AI Coding Tools Roundup — August 2026: Claude Code, GitHub Copilot, OpenAI Codex & OpenCode

Four Tools, Four Ships: A Week of Real Movement in AI Coding

It has been a genuinely busy stretch for AI coding assistants. In the space of about two weeks, Anthropic, GitHub, OpenAI, and the SST-backed OpenCode project all pushed out substantive updates rather than the usual bug-fix trickle. Below is a rundown of what actually shipped, why each change matters if you use these tools daily, and where to go read the primary sources yourself.

Claude Code: Default Models, Idle Notifications, and Session Continuity

Anthropic's CLI agent had three notable releases in quick succession.

v2.1.236 (August 19) introduced the ANTHROPIC_DEFAULT_MODEL environment variable, which sets the model a brand-new session starts on. A manual /model pick still overrides it and persists across restarts, which is useful for teams standardizing on a specific model without stopping individual developers from switching mid-session. The same release added notify_when_idle to cross-session messaging, so one Claude Code session can ask another to ping it the moment it goes idle instead of polling.

v2.1.234 (August 17) made sessions automatically continue once usage limits reset, so a long-running task no longer silently stalls until you notice and restart it. Anthropic also shipped stronger protections against credential leaks in this release.

v2.1.233 (August 14) added GitLab merge request URL support alongside GitHub, and introduced optional memory cgroup limits for Bash tool commands on Linux, preventing a runaway build or install step from stalling an entire session.

Full changelog: Claude Code Docs

GitHub Copilot: Agent Plugins 1.0 and Persistent Memory

GitHub's release cadence stayed high through August, with the headline being a new open standard for extending Copilot across clients.

Agent Plugins 1.0 is a portable plugin format that works across compatible Copilot clients, complete with marketplace installs, shared skills, MCP server packaging, and enterprise controls for managing which plugins teams can install.

JetBrains gets persistent memory (August 11): Copilot for JetBrains can now retain and recall useful information across agent chat sessions, plus local model access via Ollama and additional enterprise controls. The same update improved everyday chat workflows and fixed reliability issues across MCP servers, terminals, and customizations.

Weekly releases (August 3): the integrated browser now supports element-level feedback, letting you select a specific element on a rendered page and attach a comment to it directly. Dictation switched to a multilingual, on-device model by default.

Heads up on deprecations: six older models are being retired from Copilot on August 31, 2026, affecting Chat, inline edits, ask and agent modes, and code completions across every Copilot surface. If you have model names hardcoded in scripts or CI, this is the week to check them.

Full changelog: GitHub Changelog (Copilot)

OpenAI Codex: A Linux Desktop App and a Leaner CLI

Codex's August updates lean toward platform reach and cleanup rather than headline features.

Linux desktop app preview: Codex now has a preview desktop app for supported Ubuntu, Debian, and Fedora systems, closing a long-standing gap for Linux-first developers who previously had to rely on the CLI or browser.

Agent Plugins support was added alongside persistent, manually ordered conversation sections, and the bundled OpenAI documentation skill was improved with more targeted official-source lookups for Codex, model selection, and API workflows.

Under the hood: the MCP SDK was upgraded to 3.0.0, Ratatui to 0.30.2, and V8 to 150.4.0. macOS release notarization now goes through Azure Key Vault instead of exporting private signing keys, and the deprecated codex exec --full-auto flag was removed in favor of --sandbox workspace-write.

Also worth noting: GPT-5.4 and GPT-5.4 mini leave Codex for ChatGPT-authenticated users on August 31, 2026, though both stay available via the OpenAI API and API-key-authenticated Codex sessions.

Full changelog: OpenAI Codex Changelog

OpenCode: Faster Logins and Smarter Compaction

The open-source, SST-backed OpenCode kept its usual fast, incremental pace with three releases in the first ten days of August.

• v1.18.16 (August 10): config parsing fixes, macOS app behavior fixes, and UI interaction fixes.
• v1.18.15 (August 7): fixed message ordering, UI loading states, and broadened locale coverage.
• v1.18.14 (August 5): simplified xAI login to a single device-code flow that works better in headless and remote environments.

Across these releases, repeated compaction now preserves earlier tool-call history in summaries instead of dropping orphaned results, blob-based attachments load correctly in the web UI, and you can export full session transcripts as JSON. Structured mid-stream provider errors are also preserved so compatible providers can retry failed responses instead of failing the whole turn.

Full changelog: OpenCode Changelog

Why This Matters

The common thread across all four tools is standardization and continuity: Agent Plugins is turning into a genuine cross-vendor standard now that both GitHub Copilot and OpenAI Codex support it, which should mean less lock-in for teams building internal tooling. Meanwhile Claude Code's session auto-continue and OpenCode's improved compaction both target the same pain point from a different angle — keeping long agentic sessions alive and coherent instead of forcing a restart. If you manage a team on any of these tools, the practical to-do list this week is short: check for hardcoded model names ahead of the August 31 deprecations, and try out Agent Plugins if you maintain any internal Copilot or Codex extensions.

Resources & References

• Claude Code changelog
• GitHub Copilot changelog
• GitHub Copilot: memory and Ollama in JetBrains
• OpenAI Codex changelog
• OpenCode changelog
• GitHub: Agent Plugins 1.0 announcement (weekly releases, Aug 3)
• GitHub: August 2026 Copilot model deprecations