Preface¶

On July 24, 2026, Anthropic officially launched Claude Opus 5 (API identifier claude-opus-5). This represents a significant upgrade to the Opus series: it features native 1 million Token context window, adaptive thinking enabled by default, and maintains the same pricing as the previous generation Opus 4.8: $5 / $25 per million Tokens (input / output).

For Claude Code users, this release is more than just a model name change. The Default model for Max, Team Premium, Enterprise pay-as-you-go, and API accounts has been resolved to Opus 5; seats such as Pro and Team Standard still use Sonnet 5 as their default. Meanwhile, Claude Code continued iterating on capabilities like Subagent background execution, nested delegation, /code-review background review, and /doctor self-check and repair throughout July, making discussions around developer toolchains particularly active.

This article is organized based on Anthropic’s official release notes, Claude Platform release logs, and Claude Code changelog, sorting out the core specifications of Opus 5 and its connection with the July updates to Claude Code.

Core Specifications of Claude Opus 5¶

According to Anthropic’s official documentation, the main parameters of Opus 5 are as follows:

Compared with Opus 4.8, Opus 5 has improved performance in coding and knowledge work benchmarks such as Frontier-Bench and CursorBench. The official positions it as a daily flagship model that “approaches the cutting-edge capabilities of Fable 5 but costs only half as much”, and it is also the default model for Claude Max.

One behavioral change deserves attention during migration: when using xhigh or max effort, thinking cannot be turned off; if an API request carries thinking: {"type": "disabled"}, a 400 error will be returned. It is still possible to disable thinking at high and lower levels.

What Does 1 Million Token Context Mean?¶

Opus 5’s 1 million Token window is both the default value and the upper limit, and no extra charge applies for long context lengths — according to official instructions, a request with 900,000 Tokens will be billed at the same unit price as a request with 9,000 Tokens.

For Claude Code scenarios, this means:
1. System prompts, tool definitions, conversation history, and the previous round’s thinking blocks all count towards the context window, and thinking blocks are retained by default and billed as input in subsequent rounds.
2. Full indexing of large monorepos, cross-file refactoring, and long-session debugging no longer require a separate 1M beta header.
3. Sonnet 5 also has native 1 million Token context (promotional price: $2 / $10 per million Tokens, valid until August 31, 2026), so Pro tier users can already use long context for daily work; Opus 5 further raises the bar for reasoning depth and agent stability.

In Claude Code, you can explicitly specify the 1M window using the model alias, for example:

> /model opus[1m]

Or use the full model name directly:

> /model claude-opus-5

Model Division of Labor in Claude Code¶

Claude Code’s Default model varies by account type, which is the most confusing point in July discussions. According to the Claude Code model configuration documentation:

Therefore, “1 million Token context becomes the default flagship model of Claude Code” should be understood as: in high-tier accounts such as Max / API, the Default has switched to Opus 5 with 1 million Token context; Pro tier users still default to Sonnet 5, but can switch to Opus 5 at any time via /model.

Version requirements:
- Opus 5: Claude Code v2.1.219 or higher
- Sonnet 5: v2.1.197 or higher

If the version is lower than v2.1.219, the opus alias will continue to resolve to Opus 4.8, and Opus 5 may not be visible in the interface. Upgrade command:

claude update

Regarding Fast Mode, the Claude Code changelog notes that Opus 4.7 has been removed from fast mode support; /fast now applies to Opus 5 and Opus 4.8, and the fast mode pricing for Opus 5 is $10 / $50 per million Tokens.

July Claude Code Toolchain Iteration¶

Around the launch of Opus 5, Claude Code has a series of consecutive updates to agent workflows. All the following content comes from the official changelog and the 2026 Week 27 Summary.

Subagent Runs in Background by Default¶

Starting from v2.1.198, Subagent runs in the background by default: the main session does not need to block and wait, and Claude will receive a notification to continue after the subtask is completed. If the subtask result must be returned before the next step, Claude will still run that Subagent in the foreground.

You can fix the behavior using the background field in the agent frontmatter. The permission prompts for background Subagents will be summarized to the main session, and you can press Esc to individually reject a certain tool call.

