-native identity and billing, a US data zone for teams with residency requirements. Not a preview.

That same week, the Claude Apps Gateway shipped: a self-hosted control plane for Claude Code deployments, running as a stateless container backed by PostgreSQL, routing inference to the Claude API, Amazon Bedrock, or Google Cloud with optional failover between them. Org-level SSO, policy enforcement, per-user cost attribution, and spend limits. The gateway was built to cover Bedrock and Google Cloud. Azure Foundry added the third leg. As of this week, Anthropic has enterprise-grade deployment across all three major hyperscalers with a self-hosted control plane spanning the set.

Then Sonnet 5 dropped Tuesday morning and became the default for every Claude Free and Pro user on the same day. An agentic model with a one-million-token context window, near-Opus 4.8 benchmark performance on tool use and long-horizon tasks, introductory pricing at $2 input and $10 output per million tokens through August 31, 2026. After that, $3 and $15. That window closes in eight weeks.

This week
Jun 29 to Jul 03, 2026

Releases
26 confirmed items

Leads
Sonnet 5
Fable 5 returns
Claude Science

Signal
Agent memory: Jul 22
Lead Story
Models  ·  June 30, 2026

The Default Shift

Claude Sonnet 5 launched June 30 and became the model of record for every Free and Pro user on the same day. The agentic tier is now the everyday tier.

Key Figures $2 / $10 per MTok (through Aug 31)

$3 / $15 per MTok (standard)

1M token context window

128k max output tokens

Model ID: claude-sonnet-5
Lead Story

Sonnet 5 is positioned as an agentic model. The word "agentic" does a lot of work in Anthropic's framing this week: built for planning, tool use, multi-step tasks, and autonomous operation at a level that until recently required Opus-class compute and Opus-class pricing. On agentic benchmarks it runs close to Opus 4.8. The context window is one million tokens, native. Max output is 128,000 tokens.

The default change is the decision worth studying. Anthropic made Sonnet 5 the default for Free and Pro users on launch day. Not after a week of bake-in. Day one. That is a product bet: that the user base is ready for an agentic model as their everyday baseline, that the capabilities are stable enough to lead with, that the pricing is defensible enough to not require a paywall for most users.

The introductory pricing reinforces the bet. $2 per million input tokens through August 31 looks like a deliberate land-and-expand move: lock in long-horizon agent workflows at Sonnet-tier pricing before the standard rates kick in. Builders who find that Sonnet 5 covers their use case at $2/$10 will stay at $3/$15 rather than re-engineer. The window is real. It closes in eight weeks.

Available day one across all surfaces: Claude API with model ID claude-sonnet-5, claude.ai as the new default for Free and Pro subscribers, Claude Code after updating to v2.1.197, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The system card published the same day. One constraint worth noting: Priority Tier is not available for Sonnet 5. If your workloads depend on priority throughput, plan accordingly.

The fast-mode deprecation announced the same day sets the context. Fast mode for claude-opus-4-6 now runs at standard speed and standard billing. Fast mode for claude-opus-4-7 is deprecated, with removal scheduled July 24. Anthropic's guidance: migrate to claude-opus-4-8, where fast mode costs three times less than on prior Opus versions. The architecture is moving forward. The cheap option is the new one.

On all platforms
Claude API
claude.ai (new default)
Claude Code v2.1.197+
Amazon Bedrock
Google Cloud
Microsoft Foundry

Note
Priority Tier not available for Sonnet 5
Also Shipped The rest of the week, in full
News  ·  Jul 01
Fable 5 returns, 19 days later

Nineteen days. That is how long export controls kept Claude Fable 5 off the global market. The restrictions went on June 12 without ceremony. They came off July 1 the same way. Anthropic redeployed it across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork the same day the US reversed the order.

Pro, Max, Team, and select Enterprise plans get access within weekly usage limits capped at 50% through July 7. After that, usage credits apply. Export-controlled frontier AI is no longer an abstraction. The ceiling test for Anthropic's most capable public models just ran for three weeks in public view.

Apps  ·  Jun 30
Claude Science

A desktop research workbench, now in beta on macOS and Linux for Pro, Max, Team, and Enterprise subscribers. A coordinating agent connects to more than 60 curated skills and connectors pre-configured for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Code runs sandboxed. Results save as versioned artifacts with full provenance. Source: anthropic.com/news/claude-science-ai-workbench.

