Back to Blog
claude-code weekly-update release-notes sonnet-5 fable-5 background-agents mcp security automation community agent-patterns events

What's Up Claude Code: Week of June 29th

Olivier Legris ·

TL;DR

This was the week Claude Code stopped being a one-model story.

Claude Sonnet 5 is now the default in Claude Code as of v2.1.197: 1M-token context, stronger coding / tool-use performance, and promo API pricing of $2/M input and $10/M output through August 31. A day later, Fable 5 was redeployed after export controls were lifted, but with a very visible operational caveat: through July 7 it is included for up to 50% of weekly usage limits on eligible plans, then it moves to usage credits, and cyber-risky requests may route away from Fable to Opus 4.8.

Meanwhile, Claude Code itself shipped a lot: organization default models, safer .mcp.json approval behavior, background agents on by default, Claude in Chrome generally available, better transient-error retries, manual permission mode becoming the default wording, and a long list of fixes for background-agent daemon state.

The catch: the issue tracker is already showing the next set of sharp edges — silent model fallback, background-agent message desync, cross-talk, stuck processes, and pinned-model leakage on resumed subagents. The practical takeaway is boring and important: use the best model for the job, but design the harness as if models, resumes, and background workers can drift.

1. Sonnet 5 becomes the Claude Code default

The most important release note is short: Claude Code v2.1.197 introduced Claude Sonnet 5 as the default model in Claude Code.

Anthropic’s launch post calls Sonnet 5 its “most agentic Sonnet yet,” available across Claude, Claude Code, and the Claude Platform, with a native 1M-token context window and introductory API pricing of $2/M input tokens and $10/M output tokens through August 31, 2026 before moving to $3/$15. Source: Introducing Claude Sonnet 5.

For Claude Code users, the interesting part is not just raw benchmark movement. It is where Sonnet 5 sits in the workflow:

  • It is strong enough to be the everyday default for coding-agent loops.
  • It is cheap enough, at promo pricing, to make parallel subagents and long-context exploration more reasonable.
  • It is still not obviously the best choice for every high-effort task.

That last point matters. The latent.space AINews roundup captured the immediate community split: people liked the agentic/coding gains, but worried about effective cost-per-task because Sonnet 5 can use more output tokens and more turns than older Sonnet models or Opus at high effort.

My read: Sonnet 5 is the new sensible default, not a reason to stop routing. Use it for the broad middle of coding work; reserve Opus/Fable-style models for planning, review, risky refactors, and hard diagnosis.

2. Fable 5 is back, but the product story is now routing and credits

Anthropic also published Redeploying Claude Fable 5. The export-control restriction on Fable 5 and Mythos 5 was lifted on June 30, and Fable 5 became available again on July 1 across Claude Platform, Claude.ai, Claude Code, and Claude Cowork.

The details are what Claude Code users should actually care about:

  • Eligible Pro, Max, Team, and select Enterprise plans get Fable 5 for up to 50% of weekly usage limits through July 7.
  • After that, Fable 5 is available via usage credits.
  • Updated cyber safeguards can route some requests away from Fable to Opus 4.8.
  • Anthropic says biology / chemistry classifiers remain broad for now, and the redeployment includes work on a shared industry jailbreak-severity framework.

The smol.ai July 1 issue had the right framing: the interesting story is less “Fable is back” than how builders adapt when frontier access is conditional. Several examples converged on multi-model orchestration: use Fable for high-value planning or reasoning, then hand implementation and verification to cheaper / more available models.

That is the design pattern to steal. Treat Fable as a scarce specialist, not a blanket default. If a tool can spawn 18 subagents, it can also burn through credits faster than a human notices.

3. Claude Code release train: permissions, MCP, background agents, and Chrome

The week had a dense run of CLI releases from v2.1.196 through v2.1.201. The high-signal changes:

v2.1.196: safer MCP listing and org defaults

v2.1.196 added organization default models and a meaningful MCP security fix: claude mcp list / get no longer spawn .mcp.json servers that a repo self-approved via committed .claude/settings.json; untrusted workspaces now show ⏸ Pending approval.

That is the right direction. Listing MCP configuration should not be enough to execute repo-provided server definitions.

v2.1.198: background agents become the default path

v2.1.198 made subagents run in the background by default, shipped Claude in Chrome as generally available, added background-agent notifications via the Notification hook, added the /dataviz skill, added Claude Platform on AWS as a Gateway upstream, and changed claude agents-launched code work so background agents can commit, push, and open draft PRs from worktrees when they finish.

That is a big product-direction signal: Claude Code is becoming a multi-session workbench, not just a terminal chat.

v2.1.199 and v2.1.200: reliability cleanup after the background-agent push

v2.1.199 fixed stacked slash-skill invocation, SSL/proxy failures burning retries, partial streaming output being discarded on server errors, subagents falsely reporting API errors as success, background daemon crashes, claude stop races, hook stderr hiding, and more. It also made transient 429s retry with backoff for subscribers and raised watchdog behavior for non-capacity transient errors.

v2.1.200 changed AskUserQuestion dialogs so they no longer auto-continue by default, accepted manual as the explicit/default permission mode name, and fixed a pile of background-agent daemon / roster / stale-lock / sleep-wake issues.

