AI Coding Roundup — September 10, 2026: Claude Code v2.1.267 Caps Effort as Codex CLI and OpenCode Both Ship GPT-6-Astra
September 9 brought a fast-moving convergence in AI coding tools: Anthropic shipped a stability-focused Claude Code release the day after its previous hotfix, while OpenAI's Codex CLI and the independently maintained OpenCode project both quietly rolled out support for the same new model family within about 24 hours of each other. Here's what changed heading into September 10, 2026, and what's still sitting quiet on GitHub Copilot's side.
Claude Code v2.1.267: Capping Effort, Stabilizing the Cache
One day after the v2.1.266 hotfix for the LLM-gateway regression, Anthropic shipped Claude Code v2.1.267 on September 9, 2026 — documented in the official changelog, a release with roughly 53 changes, about two-thirds of them bug fixes.
- Added a maxEffortLevel setting (top-level or per model under modelSettings) that caps the effort level on every provider, including Bedrock, Vertex, and Foundry; users can still choose a lower level themselves.
- Added --system-prompt-snapshot off, which renders the system prompt fresh on every request instead of reusing the conversation's recorded prompt — useful when actively iterating on prompt text.
- Fixed a deeper prompt-cache stability issue: subagents and sessions started with --system-prompt or --append-system-prompt now record the system prompt and tool definitions once instead of re-rendering them on every turn.
- Fixed Workflow agent() calls with large output schemas being refused outright in auto mode instead of being checked by the safety classifier.
- Fixed Cowork scheduled tasks in the cloud failing at startup for organizations whose managed settings require sandboxing.
- Fixed managed allowedHttpHookUrls, httpHookAllowedEnvVars, and allowedChannelPlugins admitting everything instead of nothing when the value was unreadable — a fail-open bug with real security implications for locked-down organizations.
That last fix matters more than its one-line summary suggests: an allowlist that fails open when it can't be read defeats the point of having one. Combined with v2.1.265's plugin-loading and Windows-sandbox work from two days earlier, Anthropic's last three releases read as a coordinated push to make Claude Code's caching and policy layers predictable in enterprise deployments.
GPT-6-Astra Lands in Both Codex CLI and OpenCode, Hours Apart
The more interesting story this week isn't a single vendor — it's how fast a new OpenAI model spread across the ecosystem. OpenAI's own Codex CLI reached v0.154.0 on September 10, 2026, adding GPT-6-Astra to its model picker and Amazon Bedrock catalog, alongside experimental worktree support for isolated checkouts on new or forked sessions, inline question-answering while work continues, and Windows background-server daemon management. The release also brings a Vim replace mode with undo and dot-repeat, and preserves rich-text formatting when a response is copied into apps like Slack or Notion.
The community-maintained OpenCode got there almost as fast. Version 1.18.30, shipped September 9, 2026, adds Astra system-prompt support for GPT-6 models, fixes Bedrock DeepSeek model IDs that weren't resolving correctly, and adds reasoning-effort variants for GitLab GPT and Claude models — building on the OAuth model-filtering fix from v1.18.29 (September 4) that had been blocking gpt-6-astra from appearing for OpenAI subscription users in the first place.
Read together, GPT-6-Astra rolling out through Bedrock catalogs, OAuth filters, and system-prompt templates within the same 48 hours across two independently maintained tools says something about how OpenAI is distributing early access this generation — through provider integrations and third-party clients as much as through ChatGPT itself.
GitHub Copilot: Quiet Since September 4
GitHub Copilot CLI hasn't shipped a new version since v1.0.83 on September 4, 2026, which added a Windows 11 taskbar view for running sessions, Client ID Metadata Document (CIMD) support for MCP OAuth sign-in, and automatic HTTPS proxy mTLS client-certificate support. That's a longer quiet stretch than usual after a week that saw Copilot for JetBrains pick up enterprise-managed sandbox policies on September 8, plus a wave of scheduled model deprecations across the Copilot lineup — Gemini 3.5/3.6 Flash to Gemini 3.8 Flash, Kimi K2.7 Code to Kimi K3, and Claude Opus 4.7 to Claude Opus 5 — landing October 2, 2026.
Bottom Line
If you're on Claude Code, update to v2.1.267 for the maxEffortLevel guardrail and the prompt-cache fixes — teams on Bedrock, Vertex, or Foundry get the most benefit. If you use Codex CLI or OpenCode and want early access to GPT-6-Astra, both are ready today; OpenCode users should also grab the Bedrock DeepSeek ID fix. Copilot users have nothing new to install this week, but should plan around the October 2 model deprecations now.
Resources & References
Comments
Share your thoughts and join the conversation