AI Coding Roundup — September 12, 2026: Claude Code, Copilot & Codex Ship Major Updates

AI coding tools kept shipping through the first half of September 2026. Anthropic pushed Claude Code to v2.1.269 with plugin evals and output-style switching, GitHub Copilot's code review agent now resolves its own comments and runs ensemble reviews, OpenAI made GPT-6-Astra the default Codex model, and OpenCode caught up with Astra support of its own. Here's what changed and why it matters for anyone building with these agents.

Claude Code v2.1.269 — Plugin Evals, Output Styles & a Safer Gateway

The September 11 release (v2.1.269) is one of the most feature-dense in weeks. The new claude plugin eval command runs plugin eval suites and returns scored, reproducible results, giving plugin authors a way to regression-test their skills before publishing. A companion /output-style [name] command lets you list and switch output styles on the fly. Tool results for Bash commands now include a diff of any files the command changed, and OpenTelemetry metrics gained repository-attribute tagging for better fleet-wide observability.

The days before that were eventful for a different reason. v2.1.268 (September 10) added pricing controls to the Claude apps gateway via gateway.yaml and finally fixed WebFetch hanging indefinitely by adding a 300-second timeout. But v2.1.265 had briefly shipped a regression where CLAUDE_CODE_USE_GATEWAY forced Cloud-gateway sign-in and broke plain API-key configurations — a fix landed within a day in v2.1.266, a reminder to pin versions in CI if you rely on gateway or auth settings.

GitHub Copilot Code Review — Auto-Resolution and Ensemble Analysis

GitHub's September 11 update to Copilot code review closes a loop that's long been manual: once you address a comment, Copilot now resolves the thread itself, and when you apply one of its suggestions it writes a context-aware commit message instead of a generic one. Reviews can also reach for a broader set of shell tools — running build commands, executing tests, and calling targeted scripts — to validate a finding before flagging it, cutting down on false positives.

The headline number: Copilot's Lite effort level now runs an ensemble of agents rather than a single pass, and GitHub reports 47% more high-severity findings addressed and 31% more medium-severity findings resolved, while lowering review costs by 8%. Separately, GitHub confirmed model deprecations effective October 2: Gemini 3.5/3.6 Flash both roll into Gemini 3.8 Flash, Kimi K2.7 Code moves to Kimi K3, and Claude Opus 4.7 is replaced by Claude Opus 5 — worth checking now if any of your workflows pin a model by name.

OpenAI Codex — GPT-6-Astra Becomes the Default

OpenAI launched GPT-6-Astra on September 3 and it's now the recommended and bundled default model in the Codex CLI as of v0.153.4. Around it, OpenAI rebuilt several workflow pieces: /plugins now sorts remote plugins into OpenAI Curated, Workspace, and Shared-with-me sections and can proactively recommend installs mid-turn; configurable rollout token budgets track spend across agent threads with remaining-budget reminders and abort turns when exhausted; and app-server clients can now set multi-agent delegation to disabled, explicit-request-only, or proactive at the thread or turn level.

Security and usability also got attention: a new indexed web-search mode allows live searches while restricting direct page access to server-approved URLs, Codex can now receive scheduled UTC time reminders and query the current time via client-provided clocks, markdown links render as clickable labels in supported terminals, and new Interrupt hooks can run a command or MCP handler whenever an active top-level turn is cancelled.

OpenCode v1.18.30 — Astra Support Catches Up

The open-source, provider-agnostic agent didn't wait long to follow. OpenCode v1.18.30 (September 9) introduces an Astra system prompt tuned for GPT-6 models, preserves Bedrock DeepSeek model IDs including ARN-based formats, updates the Azure and OpenAI provider SDKs for compatibility, and adds reasoning-effort variants for GitLab-hosted GPT and Claude models. It follows v1.18.29 (September 4), which fixed Codex OAuth model filtering so integer GPT versions like gpt-6 are recognised correctly and resolved gpt-6-astra not appearing for OpenAI subscription users — a fast turnaround that kept OpenCode users on par with the official Codex CLI within days of GPT-6-Astra's launch.

Resources & Further Reading

Claude Code Official Changelog — full release notes for v2.1.261 through v2.1.269.

GitHub Changelog — Copilot Code Review Auto-Resolution — the September 11 announcement with ensemble-review numbers.

GitHub Changelog — Copilot Model Deprecations — which models are retiring on October 2 and their replacements.

OpenAI — Introducing Upgrades to Codex — GPT-6-Astra and the new Codex CLI workflow features.

OpenCode Changelog — v1.18.27 through v1.18.30 and the full version history.

Havoptic — Claude Code Release Tracker — automated, versioned summaries of every Claude Code release.