What changed in Claude Code during September 2026?

Claude Code shipped 27 releases in September 2026, from v2.1.257 on 2026-09-01 to v2.1.286 on 2026-09-30. The changes most relevant to agent operators were new Opus and Sonnet defaults, server-side auto mode with no classifier-overhead charge, AGENTS.md fallback, and tighter controls for subagents and unattended runs.

Update on 2026-10-06. The public changelog continues through v2.1.289, published on 2026-10-03. This post stays scoped to the September releases.

Claude Code's September 2026 release count

Twenty-seven Claude Code releases have published dates in September 2026. The release sequence starts with v2.1.257 on 2026-09-01 and ends with v2.1.286 on 2026-09-30. These are release dates, not a claim that every intervening version number exists.

Claude Code's model defaults

Claude Opus 5.5 arrived in v2.1.280 on 2026-09-22. The release note calls claude-opus-5-5 the default Opus model, with a 1M-token context, $4 per million input tokens, $20 per million output tokens, and $0.20 per million cache-read tokens. The same release changed the default model on Pro and Team Standard plans from Sonnet to Opus.

Claude Sonnet 5.5 arrived in v2.1.284 on 2026-09-28. claude-sonnet-5-5 became the default Sonnet model on the Anthropic API, with a 1M-token context, $2 per million input tokens, $10 per million output tokens, and $0.20 per million cache-read tokens. The plan-level Opus change and the API-level Sonnet change are separate defaults.

Claude Code's server-side auto mode

Version 2.1.278 changed auto mode for Claude API and Enterprise users, plus Bedrock, Vertex, Foundry, and gateway sessions, to use a server-side classifier by default. The release note says that classifier overhead is not charged. The accompanying documentation distinguishes this from agent usage itself: if a session falls back to Claude Code's own classifier requests, those requests count toward token usage.

The same release added an Auto mode server row to /status. The row reports Enabled when the server performs the classifier review and Disabled after fallback. CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out on Bedrock, Vertex, Foundry, and gateways.

Version 2.1.284 changed interactive terminal and VS Code sessions to start in auto mode when no permission mode is configured, across plans and providers. permissions.defaultMode still overrides that behavior. Version 2.1.285 made the same default apply to claude -p and Python Agent SDK sessions on third-party providers or with telemetry disabled when no permission mode is configured. An explicit --permission-mode still overrides it.

Claude Code's subagent controls

Version 2.1.257 added CLAUDE_CODE_SUBAGENT_MODEL_FORCE. It applies CLAUDE_CODE_SUBAGENT_MODEL, or the main model when no subagent model is set, to every subagent and ignores per-spawn and agent-definition model overrides:

export CLAUDE_CODE_SUBAGENT_MODEL="claude-sonnet-5-5"
export CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1

Version 2.1.271 added omitClaudeMd: true to agent frontmatter and --agents JSON. Custom and plugin subagents can then skip user, project, and local CLAUDE.md files while managed policy files still load. Version 2.1.277 changed subagent results to arrive under a subagent-output header with indented text, so returned text is not treated as the primary session's instructions. Version 2.1.260 removed the one-hour limit on background commands started by subagents.

Claude Code's shared instructions and unattended tools

Version 2.1.277 added AGENTS.md support. When a project has no CLAUDE.md, Claude Code reads AGENTS.md instead. Version 2.1.281 expanded that fallback to Amazon Bedrock, Google Vertex AI, Microsoft Foundry, LLM gateways, and sessions with telemetry disabled.

Version 2.1.259 added --permission-prompts none for unattended headless hosts. Any action that would ask for a prompt is denied automatically, while the active permission mode continues making its own decisions. Version 2.1.285 added CLAUDE_CODE_DISABLE_WEB_FETCH=1, which removes the WebFetch tool for a session.

Version 2.1.269 also added claude plugin eval. It runs a plugin evaluation suite and produces scored, reproducible JSON and HTML results. The weekly digest describes a no-plugin comparison, which gives plugin authors a concrete baseline instead of relying on a demo.

Sources

• Anthropic Claude Code changelog: https://code.claude.com/docs/en/changelog (read 2026-10-06)
• Anthropic auto mode classifier request charges: https://code.claude.com/docs/en/auto-mode-classifier-billing (read 2026-10-06)
• Anthropic What's new in Claude Code: https://code.claude.com/docs/en/whats-new (read 2026-10-06)
• Anthropic Claude Code GitHub releases API: https://api.github.com/repos/anthropics/claude-code/releases (read 2026-10-06)

Last verified: 2026-10-06.