This Week in Coding Agents: September 7, 2026

Week 37, 2026 · Published

Claude Code2.1.263
Codex CLI0.153.4
Gemini CLI0.58.0

The short version

  • Claude Code’s changelog lists 2.1.263 as the top release (September 6). 2.1.257 made Claude Fable 5.1 (claude-fable-5-1) the default Fable model. Later builds add /diff, managedMcpServers, and bashOutputMaxChars / taskOutputMaxChars.
  • Codex CLI 0.153.4 (September 4) puts Astra in the bundled model picker and makes it the bundled default when no model is set. 0.153.0 added a plugin marketplace CLI and vim undo/redo.
  • Cursor has no version number. The September 2 changelog entry is Self-hosted machines: tool execution stays on machines you run.
  • GitHub Copilot has no version number. This week it added GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash, and it will drop Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 on October 2, 2026.
  • Gemini CLI 0.58.0 shipped on September 1. The release notes I read are sandbox and ignore-path fixes, not a new model.

Claude Code

2.1.257 (September 1) through 2.1.263 (September 6). I opened the live CHANGELOG.md. The first heading is 2.1.263. There is no 2.1.262 heading on that page.

2.1.263 (September 6). The only note on the page: bug fixes and reliability improvements. No named features.

2.1.261 (September 4). /status and claude doctor now say why an organization policy failed to load (example given: a proxy not passing the endpoint through). New settings bashOutputMaxChars and taskOutputMaxChars raise how much command and background-task output Claude sees inline before the rest is saved to a file, up to 128K characters. --append-subagent-system-prompt-file reads a large subagent system prompt from a file. /skill-doctor lists loaded skills that go unused and what they cost in context. A long list of Remote Control, Bedrock, and resume fixes sits under that.

2.1.260 (September 3). A diff panel opens beside the conversation in fullscreen and shows uncommitted changes as Claude edits. Toggle it with /diff. /cost and the status-line prompt_cache field now name a likely cause for a cache miss (tool definitions or system prompt changed, idle past the TTL). /reload-plugins works in headless sessions. /advisor, /advisor <model>, and /advisor off work in the desktop app, Remote Control, and -p / Agent SDK sessions. The /model picker now shows Fable 5.1 for orgs that can use it; before this, only /model claude-fable-5-1 worked. Prompt caching on Fable 5.1 now covers context attached after tool results. /rewind no longer reports success when the backup files are missing.

2.1.259 (September 2). Organizations can ship HTTP/SSE MCP servers to every user with the managedMcpServers managed setting (same entry shape as .mcp.json; entries that name a command to run are skipped). --permission-prompts none denies anything that would prompt on an unattended host, while the active permission mode still decides. claude plugin validate --json prints a machine-readable report. Concurrent sessions no longer silently revert each other’s ~/.claude.json changes. If a managed-settings file cannot be parsed, Claude Code now refuses to start and names the source.

2.1.258 (September 1). Launch on macOS 12 (Monterey) works again after a 2.1.255 regression. Remote and scheduled sessions no longer fail with “user messages must have non-empty content” after a re-sent permission approval could not be applied.

2.1.257 (September 1). Claude Fable 5.1 (claude-fable-5-1) is the default Fable model. The changelog lists 1M context, $10/$50 per Mtok, and $0.25/Mtok cache reads. New settings: timeFormat and timeZone for the turn-end clock. Auto mode gets a Containment Escape rule so cloud metadata-credential fetches, egress evasion, and cross-tenant reach are not auto-approved unless the environment marks them expected. CLAUDE_CODE_SUBAGENT_MODEL_FORCE applies CLAUDE_CODE_SUBAGENT_MODEL (or the main model) to every subagent. s in /effort changes effort for the current session only. Auto mode now prompts before the first file read outside the working directories, with permissions.blockReadsOutsideWorkingDirectories as the block option.

Codex CLI

0.152.0 (September 1) through 0.153.4 (September 4). I opened each rust-v* release page. The releases list shows 0.153.4 as the current stable tag.

0.153.4 (September 4). Astra is visible in the bundled model picker and is the bundled default when no model is explicitly configured. Astra’s guidance now uses asynchronous questions only when that tool is available in the session.

0.153.3 (September 4). GPT-6-Astra is in the Amazon Bedrock model picker for Mantle and Runtime global/US routes. Async clarification questions for Astra use the supported tool and accept text only.

0.153.2 (September 3). The GPT-6-Astra Fast tier label now says “2x speed, increased usage” instead of “1.5x”. The notes say this is display text only.

0.153.1 (September 3). You can configure GPT-6-Astra through the API without changing the default model or showing it in the model picker.

0.153.0 (September 3). Vim mode gets undo (u) and redo (Ctrl+R), including pasted content and attachments. The plugin CLI can list, install, and remove plugins from remote marketplaces. tui.auto_recap = false turns off automatic recaps and leaves /recap in place. TUI history shows complete patches, input sent to background terminals, and individual completed commands. Plus and Team users get an earlier warning when less than half of a roughly five-hour usage window remains. tui.disable_paste_burst replaces the old top-level setting (the old key still works as a fallback). features.context_management.experimental_mode is off by default; when on, eligible ChatGPT Plus, Pro, or Pro Lite sessions on the Codex backend get token-budget context, history notes, and the new_context tool. API-key sessions, custom providers, and temporary structured threads stay out.

0.152.1 (September 1). Guardian approval review honors Node REPL policies from model metadata.

