AI Coding Roundup — September 19, 2026: Claude Code Cuts Auto Mode Costs, Codex CLI Ships a Quick Fix

Anthropic pushed a billing-relevant Claude Code release today, OpenAI's Codex CLI quietly shipped a stability fix, and GitHub Copilot got a reworked review experience alongside a heads-up on model retirements. Here's what actually changed, sourced straight from each project's own changelog.

Claude Code 2.1.278: Auto Mode Moves to a Server-Side Classifier

Anthropic shipped Claude Code 2.1.278 on September 19, and the headline change is about your bill rather than a new feature. For Claude API and Enterprise users, and for anyone running Bedrock, Vertex, Foundry, or a custom gateway, auto mode's model-routing classifier now defaults to running on Anthropic's servers instead of locally — which means the classifier's own token overhead is no longer charged. See the official auto mode classifier billing docs for the mechanics.

• What changed: auto mode's routing classifier defaults to server-side execution, removing the token cost of the classification step itself.
• Opt-out: CLAUDE_CODE_AUTO_MODE_SERVER=0 reverts to local classification on Bedrock, Vertex, Foundry, and gateway configurations — Claude API and Enterprise accounts don't get a toggle.
• Visibility: /status now shows an "Auto mode server" row confirming whether the current session's classifier is running server-side, and Claude Code warns if it silently falls back to a billed local classifier.
• Why it matters: teams that lean on auto mode heavily get a quiet cost reduction, but it's worth checking /status after upgrading — the server path isn't guaranteed on every Bedrock/Vertex/Foundry setup.

Claude Code 2.1.277: AGENTS.md Support and a Gateway Egress Flag

One version back, 2.1.277 (September 18) was a compatibility-and-infrastructure release rather than a change to daily CLI use. Full details are in the official Claude Code changelog.

• AGENTS.md fallback: projects without a CLAUDE.md now have Claude Code read an AGENTS.md file instead, extending compatibility with the cross-tool convention also used by Codex CLI and other agents.
• New gateway env var: CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 tells Claude Code that its configured gateway is the only network egress point, for apps-gateway deployments routed entirely through a forward proxy.
• Why it matters: teams standardizing instructions across multiple AI coding tools can maintain one AGENTS.md instead of duplicating a CLAUDE.md, and locked-down enterprise gateways get an explicit way to declare their network boundary.

OpenAI Codex CLI 0.155.1: A Same-Day Fix for Reasoning Summaries

Codex CLI wasn't as quiet this week as it first looked. OpenAI shipped 0.155.1 on September 18, and has already pushed several 0.156.0 alpha builds since.

• The fix: new local TUI sessions now leave reasoning summaries disabled by default, resolving requests being rejected outright by model providers that don't support the reasoning-summary field.
• Context: an earlier 0.155.x build had turned reasoning summaries on by default, which worked fine against OpenAI's own models but silently broke sessions against third-party, OpenAI-compatible providers.
• Why it matters: if you point Codex CLI at a non-OpenAI provider, update to 0.155.1 (or later) now — the 0.156.0 line is already at alpha.7 as of September 19, so a stable release should follow soon.

GitHub Copilot: A Reworked Review Experience, Plus a Deprecation Notice

GitHub shipped two Copilot-relevant changelog entries on September 18, plus an admin-facing update the day before.

• Copilot code review, improved: the review experience now shows a clearer timeline, auto-resolves suggestions it judges addressed, and generates commit messages — rolling out across the review surfaces teams already use.
• Model deprecations confirmed: GitHub confirmed a set of Copilot models will retire in mid-October (October 19) across Chat, inline edits, and code completion — worth auditing now if your team pins a specific model.
• Impact dashboard update (September 17): admins can now see feature-level engagement numbers, useful for justifying — or trimming — seats ahead of the October model changes.

OpenCode Stays Quiet

OpenCode hasn't tagged a new release since v1.18.31 on September 14. Five days without a release is unusual for a project that's shipped almost daily this month, so a larger build is likely queued up rather than the project going dormant.

What This Means for Developers

Three practical takeaways from today's changes: check /status after updating Claude Code to confirm you're actually getting the server-side auto mode classifier and its cost savings; if your team pins a specific Copilot model, start planning a migration before the October 19 deprecation lands; and Codex CLI users pointed at a third-party provider should upgrade to 0.155.1 immediately to stop reasoning-summary requests from being rejected.

Sources & Further Reading

• Claude Code Changelog (official)
• Auto Mode Classifier Billing (official docs)
• OpenAI Codex CLI releases on GitHub
• Copilot code review: An improved review experience
• Upcoming deprecation of selected GitHub Copilot models
• GitHub Changelog — September 2026 archive
• OpenCode releases on GitHub