Claude Code Summer 2026 Update: New Models, Background Agents by Default

The Models: Sonnet 5 and Opus 5

The headline change is the default model. Claude Sonnet 5 shipped with a native 1M-token context window and became the default for new sessions — no flag, no opt-in, no Max/Team/Enterprise gate to unlock the larger window. Claude Opus 5 followed a few weeks later as the new top-tier model, also with 1M context, and its fast mode now runs noticeably faster than the outgoing Opus 4.8 fast mode at the same relative cost multiplier.

If you've been pinning a specific model in settings.json via modelOverrides or ANTHROPIC_MODEL, that override still wins — Claude Code won't silently move you off a pinned model. If you haven't pinned anything, you're already on Sonnet 5 without doing a thing.

What actually changes for you

• Long sessions stop needing Max/Team/Enterprise for 1M context. If part of your workflow was "switch to Opus 4.6 when the conversation gets huge," that step is gone — the default model already has the headroom.
• /fast behavior shifted. Fast mode now applies to Opus 5 and Opus 4.8; Opus 4.7 was dropped from the fast-mode roster. If you scripted around a specific fast-mode model, check it still resolves.
• Org-restricted models still apply. If your organization pins a default model or restricts the picker, that takes priority over the global default — /model will show "Org default" or "Role default" when that's the case.

Subagents: Background by Default, Three Levels Deep

This is the change most people will feel without reading a single changelog entry. Subagents launched from your main session now run in the background automatically. You no longer watch a spinner while a research agent chews through fifteen files — it runs off to the side, and you get pinged when it's done.

The second half of this change is nesting. A subagent can now spawn its own subagents up to three levels deep, up from one. That means a top-level agent coordinating a refactor can delegate a sub-task to a helper agent, which can itself delegate a narrower sub-task, without you manually re-parenting anything.

Turning it off (partially or fully)

# Cap nesting depth back to 1 (disables deeper nesting)
export CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1

# Session-wide guardrails that now exist regardless of depth:
#   - default cap of 200 subagent spawns per session
#   - default cap of 200 WebSearch calls per session
export CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION=200
export CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION=200

Long-running MCP tool calls also move to the background automatically now — anything over two minutes gets bumped off the foreground by default, tunable with CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS. If you have an MCP tool that legitimately takes longer and you were relying on it blocking the session, raise that threshold explicitly.

/fork Moved, /code-review Moved, and /subtask Showed Up

Three related commands changed shape this summer, and they're easy to mix up if you haven't touched them recently.

/fork now copies to a background session

/fork (alias /branch) used to branch your current conversation inline. Now it copies the conversation into a brand-new background session instead, leaving your current window untouched while the fork runs independently. If what you actually wanted was a quick, disposable subagent inside your current session — not a whole new session — that's what /subtask is for. It replaces the old in-session subagent behavior that used to live under /fork.

/code-review runs as a background subagent

/code-review no longer fills up your main conversation with review output. It now runs as a background subagent and reports back, which matters if you were used to watching the review stream in real time — you'll want to check /tasks or wait for the completion notification instead. /ultrareview also got a bug fix: it used to choke on descriptive arguments like "review my auth changes" instead of a PR number, and that now works as expected.

Claude in Chrome: Now Generally Available

Claude in Chrome graduated out of preview. If you were on the waitlist or testing it behind a flag, it's simply on now — no separate opt-in required. Combined with background agent notifications (the Notification hook now fires agent_needs_input and agent_completed events from claude agents), you can kick off a background agent, switch to browsing, and get pulled back in only when Claude actually needs you.

Auto Mode Is the New Default (On Some Platforms)

Auto mode — the classifier that automatically approves low-risk commands instead of prompting for every single one — is now available without opt-in on Bedrock, Vertex AI, and Foundry. If you'd rather keep the older prompt-heavy behavior, set disableAutoMode in your settings. Auto mode's safety net also got stricter this summer: destructive git commands and infrastructure-destroy operations (Terraform, Pulumi, CDK) are blocked under auto mode unless you explicitly ask for them, and every auto-mode denial now shows up in /permissions and the transcript so you can see exactly why something was refused.

Separately — and easy to conflate with the above — the default interactive permission mode was renamed from default to Manual across the CLI, --help, VS Code, and JetBrains. Same behavior, new label. If your documentation or onboarding scripts reference the old name, it's cosmetic but worth a find-and-replace.

Smaller Additions Worth Knowing About

/dataviz — a built-in skill for charts and dashboards

A new skill ships with color-palette validation and layout guidance for anything chart- or dashboard-shaped — useful if you've been hand-rolling color scales for every analytics page you build.

Screen reader mode

Launch with claude --ax-screen-reader, set CLAUDE_AX_SCREEN_READER=1, or add "axScreenReader": true to settings. It adds structured announcements for terminal state, and now also announces deleted text on word and line deletions — a detail that matters a lot if you're navigating by ear.

Tighter sandbox controls

Two new settings worth knowing if you run sandboxed commands: sandbox.network.strictAllowlist denies any network host that isn't explicitly allowlisted, with no prompt fallback. sandbox.filesystem.disabled lets you skip filesystem isolation while keeping network egress control in place, for cases where the filesystem sandbox was getting in the way but you still want network restrictions.

DirectoryAdded hook

If you script around working directories, there's now a hook that fires after /add-dir or an SDK register_repo_root call — useful for keeping external tooling in sync with whatever directories a session has registered.

What to Actually Change Today

• Nothing, if you don't want to. All of this is designed to be invisible by default. The background-first behavior and the model upgrade both apply automatically.
• Revisit any hardcoded model references in scripts, CI configs, or modelOverrides if you were explicitly targeting Opus 4.6 or 4.7 for the 1M context window — you likely don't need to anymore.
• If you rely on synchronous subagents (watching output live instead of getting notified), get comfortable with /tasks and the new agent notifications, or set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 if nested spawning is actively unwanted in your setup.
• If you automated /verify or /code-review as an implicit end-of-task step, make it explicit again — it no longer fires on its own.
• If you're on Bedrock, Vertex, or Foundry and prefer the old per-command prompting, set disableAutoMode before auto mode surprises you mid-task.