0.152.0 (September 1). Vim mode searches drafts with / and ?, highlights matches, and repeats with n / N. Rate-limit banners link to usage, credits, limit resets, and plan management. The TUI and codex exec show credential-refresh progress, including Amazon Bedrock reauthentication. MCP server names may include :, @, /, and .. Individual MCP tools accept output_token_limit. App-server clients can set thread/shellCommand timeouts past one hour. The planning tool is off by default; the notes say to turn it on with tools.update_plan.enabled = true.

GitHub Copilot

No version number. I opened each changelog URL below.

GPT-6 Astra is generally available (September 4). Copilot Pro+, Max, Business, and Enterprise. Model picker in VS Code, Visual Studio, Copilot CLI, coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Rollout is gradual. Billed at provider list pricing under usage-based billing. Business and Enterprise admins manage access through the model policy; new models are enabled automatically unless the global default is off or this model is disabled.

Weekly releases, August 31 (September 4). Recap: Fable 5.1 for Pro+, Max, Business, Enterprise; Gemini 3.8 Flash rolling out to Pro through Enterprise; Copilot app and CLI honor content exclusions; Copilot harness generally available in JetBrains. VS Code 1.136: Agent Merge in public preview, experimental multi-root workspaces and chat backgrounds, plus chat-session hierarchy.

Upcoming deprecation of selected models (September 3). On October 2, 2026, these drop across Chat, inline edits, ask/agent modes, and completions: Gemini 3.5 Flash (move to Gemini 3.8 Flash), Gemini 3.6 Flash (same), Kimi K2.7 Code (move to Kimi K3), Claude Opus 4.7 (move to Claude Opus 5). Business and Enterprise admins may need to enable the replacements in model policy.

Gemini 3.8 Flash (September 3). Copilot Pro, Pro+, Max, Business, and Enterprise. Introductory provider pricing under usage-based billing through December 31, 2026. Picker coverage matches the other new models, plus “GitHub Copilot cloud agent”. Same default-enablement rule as Astra.

Reopening Business and Enterprise signups (September 3). Card and PayPal signups reopen over a couple of weeks. New seat assignments require payment before access. Starting October 1, 2026, existing card/PayPal Business and Enterprise seats take an upfront charge at the start of the billing cycle. Exceeding included usage may require more payment before you can keep using Copilot.

Enterprise-managed settings, any default model (September 2). Set a preferred default model for new conversations, including per enterprise team via team-mappings.json. Generally available for Business and Enterprise in the Copilot app, Copilot CLI, and VS Code.

Content exclusions in the Copilot app and CLI (September 2). App and CLI honor enterprise, org, and repo exclusion policies. Business and Enterprise only.

Expiration dates on individual user budgets (September 1). Optional expires_at on a user budget; GitHub removes it on that date and the user falls back to the next budget. Business and Enterprise. Also in the Budgets REST API.

Copilot code review can approve pull requests (September 1). Public preview for Pro through Enterprise. Every Copilot review now includes an approval assessment in the overview comment. Actual approvals that count toward required-approvals are off by default and configurable at enterprise, org, and repo (including allowed file paths). New commits dismiss the approval, same as a human review.

Claude Fable 5.1 (September 1). Pro+, Max, Business, and Enterprise. The Fable 5.1 policy is off by default. The page says Fable 5.1 requires data retention by default so Anthropic can run safety classifiers; retained data is not used to train models. Other Claude models except Fable 5 and Fable 5.1 stay on zero data retention. Eligible enterprises can get a time-bound zero-retention exemption through the end of the calendar year. Approval from GitHub does not flip the policy on; an admin still has to enable it.

Cursor

No version number. I opened Self-hosted machines (dated September 2, 2026 on that page).

Cursor now supports self-hosted machines so tool execution stays on your network. The page says your codebase, build outputs, and secrets stay on internal machines, while the agent handles tool calls locally.

My Machines connects one laptop or VM for personal work. Team pools are named queues of workers. Capacity grows as requests arrive and shrinks when workers disconnect. Pools are not tied to one repository. Idle machines can hibernate and restore inside a reconnect window.

Cloud agents can run on infrastructure you already use. The page lists AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B.

Self-hosted workers support computer use on Linux and Mac: click, type, screenshots, and browser, if the right desktop packages are installed. You can watch the desktop or take control from Cursor.

Gemini CLI

0.58.0 (September 1). I opened the v0.58.0 release. The releases list shows that tag as current.

Notes I saw: ignore-path handling now evaluates symlinks consistently; Docker and container runtime sockets and binaries are isolated in macOS Seatbelt; write-policy config declares top-level safety checkers; history rollback and retry-nudge work; the a2a-server clears a stale cancellation error on a new message turn. No new model name appears on that page.

What this means for your setup

  • If you rely on Codex picking its bundled default, 0.153.4 will select Astra unless you set a model. Pin the model you want in config before you upgrade.
  • If you used Codex’s planning tool, 0.152.0 turned it off. Set tools.update_plan.enabled = true if you still want it.
  • Claude Code users who want Fable 5.1 can type /model claude-fable-5-1. After 2.1.260 the picker also lists it. Orgs that push MCP over HTTP/SSE can use managedMcpServers instead of a local command entry. Headless hosts that must never block on a prompt can pass --permission-prompts none.
  • Copilot Business and Enterprise admins: Fable 5.1 stays off until you enable the policy, and enabling it accepts Anthropic’s retention default. Move off Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 before October 2. Leave Copilot PR approvals off unless you want an agent sign-off to count toward required reviews.
  • Cursor teams that cannot send tool execution off-network should read the self-hosted machines page and decide between My Machines and a team pool.
  • src/content/learn/ and src/content/playbooks/ are empty right now (Track B has not landed), so I have no page that states a command, flag, path, or default I would flag for a refresh.

Pages we updated this week

None this week.

Sources