Anthropic is funding up to 50 research projects with $30,000 in compute credits each. Applications close July 15; projects run September 1 to December 1. Life sciences is the declared second vertical after developer tooling. The product mirrors Claude Code: domain-specific, autonomy-forward, designed for long-horizon tasks from concise instructions.

News  ·  Jul 01
Project Glasswing: 150 new orgs

The security program added 150 organizations across 15+ countries, extending into sectors not well represented in the initial cohort: power, water, healthcare, communications, and hardware. Partners have collectively found more than 10,000 high or critical severity security flaws. Source: anthropic.com/news/expanding-project-glasswing.

Alongside Amazon, Microsoft, and Google, Anthropic proposed a cross-industry jailbreak severity rubric built on four criteria: capability gain, breadth of capability gain, ease of weaponization, and discoverability. The rubric is a proposal for a shared standard, not a policy. The coordination signal is the news.

Claude Code  ·  Jun 29 to Jul 03
Six releases, one architectural shift

v2.1.196 through v2.1.201 shipped this week across five days. Source: github.com/anthropics/claude-code/releases. The architectural change came Thursday in v2.1.200: the default permission mode flipped to Manual across CLI, VS Code, and JetBrains. Claude stops and asks before acting, where it previously did not. AskUserQuestion dialogs no longer auto-continue by default. Opt back in via /config.

Earlier in the week, v2.1.198 brought Claude in Chrome to general availability, added background agent lifecycle events (agent_needs_input and agent_completed) for programmatic monitoring, automated draft PR creation from background agents, and a new /dataviz skill for chart and dashboard design. All week Claude Code was getting more autonomous. Thursday it put on the brakes.

News  ·  Jun 29
Claude in Microsoft Foundry: generally available

Claude is generally available in Microsoft Azure Foundry as of June 29, running on NVIDIA GB300 Blackwell Ultra GPUs. Claude Opus 4.8 and Claude Haiku 4.5 are in the Messages API with prompt caching and extended thinking. Azure-native authentication, billing, and governance are built in. A US data zone is available for teams with data residency requirements. Source: anthropic.com/news/claude-in-microsoft-foundry.

With Foundry GA, Anthropic's enterprise footprint covers all three major cloud providers: AWS Bedrock, Google Cloud, and Azure Foundry. The Claude Apps Gateway, launched the same week, provides a unified self-hosted control plane across all three.

Quiet on the Wire
Signal, not noise

Both SDKs, Python v0.116.0 and TypeScript v0.110.0, shipped July 2 with a single addition each: support for the agent-memory-2026-07-22 beta header. Anthropic names its beta headers by activation date. The date embedded in that string is July 22, 2026. July 22 is 20 days out from this issue's publish date.

No public documentation. No announcement. The SDKs are wired and waiting. Managed Agents persistent memory appears to be three weeks from landing. Watch the Python and TypeScript changelogs on and after July 22.

The Close
The enterprise rails are closed.
The science workbench is open.
Manual mode is the default now, but the agents already know how to commit their own work.
· · ·
Back of Book

The Release Log

Every confirmed Anthropic release in the window: June 29 to July 3, 2026. Use it as reference. Share it with your team.

