What changed in the major AI coding CLIs?

Kimi K3, a 2.8 trillion parameter open-weight model, matches Fable 5 and GPT 5.6 on benchmarks at lower cost [5890]; it is now integrated into the Codex app with a 5-minute setup [5893]. The Claude Code team built an agent loop system of 18,000 agents, achieving 140,000 steps with 90% zero human typing [5891]. NVIDIA released DeepStream 9.1, adding 13 agentic skills for building video analytics pipelines via natural language, compatible with Claude Code and Codex [5888]. Anthropic released 'Frontend Design', an official skill for Claude Code that forces aesthetic direction before coding, with 161,000 stars [5886], and an official prompts library with ready-to-use prompts for development workflows [5885]. Moonshot AI launched Kimi Code CLI, an open-source MIT-licensed CLI that clones Claude Code with additional features [5883]. Claude for Teachers offers free premium Claude capabilities to verified K-12 educators in the US, with teaching skills and evidence-based curricula mapped to academic standards [5880]. Codex desktop app received numerous fixes: removed obsolete Codex workaround [5879], preserved composer caret after request updates [5878], resynced timeline correctly after reconnecting [5877], loaded more timeline history at once [5876], fixed project picker positioning crash [5875], kept permission auto-accept state separate for each server [5874], fixed connected project picker positioning when anchor is still loading [5873], fixed spacing between model provider sections in Settings [5872], fix(core): tolerate AlreadyExists in FSUtil.ensureDir [5871], hid drawer close button on Windows [5870], stopped subagents from launching nested subagents by default with configurable subagent_depth limit [5869], home page command palette search can find and open sessions [5868], fixed custom agent selector visibility [5867], fixed startup readiness including WSL server loading [5866], improved dangerous-command detection [5865], fixed plan approvals without edits being labeled "(edited by user)" [5864], fixed hook callback timeout misreported as user rejection [5863], fixed session crash when tool result renderer returned bigint or plain text [5862], fixed paste markers leaking into external editors [5861], fixed rendered text fragment leaking into crash telemetry [5860], fixed ultracode keyword opt-in firing on non-human-originated input [5859], added startup warning for Write/NotebookEdit/Glob permission rules [5858], added live elapsed-time counter to collapsed tool summary line [5857], fixed edits that leave input as "?" being silently swallowed [5856], fixed file upload validation for DOS device suffixes and multiple hard links [5855], fixed Claude Code on Vertex and Bedrock attempting default Opus model at startup [5854], fixed auto mode overriding PreToolUse hook's ask decision for unsandboxed Bash [5853], fixed permission previews not neutralizing bidirectional-override characters [5852], fixed shell mode not executing commands containing file paths while autocomplete popup was open [5851], fixed SIGTERM during Bash tool orphaning process tree [5850], fixed continue:false hook's halt being dropped [5849], fixed plan mode auto-running file-modifying Bash commands without permission prompt [5848], added session-wide limit on WebSearch tool calls (default 200) [5847], added message.uuid, client_request_id, and tool_source to OpenTelemetry log events [5846], added ISO modified timestamp to memory file frontmatter [5845], fixed Bash permission checks for help/man commands [5844], fixed Bash permission checks for zsh variable subscripts [5843], fixed Bash permission checks for very long commands [5842], fixed Bash permission checks for file-descriptor redirect forms [5841], fixed permission-check bypass in Windows PowerShell 5.1 [5840], fixed single-segment dir/** allow rules [5839].[49]

Kimi K3, a 2.8 trillion parameter open-weight model, matches Fable 5 and GPT 5.6 on benchmarks at lower cost [5890]; it is now integrated into the Codex app with a 5-minute setup [5893]. The Claude Code team built an agent loop system of 18,000 agents, achieving 140,000 steps with 90% zero human typing [5891]. NVIDIA released DeepStream 9.1, adding 13 agentic skills for building video analytics pipelines via natural language, compatible with Claude Code and Codex [5888]. Anthropic released 'Frontend Design', an official skill for Claude Code that forces aesthetic direction before coding, with 161,000 stars [5886], and an official prompts library with ready-to-use prompts for development workflows [5885]. Moonshot AI launched Kimi Code CLI, an open-source MIT-licensed CLI that clones Claude Code with additional features [5883]. Claude for Teachers offers free premium Claude capabilities to verified K-12 educators in the US, with teaching skills and evidence-based curricula mapped to academic standards [5880]. Codex desktop app received numerous fixes and improvements as listed in the current state.[8]

No upcoming catalysts are mentioned in the provided claims.