AI Coding Roundup — September 13, 2026: Claude Fable 5.1, Copilot's HydraFusion, and GPT-6 Astra

Four of the most closely watched AI coding tools all shipped meaningful updates in the same two-week stretch: Claude Code pushed out roughly a release a day between September 1 and 12, making Claude Fable 5.1 its default Fable model and adding a live diff panel; GitHub Copilot rolled Jira into its app, put an experimental model router called Project HydraFusion into its CLI, and added agent task scheduling to VS Code; OpenAI's Codex CLI reached version 0.154.0 with GPT-6-Astra now visible in its model picker and on Amazon Bedrock; and OpenCode spent the same window chasing all three vendors' GPT-6 rollout to keep its own model picker in sync. Here's what actually shipped, verified against each project's own changelog, as of September 13, 2026.

Executive Summary

• Claude Code shipped 12 releases in 12 days (2.1.257 → 2.1.270), made Claude Fable 5.1 the default Fable model, added a live /diff panel and claude plugin eval, and fixed two self-inflicted regressions within a day of each landing.
• GitHub Copilot's Sep 7 weekly release brought Jira integration to the Copilot app, an experimental adaptive model router ("Project HydraFusion") to Copilot CLI, and agent task scheduling plus voice mode to VS Code 1.137.
• Copilot code review now auto-resolves its own comments once you address them, and writes commit messages when you apply its suggestions.
• OpenAI Codex CLI hit 0.154.0 (Sep 9) with GPT-6-Astra live in the model picker and on Amazon Bedrock, plus experimental worktree support and inline question-answering mid-task.
• OpenCode shipped v1.18.26 through v1.18.30 over two weeks, mostly patching GPT-6-Astra visibility and provider compatibility (Bedrock, Azure, GitHub Copilot auth) to keep pace with the upstream vendors.
• Watch your model list: Copilot is deprecating Gemini 3.5/3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 on October 2, 2026.

Claude Code: A Release a Day, and Fable 5.1 Becomes the Default

Claude Code shipped twelve numbered releases between September 1 and 12 — roughly one a day, unusually fast even by its own pace. The headline landed on day one: 2.1.257 (Sep 1) made claude-fable-5-1 the default Fable model, with a 1M-token context window and pricing of $10/$50 per million tokens ($0.25/Mtok for cache reads), and added a Containment Escape rule to auto mode that blocks cloud metadata-credential fetches and other egress-evasion attempts.

• 2.1.260 (Sep 3) added a diff panel beside the conversation in fullscreen mode, toggled with /diff, showing uncommitted changes as Claude edits.
• 2.1.261 (Sep 4) added /skill-doctor to show which loaded skills go unused and their context cost.
• 2.1.265 (Sep 8) added support for pointing --plugin-dir at a folder of plugins and capped tool results saved to disk at 1 GB.
• 2.1.267 (Sep 9) added maxEffortLevel to cap reasoning effort across every configured provider.
• Two regressions, fixed fast: 2.1.266 (Sep 8) reverted a bug where the undocumented CLAUDE_CODE_USE_GATEWAY variable started forcing Cloud-gateway sign-in on its own, breaking API-key and custom-auth setups; 2.1.268 (Sep 10) fixed every turn failing with HTTP 400 on third-party Anthropic-compatible endpoints (a 2.1.265 regression) and gave WebFetch a 300-second timeout so it stops hanging on servers that never close the response.
• 2.1.269 (Sep 11) added claude plugin eval for scored, reproducible plugin test runs, plus /output-style [name] support over Remote Control and headless sessions; 2.1.270 (Sep 12) then fixed a same-day regression where read-only git commands started asking for permission unexpectedly.

The full list — including hundreds of smaller fixes across permissions, MCP servers, and the VS Code extension — is in Anthropic's own changelog.

GitHub Copilot: Jira, HydraFusion, and Scheduled Agents

GitHub Copilot's September 7 weekly release spread new capability across every surface it ships on. The Copilot app can now pull Jira issues into a shared canvas so you decide what moves forward while Copilot handles investigation and pull-request prep. Copilot CLI gained an experimental feature called Project HydraFusion, which routes each request across local, cloud, and compound models automatically to balance latency, cost, and quality.

• VS Code 1.137 added agent task scheduling (hourly, daily, weekly, or on demand) in public preview, an experimental voice mode, and the ability to review issues and pull requests from the Agents window regardless of which repository they belong to.
• JetBrains integrations gained centralized enterprise controls for sandbox policy — enablement, filesystem/network access, and proxy settings — managed from one place.

