Foreword¶

In late July 2026, Anthropic rolled out multiple version updates for Claude Code consecutively. As of July 24, the latest stable version is 2.1.220. Version 2.1.219 introduced the two most critical changes of this cycle: switching the default model to Claude Opus 5, and relaxing the default nested Subagent depth from 1 layer to 3 layers. Version 2.1.218 changed /code-review to run as a background Subagent. In early August, multiple evaluation institutions compared Claude Code + Opus 5 with OpenAI’s GPT-5.6-Sol + Codex on benchmarks like Terminal-Bench, further heating up the “arms race” for AI programming Agent models and orchestration layers.

Claude Code is Anthropic’s AI programming Agent for terminal scenarios, supporting reading and writing code in local shells, calling tools, and dispatching Subagents to complete subtasks. Compared with previous generations, the focus of this round of updates is not simply about “switching to a stronger model”, but rather advancing three aspects together: model capabilities, Agent orchestration, and workflow isolation. This article sorts out the key changes of version 2.1.220 and related releases based on the official CHANGELOG and public evaluation materials, and provides actionable configuration instructions.

Version Timeline: 2.1.218 → 2.1.220¶

According to the Claude Code official CHANGELOG, the consecutive releases in late July can be understood as follows:

Version 2.1.220 itself is mainly for bug fixes, but external discussions usually use “2.1.220 with default Opus 5 + 3-layer Subagent” as the overall label, because the functional changes in 2.1.219 are the actual upgrade points. The following content will be expanded by functional dimensions, without deliberately distinguishing patch numbers.

Claude Opus 5 Becomes the Default Model¶

Version 2.1.219 introduced claude-opus-5 and set it as the default model in the Opus series. The key points of the official description are as follows:
- Context Window: 1M tokens
- Fast Mode Pricing: $10 / $50 per million tokens for input / output
- Migration Path: The claude-api skill now points to Opus 5 by default, and there is official guidance for migration from Opus 4.8
- Fast Mode Adjustments: Opus 4.7 no longer supports fast mode; the /fast command now applies to Opus 5 and Opus 4.8

For daily users, no manual configuration changes are needed after upgrading, and new sessions will use Opus 5 by default. If a project is sensitive to cost or latency, you can still select other Opus variants via /model, or explicitly specify the model ID in API calls.

Horizontal Comparison with GPT-5.6-Sol¶

In early August, NeuralCoreTech’s evaluation article summarized the benchmark data of mainstream AI programming Agents at that time. It should be noted that: The following Terminal-Bench scores come from the “pure model” test of third-party institution Artificial Analysis, not the official product-level rankings of Claude Code or Codex, and they had not yet appeared on the tbench.ai public Agent leaderboard as of the publication date.

The gap between the two on Terminal-Bench is less than 0.5 percentage points. The choice mainly depends on workflow habits, subscription cost structure, and Agent orchestration layers (Subagent, MCP integration, etc.) rather than a single score.

Nested Subagent Depth Expanded to 3 Layers¶

Subagent is the core mechanism for Claude Code to split complex tasks into parallel or serial subtasks. The main Agent dispatches Subagents through the Agent tool. Each Subagent executes in an independent context and returns the results after completion.

Depth Changes and Historical Background¶

According to the official CHANGELOG, version 2.1.219 adjusted the default nested depth from 1 to 3:

Subagents can now spawn nested subagents up to depth 3 by default (was 1)

This means:
- Depth 0: Main chat Agent
- Depth 1: Subagent dispatched by the main Agent
- Depth 2: Sub-subagent dispatched by a Subagent
- Depth 3: Third-level nesting (new default limit)

Prior to this, Claude Code had made multiple adjustments to the nesting strategy: early versions allowed deeper nesting, version 2.1.213 once tightened the default depth to 1 and added concurrency limits, and version 2.1.219 reopened it to 3 layers under controllable conditions. Anthropic’s intention is clear – it leaves room for a three-level pipeline like “main Agent → review Agent → line-by-line verification Agent” for scenarios such as Code Review and multi-file refactoring, while avoiding unrestricted recursion.

