Claude Code August 2026: Sonnet 5 Default, Fork Subagents, Dataviz Skill, Auto Draft PRs

Sonnet 5 as default — what changes in practice

Sonnet 5 became the default Claude Code model in August 2026 with a 1M-token native context window. The previous default (Opus 4.8) had a 200K context ceiling before chunking was required.

For solo projects, the difference is marginal. For large codebases, it is significant: a 1M-token context holds approximately 750,000 words of code, or the entirety of most mid-sized production codebases including test files, documentation, and configuration. The practical effect is fewer mid-session context drops and more coherent reasoning about cross-file dependencies.

Pricing context: Sonnet 5 is positioned as the workhorse model — priced below Opus 5 while delivering competitive coding performance. Check the current rate card at the Anthropic pricing page; pricing for Claude Code models has been updated since the initial Sonnet 5 launch.

Fork subagents on by default (v2.1.232)

The biggest behavioral change for power users: subagent_type: "fork" is now the default for non-teammate agent spawns in interactive sessions. A fork subagent inherits:

• The full conversation history up to the fork point
• The active prompt cache (no re-tokenization cost)
• The current working directory and file context

What this means in practice: when you ask Claude Code to “spin up an agent to handle the tests while I keep working on the main logic,” it forks the conversation cleanly rather than starting a fresh context. The fork sees exactly what you see at that moment.

Toggle: CLAUDE_CODE_FORK_SUBAGENT=0 to restore the previous behavior (fresh context per agent spawn).

This is additive — existing workflows that explicitly set subagent_type to other values continue working. The change only affects the default when no explicit type is set.

The /dataviz skill (v2.1.198)

The /dataviz skill is the first officially bundled skill specifically for data visualization and chart design. Activate it with /dataviz in any session.

What it does:

1. Chart type selection — given a description of your data and goal, recommends the appropriate chart type with a rationale (bar vs line vs scatter vs treemap vs etc.)
2. Color palette guidance — generates an accessible palette with contrast ratios validated for WCAG 2.1 AA
3. Runnable color-palette validator — this is the most technically specific part: it ships a small validator script that checks your palette programmatically rather than relying on visual inspection
4. Dashboard layout principles — for multi-chart dashboards, provides layout patterns and hierarchy guidance

The color-palette validator fills a real gap in existing dataviz workflows. Most chart libraries let you plug in arbitrary hex values with no feedback on whether the resulting combination is accessible or visually coherent. The validator catches this before it reaches a PR.

Background agents auto-commit and open draft PRs (v2.1.198)

Previously, background agents in worktrees finished their work and stopped, leaving changes uncommitted. You had to manually commit, push, and open a PR.

As of v2.1.198: when a background agent running in a git worktree finishes code work, it automatically:

1. Stages all changes (git add -A equivalent)
2. Commits with a generated message describing what was done
3. Pushes the branch to origin
4. Opens a draft PR against the base branch

The draft PR state is intentional — the agent signals “review required” rather than “ready to merge.” The description field of the auto-opened PR includes the agent’s summary of what it did and why, pulled from the conversation context at the point of completion.

This behavior is on by default for worktree sessions. No configuration required if you use claude agents with worktrees. If you prefer to keep the manual commit step, set CLAUDE_CODE_AUTO_PR=0.

Cross-session @ mentions

Type @session-name in a prompt to route a message to another named Claude session. Claude uses the internal SendMessage tool to deliver it and incorporates the response inline.

Practical use: “Ask @backend-agent if the /users endpoint has been migrated yet, then continue with the frontend implementation based on its answer.” The two sessions communicate mid-task without manual copy-paste.

Session naming: use /name <session-name> at the start of a session to give it a persistent name within the current Claude Code process. Named sessions are visible to other sessions in the same process.

Task-tracking tools removed from newer models

The TaskCreate, TaskGet, TaskUpdate, TaskList, and TodoWrite tools are no longer available by default on Opus 4.8, Sonnet 5, Fable 5, and newer models.

To re-enable: set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 in your environment.

The rationale: for most workflows, the overhead of explicit task tracking tools added latency and token cost without meaningful benefit over the model’s in-context planning. Teams using these tools heavily (particularly in multi-session orchestration workflows) should set the env var and continue as before.

Related

• Claude Skills Directory — the growing ecosystem of skills that extend what Claude Code can do, including the new /dataviz skill context
• Agent Plugins 1.0 — the open standard that codifies how skills distribute across agents