AI Coding Tools Roundup: September 22, 2026

September 22, 2026 — This week's AI coding tools cycle brought a genuinely useful mix of changes: Claude Code made its server-side auto-mode classifier free for API and Enterprise users and added AGENTS.md fallback support, GitHub Copilot pushed a multi-model orchestration research preview called Project HydraFusion into its CLI, OpenAI retired the fastest Codex model it has ever shipped after just seven months, and OpenCode kept its usual rapid pace of provider and stability fixes. Here's what actually matters.

Claude Code: Auto-Mode Goes Free, AGENTS.md Support Lands

Version 2.1.278 (September 19) changes how auto mode is billed. API, Enterprise, Bedrock, Vertex, Foundry, and gateway users now hit the server-side classifier by default, which no longer charges for classifier overhead. A new "Auto mode server" row was added to /status so you can confirm which mode a session is running in; setting CLAUDE_CODE_AUTO_MODE_SERVER=0 opts back out on non-Anthropic backends. The same release removed the TaskOutput tool — background task output now flows through the standard Read tool instead.

A day earlier, version 2.1.277 (September 18) added AGENTS.md support: Claude Code now reads it as a fallback whenever a project has no CLAUDE.md, closing a compatibility gap with repos that standardized on the cross-agent AGENTS.md convention. It's not yet available on Bedrock, Vertex, or Foundry. Full details are in the official changelog.

Other fixes shipped this cycle:

• Sessions that used to hang with no result after an internal error now surface the error and exit cleanly.
• A bug causing conversations to fail with "text content blocks must be non-empty" errors was fixed.
• Unexpected logouts caused by older Claude Code builds running on the same machine no longer happen.

GitHub Copilot: Project HydraFusion Multi-Model Orchestration

The headline Copilot news is Project HydraFusion, a research preview now live in Copilot CLI via /experimental on every plan. Instead of routing a task to one fixed model, HydraFusion builds a per-task execution plan and picks between three patterns — Single (route to one model), Cascade (try an efficient model first, then escalate through a quality gate if needed), and full multi-model orchestration across providers to draft, critique, and revise.

On TerminalBench 2.1, GitHub reports a 4.9 percentage-point improvement in verified task quality alongside a 67% reduction in estimated cost compared with running Claude Opus 5 alone — a meaningful efficiency claim if it holds up outside GitHub's own benchmark.

Beyond HydraFusion, Copilot shipped a Jira integration in the Copilot app, tightened enterprise controls so the Copilot app and CLI now respect content exclusion policies set by org and repo admins, and updated the impact dashboard to show feature-level engagement so enterprise admins can see which Copilot experiences are actually being used. Copilot code review also got smarter: it now resolves previously addressed comments on subsequent passes and suggests commit messages when a suggestion is applied.

On the business side, GitHub reopened Copilot Business and Enterprise sign-ups for credit-card and PayPal customers on September 1, and plans to relaunch Copilot Chat on github.com, GitHub Mobile, and the cloud agent as a single unified experience no earlier than September 28. Full detail in the weekly release notes.

OpenAI Codex: GPT-5.3-Codex-Spark Retired After Seven Months

OpenAI deprecated GPT-5.3-Codex-Spark on September 14, pulling it from the ChatGPT desktop app, Codex CLI, and IDE extension. Spark was OpenAI's fastest model ever — over 1,000 tokens per second — and its first production model served on non-Nvidia hardware, running on Cerebras WSE-3 wafer-scale chips. Its reason for existing largely evaporated in August when Cerebras shipped an Ultrafast mode that runs the flagship GPT-5.6 Sol at 750 tokens per second, closing most of the speed gap.

Elsewhere in the Codex changelog, this cycle added:

• Experimental voice conversations with live transcripts and microphone controls, enabled through /experimental.
• Live reasoning summaries in the TUI status row, plus completion timestamps after successful turns.
• Task hiding, archiving, and deletion in the agents overview, with worktree ownership details and confirmed deletion of clean managed worktrees.
• Touch ID verification for MCP requests in local TUI sessions on supported Macs.
• ExternalMessage support for sync and async run/turn calls, and new resume and history options for turn selection, service tier, and source metadata.

OpenCode: Cloudflare AI Gateway Passthroughs and a Stability Sweep

OpenCode added native OpenAI and Anthropic passthroughs for Cloudflare AI Gateway models this cycle, tightened Codex rate-limit handling, and pushed a run of provider, pricing, residency, and web-search reliability fixes. Recent releases from the GitHub release history:

• September 14 — Fixed ACP session state persistence when loading, resuming, or forking sessions.
• September 9 — Added an Astra system prompt tuned for GPT-6 models.
• September 4 — Fixed OAuth model filtering for integer GPT version numbers.
• September 2 — Raised the default provider header timeout to five minutes to cut down on startup failures for slow models.
• September 1 — Claude 5 sessions now tolerate stale thinking blocks instead of failing after prompt or tool changes.

TL;DR — What Actually Matters

• Claude Code v2.1.278/2.1.277 — Server-side auto-mode classifier is now free for API/Enterprise; AGENTS.md is read as a CLAUDE.md fallback; TaskOutput tool removed.
• GitHub Copilot — Project HydraFusion multi-model orchestration hits CLI as a research preview (+4.9pp quality, -67% cost on TerminalBench 2.1); Jira integration; content-exclusion enforcement in app/CLI.
• OpenAI Codex — GPT-5.3-Codex-Spark retired September 14 after Cerebras Ultrafast mode closed the speed gap; voice mode, live reasoning summaries, and Touch ID for MCP shipped.
• OpenCode — Native Cloudflare AI Gateway passthroughs, tighter Codex rate limits, and a steady stream of provider/session stability fixes through September.

Resources & References

• Claude Code Official Changelog
• Claude Code Stops Charging for Auto-Mode Safety Checks — Mixed News
• Claude Code 2.1.277 AGENTS.md Support — ccleaks
• Project HydraFusion: Frontier Quality via Multi-Model Orchestration — The GitHub Blog
• GitHub Copilot Weekly Releases — September 14
• OpenAI Codex Changelog
• OpenAI to Retire GPT-5.3-Codex-Spark — BigGo Finance
• OpenCode GitHub Releases
• OpenCode Changelog — gradually.ai