AI Coding News — August 26, 2026: Claude Code's Three-Release Day
Yesterday's roundup closed with Claude Code sitting still at v2.1.241, unchanged since August 23. That didn't last. Over the following two days Anthropic pushed three more releases in a single 24-hour window, OpenCode quietly fixed a Cloudflare AI Gateway bug that was silently breaking third-party model routing, and OpenAI flagged the Codex CLI's mcp-server command for deprecation. Here's what actually shipped since our last check-in, and what's worth acting on today.
Claude Code's three-releases-in-a-day stretch
All three land on August 25 in the official Claude Code changelog. v2.1.243 is the biggest of the trio: it appends a Loops breakdown to /usage — per-loop run count, total tokens, tokens per run, and last-run timestamp — aimed at anyone running unattended /loop tasks and wondering where the token budget went. The same release adds a modelPicker setting for curating an ordered, labeled list in the /model picker, a promptCacheTtl / subagentPromptCacheTtl pair for tuning the 1-hour prompt cache, and keyless sign-in through /login → Anthropic Console.
- v2.1.243 (Aug 25): Loops breakdown in /usage, modelPicker curation setting, promptCacheTtl/subagentPromptCacheTtl, keyless Console sign-in, managed modelPricing setting for contracted per-model rates.
- v2.1.245 (Aug 25): a one-line fix for a startup crash on Linux distributions shipping glibc 2.44, including Arch Linux, CachyOS, and Fedora Rawhide.
- v2.1.246 (Aug 25): a wave of stability fixes covering background sessions failing after 45 seconds, severe transcript slowdown on very long lines like base64 blobs, plugin cache duplicating SHA-named directories, plus a new Auto mode tab in /permissions.
Earlier in the week, v2.1.236 added the ANTHROPIC_DEFAULT_MODEL environment variable for pinning which model new sessions start on, plus notify_when_idle for cross-session SendMessage. v2.1.238 added claude self-hosted-runner --defer-shutdown-max-min and a headersHelper mechanism so plugin marketplaces can mint HTTP headers on install. If you're on Linux and hit a startup crash this week, v2.1.245 is the fix — update now rather than working around it.
OpenCode patches Cloudflare AI Gateway and GitHub auth
OpenCode shipped four point releases across the same window. v1.18.20 (Aug 20) made failed subagent tool calls resumable via a task_id and hardened retry logic for network and capacity errors. v1.18.21 (Aug 21) fixed handling of unknown model finish-reasons and routed Vertex AI multi-region Gemini requests through REP endpoints. v1.18.22 (Aug 24) fixed device-login links breaking on servers that return relative verification URLs.
The one worth updating for is v1.18.23 (Aug 25), which fixes AI Gateway routing for third-party providers so non-Workers models actually work through Cloudflare's REST API, converts Anthropic's dotted model IDs like claude-haiku-4.5 to the dashed slug the gateway expects, and resolves a GitHub auth failure tied to immutable OIDC subject tokens. If you route Anthropic models through Cloudflare AI Gateway in OpenCode, this is the update that stops calls from silently failing.
OpenAI Codex CLI: an agents dashboard, then a deprecation notice
Codex CLI's stable channel moved from v0.148.0 (Aug 18) to v0.149.0 (Aug 20) to a v0.149.1 patch (Aug 24), per the Codex changelog. The headline feature is v0.149.0's interactive agents dashboard for searching, starting, opening, renaming, and stopping tasks, alongside new /cd, /pwd, and /cwd commands for managing the working directory mid-session, an expanded codex doctor that now diagnoses endpoint protection and network/proxy failures, and a codex queue command for messaging existing sessions. v0.148.0 had already added /export to Markdown, codex exec fork for session forking, and an archive/restore option in the resume picker.
The one thing worth acting on: OpenAI deprecated the codex mcp-server command on August 24. If anything in your pipeline still shells out to it directly, it's time to migrate to the supported path before it's removed.
GitHub Copilot: still riding last week's model wave
Copilot's weekly release cadence hasn't posted a new roundup since August 10, so nothing new landed in the stable changelog this week. What's still rolling out from that release is worth a reminder if you haven't tried it: Kimi K3 and MAI-Code-1.1-Flash arrived as new model options, Agent Plugins 1.0 went generally available across VS Code, the CLI, the SDK, and the Copilot app, and JetBrains picked up cross-session project memory. Combined with last week's Slack and Microsoft Teams integrations, Copilot's roadmap this month has been about surface area — more places to run the same agent — rather than new underlying capability.
What it means for developers
Two changes this week are worth acting on immediately rather than filing away: Claude Code v2.1.245 if you've hit a startup crash on a glibc 2.44 Linux box, and OpenCode v1.18.23 if you route Anthropic models through Cloudflare AI Gateway. Everyone else can treat this as routine maintenance — pull the latest release, skim the /usage and /model picker changes, and note that Codex's mcp-server path is on its way out. Nothing this cycle is a new model or a new agent, just the unglamorous work of three tools getting more reliable.
Sources & further reading
Comments
Share your thoughts and join the conversation