AI Coding Roundup — September 5, 2026: GPT-6 Astra Rolls Out Across Copilot, Codex, and OpenCode

Five separate AI coding tools shipped updates in the past 48 hours, and the pattern is unmistakable: GPT-6 Astra is now available almost everywhere you write code, while Anthropic's Claude Code spent its latest release polishing the operational tooling teams need to run agents at scale rather than adding another flashy feature. Here's what changed and why it matters.

GPT-6 Astra becomes the default almost everywhere

OpenAI's GPT-6 Astra spent the last few days rolling out across nearly every major coding assistant. GitHub Copilot marked the model as generally available for all users on September 4, following Gemini 3.8 Flash's arrival in Copilot just a day earlier — a sign of how quickly Copilot's multi-model lineup keeps widening. Full details are in the GitHub Copilot changelog.

OpenAI's own Codex CLI caught up too. Version 0.153.3 added GPT-6-Astra support to the Amazon Bedrock model picker, and version 0.153.4 fixed a bug that had kept Astra hidden from the bundled model picker for some users — see the Codex CLI changelog for the full release notes.

OpenCode, the provider-agnostic terminal agent from SST, shipped two patches in the same window: v1.18.28 added a GitHub Copilot session-ID header for better request tracing, and v1.18.29 fixed OAuth model filtering so integer GPT versions like gpt-6 are recognized correctly, resolving why gpt-6-astra wasn't showing up for OpenAI subscription users. Details in the OpenCode changelog.

Claude Code 2.1.261: quieter, but built for teams running at scale

Anthropic skipped flashy features this round and focused on operations. Release 2.1.261 adds an "Organization policy" line to /status and claude doctor so admins can confirm which managed settings are active, introduces bashOutputMaxChars and taskOutputMaxChars settings that raise how much command and background-task output Claude can see inline before it's spilled to a file (up to 128K characters), and ships --append-subagent-system-prompt-file for teams handing subagents long custom instructions without hitting prompt-length limits. A new /skill-doctor command also surfaces unused skills and how much context budget they're quietly consuming. The full list is in the Claude Code changelog.

Why it matters

None of this week's changes are headline-grabbing on their own, but together they point to two converging trends. Model choice is becoming table stakes — every major agent now ships the same frontier models within a day or two of each other — so the real competitive battleground is shifting to operational plumbing: output limits, policy visibility, session tracing, and context-budget tooling that only matter once a team is running these agents at scale rather than as a solo CLI experiment.

Resources & references

• Claude Code changelog
• GitHub Copilot changelog
• OpenAI Codex CLI changelog
• OpenCode changelog