Models
1 release
MODEL
Claude Sonnet 5
New mid-tier flagship and the default model for all Claude Free and Pro users. Introductory pricing: $2/$10 per million tokens through August 31, 2026, then $3/$15 standard. Native 1M-token context window. 128k max output. Agentic performance close to Opus 4.8 on planning, tool use, and multi-step tasks. System card published same day. Priority Tier not available.
How to use: API model ID: claude-sonnet-5 (or claude-sonnet-5-20260630). New default on claude.ai. Update Claude Code to v2.1.197 or later. Available on Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
Why it matters: The agentic tier is now the mass-market default. The introductory pricing window closes August 31.
API & Platform
1 release
API
Fast mode removed for Opus 4.6; Opus 4.7 deprecated
Requests to claude-opus-4-6 with speed: "fast" now run at standard speed and are billed at standard rates. No error is returned. Fast mode for claude-opus-4-7 is deprecated; removal is scheduled July 24, 2026. After July 24, fast requests to Opus 4.7 return an error.
How to use: Migrate fast-mode Opus 4.6 and 4.7 workflows to claude-opus-4-8. Fast mode on Opus 4.8 costs three times less than on prior Opus models.
Claude Code
6 releases
Six builds in five days. v2.1.200 is the architectural one.
CODE
Claude Code v2.1.201
Sonnet 5 sessions no longer inject mid-conversation system role messages for harness reminders.
CODE
Claude Code v2.1.200
Default permission mode changed to "Manual" across CLI, VS Code, and JetBrains. AskUserQuestion dialogs no longer auto-continue by default. Eight bug fixes: background sessions no longer silently stop after sleep/wake; background-agent daemon handover and roster issues resolved; subagents cut off by rate limits fixed; tmux 3.4+ rendering flicker fixed; startup crash on non-array disabledMcpServers / enabledMcpServers fixed; screen-reader output and install script messaging improved.
How to use: claude update or reinstall. Manual mode takes effect immediately. To restore idle auto-continue, run /config and enable the timeout option.
CODE
Claude Code v2.1.199
Stacked slash-skill invocations: /skill-a /skill-b your prompt loads all leading skills, up to five. SSL certificate errors fail immediately with fix hints instead of burning retries. Streaming responses no longer discard partial output on mid-stream errors. Background-agent daemon corruption on Linux fixed. macOS SSH background agent sessions fixed. claude stop no longer undone by background-agent respawn.
How to use: claude update. Stacked skills syntax: /skill-a /skill-b do XYZ
CODE
Claude Code v2.1.198
Claude in Chrome is now generally available. Background agents gain two lifecycle events: agent_needs_input and agent_completed, exposed in the SDK for programmatic monitoring. Background agents now auto-commit, push, and open draft PRs rather than waiting for user input. New /dataviz skill for chart and dashboard design. Syntax highlighting upgraded to highlight.js 11.
How to use: claude update to get v2.1.198.
CODE
Claude Code v2.1.197
Sonnet 5 is now the default model. Org-configured default models, readable session names, clickable file attachments, and a smoother agents view. Background session reliability and Remote recovery improved. Streaming idle watchdog on by default for all providers, aborting and retrying if no events arrive in five minutes. Org-configured model restrictions appear in the model picker with a "restricted by organization" label.
How to use: claude update or reinstall.
CODE
Claude Code v2.1.196
Org admins can set default models in the console, surfaced as "Org default" or "Role default" in the model picker. Sessions get readable names on start. File attachments are clickable in chat (Cmd/Ctrl-click opens in Finder or Explorer). MCP security tightened: self-approved repo servers no longer auto-spawn in untrusted workspaces. Streaming idle watchdog on by default. Per-frame render work cut by roughly 37%.
Key fixes: Background session crash on transcript probe. Rate-limit warning flickering. MCP OAuth scope requests on enterprise IdPs now request only required scopes. PowerShell git diff / git grep false-positives on exit 1. Keyboard focus stuck in agents panel. Set CLAUDE_ENABLE_STREAM_WATCHDOG=0 to disable the watchdog.
Claude Apps
2 releases
APPS
Claude Science (beta)
A desktop application for scientific research, in beta on macOS and Linux for Pro, Max, Team, and Enterprise subscribers. A coordinating agent connects to 60+ curated skills and connectors pre-configured for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Claude writes and runs Python, R, or shell code in a sandboxed environment, with versioned artifacts and full audit trails. Up to 50 research grant projects receiving $30,000 in credits each; applications open through July 15, 2026.
How to use: Access at claude.com/science. Available on Pro, Max, Team, and Enterprise plans. Discounted seats for academic and nonprofit labs.
Why it matters: Life sciences is Anthropic's declared second vertical after developer tooling.
APPS
Claude Apps Gateway for Amazon Bedrock and Google Cloud
