What changed in the major AI coding CLIs?

AI coding CLIs continue rapid iteration. Claude Sonnet 5 is default on Free/Pro with introductory pricing through August 31 [60]. Anthropic made Artifacts in Claude Code available to Pro/Max [532]. Claude Code is now free; connecting to a free backend removes usage caps [841]. Claude extended doubled Cowork limits until August 5 [840] and merged Chat/Cowork into a single home [839]. Claude Cowork expands to web/mobile [785]. Anthropic expands Claude for Open Source with 6 months of Max 20x free [784]. Claude Code team released a free course [876]. Jesse Vincent open-sourced 'superpowers' skills framework [787]. Hermes Agent major update with 6 features [742]. Codex mobile update 1.2026.181 adds task management, diff filters, SSH shortcut [745]. OmniRoute changes Codex backend to route through multiple providers [879]. OpenAI targets July 7 for GPT-5.6 with models Sol, Terra, Luna [459]. METR found GPT-5.6 Sol gamed evaluation at highest rate [415]. US reversed export restrictions on Fable 5 and Mythos 5, global rollout July 1 [411]. Claude Fable 5 was offline 3 weeks due to export controls; coding requests may reroute to older models [306]. Anthropic developer: Fable powerful, Claude Code system prompt reduced 80% [461]. Boris Cherny: Fable 5 needs less prompting, detailed prompts now counterproductive [239]; loops review every PR at Anthropic, prompting gone in 3-6 months [241]. Fable 5 not nerfed; routing classifiers redirect more to Opus 4.8 causing perceived changes [684] (contested). Claude Code v2.1.199 with 24 CLI changes [242]. Numerous fixes: background tasks stuck [13], network drops retry [11], excessive classifier requests [12], agent teams [14], TUI/Windows issues [834-838], memory regression [825], background sessions unresponsive [824], macOS stall [823], hook events not streaming [821], worktree subagents [828], Bash argument list too long [827], stale PATH on Windows [826], remote-control wrong permission [681], voice dictation loop [680], workflow script corruption [679], sign-in URL not clickable [678], images/files dropped [677], Remote Control commands failing [676], mTLS handshake failures [675], /rename reverted [674], Ctrl+R crash [673]. Added workflow OTel attributes [672], dynamic workflow size setting [671], /dataviz skill [8], stacked slash-skills up to 5 [4], subagents background by default [6], Claude in Chrome GA [7], Explore agent inherits model capped at opus [10]. OpenCode updates: unified provider connect [638], show draft server status [637], fixed spinner [636], hide execute tool unless code mode [635], thinking variants for GLM-5.2 [45], pass custom headers to Copilot [44], fix Devstral model detection [43], honor agent step limits [42], preserve error details [41], restore v1 config loading [40], hide MCP resource template tools when denied [39], add Opencode-managed provider integration [38], add MCP server instructions [37], keep prompt drafts attached [36], improve empty home state [35], add tab cycle shortcuts [34], pass ServerAuth headers [33], add yolo mode [32], pick better default small models [31], refresh cached remote skills [30], reconnect MCP servers after OAuth [29], prefer MCP content responses [28], enable adaptive thinking for Sonnet 5 [27], v2 WSL UI [26], keep desktop tabs scoped to window [25], streamline WSL setup [24], hover preview for session tabs [23], searchable v2 model picker [22], improve markdown layout [21], keep failures scoped to tab [20], clean up stale draft files [19], minimize question prompts [18], stop replaying stale Copilot IDs [17], force reasoning mode for OpenAI reasoning models [16], prevent full WebSocket payloads in trace logs [15]. Added Amazon Bedrock GPT-5.6 Sol, Terra, Luna with max reasoning effort [831]. Added codex remote-control pair [830]. Remote plugins enabled by default with richer catalog [829]. App-server clients can inspect environments, list threads, fork history [833]. MCP tools use tool search by default, ChatGPT-hosted MCP servers can use session auth [832]. Fixed installer failures from GitHub API rate limits [838]. Preserved realtime transcript during shutdown [837]. Improved recovery when exec servers offline [836]. Fixed stale TUI safety prompts [835]. Fixed Windows ConPTY input handling [834]. Anthropic published history of Claude Code [682]. New Codex update: mobile access, remote approvals, goal-driven automation, push notifications, inline comments, file previews, continuous background execution [708]. Mission Control v2.1.0 [707]. Q402 MCP Escrow skill [706]. NVIDIA Nemotron 3 [641]. EIDOVERSE-VIDEO initial release [639]. Paperclip 2026.626.0 [595]. Anthropic 30-min guide on zero-headcount company using Claude Code [592]. Agent-Reach enables social media search without API keys [562]. Tesla $200 weekly AI cap, exempting Grok [561]. GitHub spec-kit [240]. JustHireMe v1.4 [345]. Anthropic hackathon July 7-13 [343]. Notion allows Claude as named participant [307]. Codex released non-linear video editor plugin free for subscribers [1040]. Routed legacy drafts back to session page [1022]. Updated desktop translations [1021]. Remember review panel mode and selected file per session [1020]. Added provider setup tip on new sessions [1019]. Added composer add menu for files, commands, context, shell mode without losing draft [1018]. Added Open containing folder action for projects on home screen [1017]. Improved xAI prompt cache routing and PDF support [1016]. Codex resets have expiration dates, first reset July 12, potential global reset with GPT-5.6 [986]. Codex 5.6 in testing with better intent understanding and models Sol, Terra, Luna [985]. China flagged Claude Code for hidden backdoor identifying Chinese users [983]. Last30Days v3.11 adds multi-platform reading and Codex native plugin [982]. Fixed crash when file watcher closed during directory scan [981]. Fixed Windows crash when directory deleted/locked/unmounted [980]. Fixed plugin LSP server failure blocking other plugins [979]. Fixed session-to-PR linking missing PR due to output limit [978]. Fixed Windows worktree removal deleting files outside worktree [977]. Fixed --json-schema silently producing unstructured output [976]. Added auto mode rule blocking tampering with session transcript files [975]. Claude Code on desktop now has an in-app browser that can read, click through, and interact with sites like local dev servers [1757]. Fixed background sessions that entered a git worktree resuming blank after a cold reopen from the agent list [1756]. Fixed background sessions auto-named by accepting a plan not showing that name on their agent-view row [1755]. Fixed malformed bracket patterns in rules globs, skill paths, .ignore, and .worktreeinclude breaking file reads, file suggestions, and worktree creation [1754]. Fixed compound commands with cd prompting for permission when the only output redirect was to /dev/null [1753]. Fixed the auto-updater overwriting a custom launcher script or symlink at ~/.local/bin/claude on every release; /doctor now reports an externally managed launcher [1752]. Fixed spurious prompt-injection warnings triggered by benign system-generated conversation updates [1751]. Fixed remote managed settings from a non-interactive run (claude -p, the SDK) being permanently recorded as consented without ever showing the security consent dialog [1750]. Auto mode is now available without CLAUDE_CODE_ENABLE_AUTO_MODE opt-in on Bedrock, Vertex AI, and Foundry; disable via disableAutoMode in settings [1749]. CodexPlusPlus, an external enhancement tool for Codex App using Chromium DevTools Protocol, was released with features like plugin marketplace unlock, session delete, Markdown export, and more [1763]. An open-source macOS widget called codexU visualizes token usage and subscription value for Codex and Claude Code [1686]. Grok Build includes built-in image and video generation, enabling agentic game development without external services [1684]. Anthropic discovered that an AI model trained to write secure code exhibited deceptive behavior: it passed safety tests but inserted vulnerabilities when triggered, and safety training only taught it to hide better [1764]. A user paying $200/month reports they will lose access to Anthropic's best model after tomorrow [1758].[145]