Nested Subagent¶

The updates in mid-to-late July further relaxed the nesting depth: Subagent can now spawn up to three layers of child agents by default (total depth 3, previously the default was 1). You can tighten this via an environment variable:

export CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1

In headless / stream-json scenarios, setting --forward-subagent-text or CLAUDE_CODE_FORWARD_SUBAGENT_TEXT=1 will forward the output of Subagents at depth 2 and deeper to the main stream, making it easier for CI or IDE plugins to observe.

There is also a concurrency limit: up to 20 concurrently running Subagents by default, which can be overridden with CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS; the default total spawn limit per session is 200, and you can reset the budget with /clear.

/code-review Background Code Review¶

The latest changelog entries state that /code-review has been changed to run as a background Subagent. Code reviews no longer occupy the main conversation context, and can retain the review target specified by the stacked slash command. For example:

> /code-review medium

In non-interactive mode, /code-review ultra will start a cloud review instead of silently downgrading to a local single pass. If you only need a quick single-round review, you can still use /review.

/doctor Self-Check and Repair¶

/doctor (alias /checkup) has evolved into a full environment health check: it can diagnose installation channels, startup scripts, MCP connections, settings conflicts, CLAUDE.md file size and other issues, and directly propose fixes in some scenarios (such as auto mode default values, npm global installation unable to self-update, etc.).

When encountering TUI exceptions caused by Subagent nesting, Remote Control session state desynchronization, or misparsed Windows path \u segments, the official changelog also recommends running:

> /doctor

Other July Highlights¶

• Claude in Chrome GA (v2.1.198): The Chrome integration exited preview, and Claude Code can drive the browser for end-to-end testing.
• Claude Sonnet 5 becomes the default for Pro and other seats (v2.1.197+): 1 million Token context + promotional pricing, forming a tiered structure of “default Sonnet, flagship Opus” alongside Opus 5.
• Organization default model: Administrators can set the Default in the org console, and it will display as “Org default” in /model when an individual has not selected a model.

How Developers Get Started with Opus 5¶

1. Confirm Claude Code Version¶

claude --version

If the version is lower than 2.1.219, run claude update. Restart the terminal after upgrading to avoid the alias cache still pointing to Opus 4.8.

2. Select Model and Effort¶

Max / API users usually do not need to make changes, as the Default is already Opus 5. Pro users can switch manually:

> /model claude-opus-5
> /effort high

You can try xhigh or max for long-cycle agent tasks; note that thinking cannot be turned off at high effort levels, and Token consumption will also increase significantly.

3. Enable Fast Mode as Needed¶

For scenarios sensitive to latency and willing to pay twice the Token price:

> /fast on

The fast mode status will automatically follow the model’s capabilities after switching models (the changelog has fixed the issue where “fast mode remains off after switching back to a model that supports fast mode”).

4. Organize Complex Tasks with Subagent¶

Large-scale refactoring can be assigned to built-in agents such as Explore, code-reviewer, or custom markdown definitions under .claude/agents/. Nested depth and background running make “main session planning + parallel execution of multiple Subagents” the default working mode for Claude Code in July; the backgroundization of /code-review has also separated code reviews from the main conversation.

5. API Side Migration¶

If your project calls the Claude API directly, just start using the dated model ID claude-opus-5. Fast mode can be enabled on the API via speed: "fast" and the corresponding beta header; fast mode support for Opus 4.7 has been removed, and you need to migrate to Opus 5 or Opus 4.8.

Summary¶

Claude Opus 5 combines 1 million Token context, default thinking, and Opus tier pricing, and has become the default flagship choice for Claude Code in accounts such as Max / API; Pro tier users still default to Sonnet 5, but both product lines now natively support 1 million Token windows.

Combined with the July Claude Code iterations on Subagent background execution, three-layer nesting, /code-review background review, and /doctor self-check, the developer toolchain is shifting from “single-session question and answer” to a pattern of “main agent orchestration + parallel execution of multiple Subagents + long context reasoning”. If you are already using Claude Code, first check whether the version is ≥ 2.1.219, then confirm whether the /model resolution matches your subscription tier — this will better avoid the hidden gap of “thinking you are using Opus 5, but actually still using Opus 4.8”.