This is a healthy but revealing pattern: once background agents become default, every edge in daemon state, transcript state, socket auth, stale locks, and resume semantics becomes user-facing.

v2.1.201: harness reminder cleanup

v2.1.201 is tiny but telling: Sonnet 5 sessions no longer use the mid-conversation system role for harness reminders. That should reduce some weirdness where the harness itself feels like it is intruding into the conversation.

4. The bug cluster to watch: model drift and background-agent boundaries

The public issue tracker is noisy, but the cluster is clear: background agents and model routing are now the hot path.

Notable reports:

  • #74598: subagent model pins can allegedly be lost on injected wake / resume turns, so a resumed subagent may run and bill at the waker’s model rather than the model it was spawned with.
  • #74637: on v2.1.201, queued user messages and background-agent notifications can allegedly desync, with assistant turns starting before queued message delivery and messages being silently removed.
  • #74653: a background subagent reportedly returned another conversation’s content as its first completion, ignored its task prompt, and made no tool calls.
  • #74638: experimental agent teams / background teammates reportedly fail to terminate cleanly, leaking processes until the user kills them manually.
  • #74646: the desktop “working” indicator can allegedly stay on after all processes finish because stale background-task IDs remain in local storage.
  • #74691: a user reports silent downgrade from claude-fable-5 to Opus without user-visible notification.
  • #74685: ScheduleWakeup in --print / headless mode can accept a wakeup, claim it is scheduled, then exit with no persistent session to fire it.
  • #74692: a Fable 5 / v2.1.201 report of assistant output leaking past the turn boundary and hallucinating the next user: message inside the assistant text.

Do not overread any single issue before Anthropic triages it. But do take the class seriously.

Practical mitigations:

  1. Treat model choice as state that can drift. Log it, surface it, and re-check it before expensive runs.
  2. Do not run high-cost Fable subagent fanout without concurrency and spend limits.
  3. Prefer explicit worktrees and clean branch state for background agents.
  4. For scheduled / headless automation, verify that the process model actually persists. A tool saying “scheduled” is not enough.
  5. For long-running agents, track OS processes separately from UI status.

5. Community projects worth watching

A few useful pieces from the broader agent-tooling ecosystem this week:

  • OpenWiki — LangChain’s CLI for generating and maintaining agent-consumable codebase documentation. This fits a real Claude Code pain: agents do better with maintained, inspectable memory than with random transcript archaeology.
  • SkillComposer and the related SkillsBench benchmark — the research direction is explicit skill selection/composition instead of dumping every tool and instruction into the context. Very relevant for Claude Code plugin and skill hygiene.
  • Devin Security Swarm — Cognition’s security workflow uses a swarm / Agentic MapReduce pattern to fan out bounded agents across a codebase, validate exploitability, and aggregate findings. Even if you never use Devin, the architecture is the point.
  • Weaviate Engram — production-grade agent memory keeps moving from “vector search over notes” toward governed, reconciled, shared memory. Claude Code teams will need the same pattern for project memory and permission-aware context.

The through-line: agent tooling is moving from prompt tricks to operating systems for agents — memory, skill routing, evals, security, and process control.

6. smol.ai / latent.space community signal: orchestration beats model maximalism

The AINews signal this week matched the Claude Code release train almost too neatly.

  • Sonnet 5 today, and Fable 5 tomorrow captured the launch split: Sonnet 5 is broadly useful and production-friendly, but token/turn usage makes effective cost more complicated than the sticker price.
  • smol.ai’s July 1 issue framed Fable 5’s return as a routing story: cyber safeguards, Opus fallbacks, usage-credit economics, and builder strategies that reserve Fable for high-value planning.
  • The same issue highlighted agent infrastructure themes: wiki memory, skill composition, Agentic MapReduce, and standardized AI-flaw reporting via FLARE-AI.

The lesson for Claude Code readers: the model is only one layer. The durable advantage is in the harness — model routing, permission boundaries, memory, evals, recovery, logs, and spend control.

7. Community events

The Luma sync ran before this post and PR #54 has been merged, so the site event data is current.

This week (Jul 6–12):

Next week (Jul 13–19):

Later highlights: Berlin, Seattle, San Diego, Taipei, Kolkata, Tokyo, Adelaide, Toronto, Osaka, Mishima, and more are now in the events data for late July / early August.

Host your own on Luma →

8. What I’m watching

  1. Whether Sonnet 5 stays the practical default after the promo period — cost-per-solved-task matters more than price-per-token.
  2. Whether Fable usage credits make it a specialist model — especially for code review, planning, and security workflows.
  3. Whether Anthropic fixes the background-agent boundary bugs quickly — model pinning, queued messages, cross-talk, stale process state, and wake semantics are now core product reliability issues.
  4. Whether MCP approval and plugin loading keep tightening — the v2.1.196 .mcp.json change is good, but repo-provided tools remain a supply-chain boundary.
  5. Whether skill composition becomes mainstream — Claude Code’s plugin / skills layer needs selection, auditing, and routing, not just accumulation.