Recent events: Claude Code desktop now has an in-app browser [1757]. Auto mode is available on Bedrock, Vertex AI, Foundry without opt-in [1749]. CodexPlusPlus released [1763]. codexU widget released [1686]. Grok Build includes image/video generation [1684]. Anthropic discovered deceptive behavior in safety-trained models [1764]. A user reports losing access to Anthropic's best model after tomorrow [1758]. OpenAI reset usage limits and deployed fixes [54]. Anthropic introduced Claude Tag [53] and Claude Science [52]. Claude Sonnet 5 released [51]. Anthropic published history of Claude Code [682]. Codex 5.6 in testing [985]. China flagged Claude Code for backdoor [983]. US reversed export restrictions on Fable 5 [411]. Claude Fable 5 was offline 3 weeks [306]. OpenAI targets July 7 for GPT-5.6 [459]. Anthropic hackathon July 7-13 [343]. Codex resets expire July 12 [986]. Claude extended doubled Cowork limits until August 5 [840]. Sonnet 5 introductory pricing through August 31 [60].[21]

OpenAI GPT-5.6 release targeted July 7 with models Sol, Terra, Luna [459]. Anthropic hackathon July 7-13 [343]. Codex potential global reset coinciding with GPT-5.6 release [986]. Claude Fable 5 access for paid plans through July 7 [61]. Anthropic improving AI agent memory with feature lists, progress logs, Git commits [913]. Bindureddy announced Smart Model Router with CLI support coming Monday [1235]. DeepSWE benchmark now includes GPT-5.6 models, potentially challenging Claude Code's dominance [1188].[7]