A self-hosted control plane for Claude Code deployments on enterprise cloud infrastructure. Runs as a single stateless container on Linux, backed by PostgreSQL. Gives organizations: corporate SSO login, centrally enforced policy, role-based access, per-user cost attribution, and daily/weekly/monthly spend limits per org, group, or user. Routes inference to the Claude API, Bedrock, or Google Cloud with optional failover. Does not forward inference traffic or usage data to Anthropic unless configured to use the Claude API directly.
How to use: See the Claude apps gateway docs. Deploy as a container, configure upstream provider credential, set policies and spend limits.
Why it matters: With the gateway covering Bedrock and Google Cloud, and Claude in Microsoft Foundry going GA the same week, Anthropic has enterprise-grade deployment across all three major hyperscalers with a self-hosted control plane for each.
Agent SDK, Python
5 releases
SDK-PY
anthropic-sdk-python v0.116.0
Added support for the agent-memory-2026-07-22 beta header. SDK-level plumbing for Managed Agents persistent memory, ahead of the July 22 feature rollout.
How to use: pip install --upgrade anthropic
SDK-PY
anthropic-sdk-python v0.115.1
Removes nonfunctional types that shipped in v0.115.0.
How to use: pip install --upgrade anthropic
SDK-PY
anthropic-sdk-python v0.115.0
Managed Agents event delta streaming, agent overrides, reverse pagination, vault credential injection scoping, and agent and deployment webhook events.
How to use: pip install anthropic==0.115.0
SDK-PY
anthropic-sdk-python v0.114.0
Support for claude-sonnet-5. Agent toolset now allows absolute paths that resolve inside workdir.
How to use: pip install anthropic==0.114.0
SDK-PY
anthropic-sdk-python v0.113.0
Adds SDK-level support for the web fetch and support tools introduced in API version 20260318. Fixes a missing output_format / output_config merge block in async count_tokens. Accepts user profile IDs when counting tokens.
How to use: pip install anthropic==0.113.0
Agent SDK, TypeScript
8 releases
SDK-TS
anthropic-sdk-typescript v0.110.0
Added support for the agent-memory-2026-07-22 beta header. Parallel to the Python v0.116.0 release.
How to use: npm install @anthropic-ai/sdk@latest
SDK-TS
anthropic-sdk-typescript v0.109.1
Removes nonfunctional types from v0.109.0. Parallel patch to Python v0.115.1.
How to use: npm install @anthropic-ai/sdk@latest
SDK-TS
anthropic-sdk-typescript v0.109.0
Managed Agents event delta streaming, agent overrides, reverse pagination, vault credential injection scoping, and agent and deployment webhook events. Parallel to Python v0.115.0.
How to use: npm install @anthropic-ai/sdk@0.109.0
SDK-TS
anthropic-sdk-typescript v0.108.0
Support for claude-sonnet-5. Agent-toolset absolute path fix. Parallel to Python v0.114.0.
How to use: npm install @anthropic-ai/sdk@0.108.0
SDK-TS
@anthropic-ai/vertex-sdk v0.19.0
google-auth-library bumped to v10.2.0.
How to use: npm install @anthropic-ai/vertex-sdk@latest
SDK-TS
@anthropic-ai/bedrock-sdk v0.32.0
Client logger passed to AWS credential provider chain.
How to use: npm install @anthropic-ai/bedrock-sdk@latest
SDK-TS
@anthropic-ai/aws-sdk v0.6.0
Client logger passed to AWS credential provider chain. Parallel to bedrock-sdk v0.32.0.
How to use: npm install @anthropic-ai/aws-sdk@latest
SDK-TS
anthropic-sdk-typescript v0.107.0
Web fetch and support tools added for API version 20260318. Agent-toolset symlink hops are now bounded and glob matches validated before confinement checks. BatchCreateParams.Request.params, dropped by a codegen merge, is restored. Token counting accepts user profile IDs.
How to use: npm install @anthropic-ai/sdk@0.107.0
News & Partnerships
3 items
NEWS
Redeploying Claude Fable 5 globally
Claude Fable 5 is available globally as of July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. The US reversed the export restrictions imposed June 12. On Pro, Max, Team, and select Enterprise plans, usage counts toward up to 50% of weekly limits through July 7; after that, usage credits apply.
Why it matters: The restrictions lasted 19 days. Frontier AI subject to export control is the new ceiling test for Anthropic's most capable public models.
NEWS
Project Glasswing expansion
150 new organizations across 15+ countries join the security program, extending into power, water, healthcare, communications, and hardware. Partners have surfaced more than 10,000 high or critical severity security flaws. Anthropic proposed a cross-industry jailbreak severity rubric alongside Amazon, Microsoft, and Google, built on four criteria: capability gain, breadth of capability gain, ease of weaponization, and discoverability.
NEWS
Claude in Microsoft Foundry: generally available
Claude Opus 4.8 and Claude Haiku 4.5 are generally available in Microsoft Azure Foundry, running on NVIDIA GB300 Blackwell Ultra GPUs. Azure-native authentication, billing, and governance built in. US data zone available. The "Hosted on Anthropic infrastructure" option remains in preview; the "Hosted on Azure" option is GA.
How to use: Access through your Azure subscription. See platform.claude.com/docs/en/build-with-claude/claude-in-microsoft-foundry for setup.
Why it matters: Anthropic's enterprise footprint now covers all three major cloud providers. The Claude Apps Gateway, announced the same week, provides a unified self-hosted control plane across all three.
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.