Supporting Restriction Mechanisms¶

Having only a “depth of 3” is not enough. The official also provided multi-layer safety brakes:
1. Concurrent Subagent Limit: Up to 20 running at the same time by default, which can be overridden via the environment variable CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS
2. Per-Session Dispatch Limit: 200 times by default, which can be adjusted via CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION; /clear can reset the budget
3. Dynamic Workflow Size Guidance: Version 2.1.219 introduced the workflowSizeGuideline setting, with the default value of medium (recommend fewer than 15 Agents)

How to Disable or Limit Nesting¶

If your team wants to maintain a flat structure and avoid overly long Subagent chains, you can add the following configuration in settings.json:

{
  "env": {
    "CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH": "1"
  }
}

Setting CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH to 1 will restore the behavior where Subagents cannot dispatch sub-Agents, which is consistent with the tightened strategy in version 2.1.213.

For scenarios that need to observe Subagent output in CI or headless environments, version 2.1.219 also supports the --forward-subagent-text flag and the CLAUDE_CODE_FORWARD_SUBAGENT_TEXT environment variable, which will make the text of depth-2 and above Subagents appear in the stream-json output.

/code-review Changed to Run as a Background Subagent¶

A workflow change in version 2.1.218 is highly noticeable in actual use: /code-review has been changed to run as a background Subagent.

Before the change, the output of Code Review would be directly written to the main chat, and long diff analysis would easily take up context and interrupt ongoing coding tasks. After the change, the review runs in a background Subagent, keeping the main chat clean; if you use stacked slash commands, the review will still target the correct objects.

The official also fixed the issue where /code-review ultra would silently downgrade to local review in non-interactive sessions – now it will correctly start cloud review. For PR review scenarios, you can specify a Pull Request via /code-review <pr#>, and use --comment to post inline comments to GitHub.

This change complements the Subagent depth expansion: the review itself runs in the background, and the review Agent can dispatch verification Subagents internally, forming a combination of “no pollution to the main chat + in-depth review”.

Developer Practice Recommendations¶

Combined with official documentation and community feedback, after upgrading to version 2.1.220 series, you can verify the environment in the following steps:

1. Confirm the Version

claude --version

The output should show 2.1.220 or higher.

2. Check the Default Model
Run /model in a Claude Code session, and the Opus line should display Opus (1M context), which corresponds to claude-opus-5.

3. Evaluate the Subagent Strategy
- Large monorepo refactoring: Keep the default depth of 3 to leverage multi-level division of labor
- Cost-sensitive or simple script modification: It is recommended to set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to reduce token consumption
- High-concurrency scenarios: Pay attention to CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS to avoid triggering too many background Agents in a single message

4. Stay Objective When Comparing Evaluations
Benchmarks like Terminal-Bench and SWE-bench have different test harnesses. The NeuralCoreTech article clearly pointed out that the 89%+ scores of GPT-5.6-Sol and Claude Opus 5 are third-party “pure model” tests, and Claude Code + Fable 5 scored 83.1% on Terminal-Bench in the public Agent leaderboard. When selecting tools, you should distinguish between “model intelligence” and “Agent product completion”, which are not equivalent.

Summary¶

The Claude Code 2.1.220 series update marks that Anthropic is advancing three lines simultaneously in the AI programming Agent track: stronger default model (Opus 5), more flexible Subagent orchestration (3-layer nested depth + concurrency/session limits), and cleaner main chat experience (background Code Review). The narrow gap with OpenAI Codex + GPT-5.6-Sol on Terminal-Bench shows that the competitive focus of AI programming Agents in the second half of 2026 is shifting from “whose model has a higher score” to “whose Agent harness is more stable and whose workflow is smoother”.

For ordinary developers, you can enjoy Opus 5 by default after upgrading without extra operations; if there are too many Subagent dispatches or review outputs interfere with the main process, you can finely adjust through the aforementioned environment variables and settings. It is recommended to follow the official CHANGELOG to track subsequent versions – Claude Code released dozens of minor versions alone in July, with an extremely fast iteration pace.