Copilot code review also picked up two workflow changes on Sep 11: it now auto-resolves its own comments once you've addressed them, and writes commit messages for you when you apply one of its suggested changes.

On the business side, GitHub reopened Copilot Business and Enterprise signups for credit-card and PayPal customers on September 1, but existing customers move to upfront per-seat billing on October 1. The same announcement confirms model deprecations landing October 2: Gemini 3.5 Flash and Gemini 3.6 Flash both retire in favor of Gemini 3.8 Flash, Kimi K2.7 Code is replaced by Kimi K3, and Claude Opus 4.7 is replaced by Claude Opus 5.

OpenAI Codex: GPT-6 Astra Goes Wide

Codex CLI reached 0.154.0 on September 9, and its headline is GPT-6-Astra becoming visible in both the standard model picker and Amazon Bedrock's model catalog, after a pair of hotfixes (0.153.2–0.153.4, Sep 3–4) corrected Astra's visibility in the bundled picker. The same release added experimental worktree support for isolated checkouts, the ability to keep answering questions inline while Codex keeps working on a task, a background Codex server with daemon lifecycle commands on Windows, and a vim-style replace mode with undo and dot-repeat.

• The Python SDK followed a day later at 0.154.0 (Sep 10) with max and ultra reasoning-effort values, an ExternalMessage type for turn integration, and include_turns for finer history selection on resume or fork.
• The ChatGPT apps moved in step: iOS 1.2026.244 (Sep 8) added @ mentions to reference other tasks from the composer, live question-answering during active work, and background worktree setup with progress shown in a Live Activity; the Windows desktop app (26.908, Sep 11) added a quick-chat mode and app-window sharing via "appshots."

Full details are in OpenAI's own Codex changelog.

OpenCode: Playing Catch-Up Across Every Provider

OpenCode, the open-source terminal agent, spent the same two weeks almost entirely on provider compatibility rather than new features — a sign of how much churn the GPT-6-Astra rollout caused upstream. v1.18.29 (Sep 4) fixed gpt-6-astra not showing up for ChatGPT-subscription users by improving Codex OAuth model filtering to recognize integer GPT versions, and v1.18.30 (Sep 9) added Astra's system prompt for GPT-6 models, fixed Bedrock DeepSeek model-ID resolution, and updated the Azure and OpenAI provider SDKs.

• Smaller fixes in the same window: v1.18.28 (Sep 4) sends a GitHub Copilot interaction header for better session tracking; v1.18.27 (Sep 2) extended timeout defaults to five minutes for slow model startups and fixed unhandled errors on cancelled streams; v1.18.26 (Sep 1) removed outdated pricing messaging and fixed device-login links that broke on servers using relative verification URLs.

Full entries are on OpenCode's changelog.

The Common Thread: GPT-6 Astra Is the Model Every Tool Is Racing to Support

Read across all four projects and one story repeats: GPT-6-Astra is becoming the model tools reach for by default, and getting there broke things before it worked. Codex needed two hotfixes just to make Astra visible in its own picker; OpenCode needed a dedicated release to fix Astra not appearing for ChatGPT-subscription users and another to add its system prompt; and Claude Code's own gateway config gained pricing controls in the same window its Cloud-gateway auth had a same-week regression. If you route requests through more than one of these tools, or through a shared gateway config, it's worth explicitly checking which model each one falls back to when none is configured — several of this month's bugs were exactly that silent-default problem.

What to Prioritize

• Update Claude Code past 2.1.270. Two back-to-back regressions (2.1.266's gateway-auth bug and 2.1.269's permission-prompt regression) mean anything between 2.1.265 and 2.1.269 is worth skipping in favor of the latest build.
• Audit your Copilot model routing before October 2. Gemini 3.5/3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 all retire that day — anything pinned to those model IDs needs a new default.
• Confirm GPT-6-Astra actually resolves where you expect it to. Codex, OpenCode, and Claude Code's Bedrock/gateway paths were all mid-migration this month; verify Bedrock ARNs, OAuth model filters, and gateway pricing config rather than assuming the picker shows the model you're actually billed for.
• Turn on Copilot's self-resolving code review. It closes its own comments once addressed and writes commit messages for applied suggestions — a small but real reduction in review overhead if you haven't enabled it yet.
• Verify changelog claims against the vendor's own page. Third-party aggregators are useful for a first pass, but versions and dates drift; every claim above was cross-checked against Anthropic's, GitHub's, OpenAI's, and OpenCode's own changelog pages linked throughout.