Skip to content
Antique Cartographer's Map style editorial illustration for the news article: Claude Sonnet 5 1M context, Cursor cloud agents, and Codex Remote GA (mo

Claude Sonnet 5 1M context, Cursor cloud agents, and Codex Remote GA (model launch + agent tools + mobile hook)

By Harsh Desai
Share

TL;DR

Claude Code, Cursor, and Codex pushed agent automation, Sonnet 5 defaults, iOS betas, and remote control features across the week.

What shipped

Across this week Cursor, Claude Code, and OpenAI Codex released updates centered on cloud agents, background execution, and mobile access. Claude Code made Sonnet 5 the default model with a native 1M-token window and promotional pricing. The releases add practical controls for workflows that span local machines, cloud instances, and mobile devices.

Cursor

Cursor shipped ten updates that expand cloud agent control, automation triggers, and team marketplace features. The changes let users spin up isolated environments quickly and manage plugins from a single page. Mobile beta access and organization-level MCP distribution round out the set.

  • Cloud agents in Agents Window Cursor lets users run and manage cloud agents directly in the Agents Window with isolated environments set up in under 10 minutes and session switching via /in-cloud.
  • Automate skill and triggers Cursor added the /automate skill so users build workflows triggered by Slack emojis or five GitHub events while cloud agents access virtual computers for demos.
  • Customize page and plugin canvases Cursor introduced a central Customize page with marketplace leaderboard, prebuilt canvases, and GitLab or Bitbucket imports for team plugin management.
  • iOS public beta Cursor for iOS reached public beta on paid plans, letting users launch cloud agents, direct desktop agents via Remote Control, and track progress with Live Activities.
  • Team MCPs and groups Cursor allows admins to configure Team MCP servers once for distribution across agents, IDE, and CLI with access restricted by organization groups.

Claude Code

Claude Code delivered 49 updates dominated by Sonnet 5 as default model, background subagents, and permission controls. The releases add stacked skills, dynamic workflow sizing, and Chrome availability while tightening security and telemetry. Many fixes target stability for remote and headless sessions.

  • Background subagents default Claude Code now runs subagents in the background by default with notifications and adds a /dataviz skill plus AWS Claude Platform support.
  • Stacked slash-skills support Version 2.1.199 loads up to five stacked slash-skills and gives immediate guidance on SSL errors while preserving partial streams during overloads.
  • Manual permission default Claude Code sets Manual permission mode as default across CLI and IDEs and disables auto-continue for AskUserQuestion dialogs.
  • Sonnet 5 role reminders removed Claude Sonnet 5 sessions drop mid-conversation system role reminders to streamline flow.
  • Dynamic workflow size setting Claude Code adds a Dynamic workflow size option in config and includes workflow attributes in OTel telemetry.
  • Login expiration warnings Version 2.1.203 shows login expiration warnings, a manual permission badge, and session directories for MCP roots.
  • Hook event streaming fix Claude Code fixes hook event streaming in headless sessions to prevent remote workers from being idle-reaped mid-hook.
  • Auto mode security rules Version 2.1.205 blocks session transcript tampering and streams auto-update downloads to disk to cut memory use by 400 MB.
  • Directory suggestions and doctor Claude Code 2.1.206 adds directory path suggestions to /cd and a /doctor check that trims CLAUDE.md files.
  • Organization default models Claude Code supports admin-set organization default models and clickable file attachments in chat.
  • Rewind command added Claude Code added /rewind to restore conversations cleared by /clear.
  • ClassifyAllShell setting Claude Code added the autoMode.classifyAllShell setting to route all shell commands through the classifier.
  • Mouse click toggle and dictation CLAUDE_CODE_DISABLE_MOUSE_CLICKS option added and voice dictation fixes applied.

OpenAI Codex

OpenAI Codex released 36 updates that integrate Codex into ChatGPT desktop and iOS, add remote control, and expand to new regions. Record & Replay, branch management, and rate-limit banking give users more workflow reuse and control. CLI releases add proxy support and usage tools.

  • ChatGPT iOS 1.2026.153 ChatGPT for iOS 1.2026.153 adds branch selection, worktree creation, environment scripts, Codex profile screen, and /goal support.
  • Codex app 26.608 Codex app 26.608 introduces import flows from Claude Code, a revamped plugins screen, and expanded Settings search.
  • Codex app 26.609 Codex app 26.609 adds rate-limit reset banking, Developer mode, the /init command, and customizable macOS Dock icons.
  • ChatGPT iOS 1.2026.160 ChatGPT for iOS 1.2026.160 adds workspace file browser, directory picker, diff controls, MCP approval choices, and LaTeX rendering.
  • EEA UK Switzerland rollout Codex app features including Computer Use and Chrome extension now reach users in EEA, UK, and Switzerland.
  • Codex app 26.616 Codex app 26.616 introduces Record & Replay on macOS, bulk actions for history, and thread handoff between hosts.
  • ChatGPT iOS 1.2026.167 ChatGPT for iOS 1.2026.167 adds per-host personality settings, direct goal editing, and links from forked conversations.
  • Codex Remote GA Codex Remote reached general availability with secure QR pairing and a new DigitalOcean plugin for remote hosts.
  • ChatGPT iOS 1.2026.181 ChatGPT for iOS 1.2026.181 adds Codex task management from conversations, branch filters, and SSH shortcuts.
  • ChatGPT desktop integration Codex is now integrated into the ChatGPT desktop app for Markdown editing and GitHub PR review.
  • Codex CLI 0.142.0 Codex CLI 0.142.0 adds usage credit redemption via /usage and rollout token budgets.
  • Codex CLI 0.142.1 Codex CLI 0.142.1 adds opt-in Windows system proxy support for authentication.
  • Codex CLI 0.142.2 Codex CLI 0.142.2 enables tool search by default for MCP tools and adds macOS proxy support.
  • Codex CLI 0.142.5 Codex CLI 0.142.5 prevents full WebSocket payloads from being written to trace logs.
  • Sites preview launch Codex app offers Sites in preview for creating and deploying websites hosted by OpenAI.

What this means for you

For Vibe Builders: You can now run cloud agents from Cursor iOS beta and trigger automations with Slack emojis or GitHub events. Claude Code background subagents and Codex Remote let you keep working while tasks finish on other machines. These releases move agent workflows from chat to scheduled, multi-device execution you can monitor without staying at the keyboard.

For Non-techies: Codex Remote and Cursor iOS beta let you start and approve coding tasks from your phone while agents run in the background. Claude Sonnet 5 default gives longer context for business documents without extra setup. Record & Replay in Codex turns repeated steps into reusable skills you can trigger later.

For Developers: Claude Code now defaults to Manual permission mode and adds OTel attributes plus dynamic workflow sizing for production agent runs. Cursor Team MCP distribution and Codex CLI proxy support give tighter control over enterprise environments. Evaluate Sonnet 5 1M context against your current stack and test stacked slash-skills before rolling into CI pipelines.

What to watch next

Watch Claude Code for further Sonnet 5 prompt tuning and Cursor for deeper GitLab marketplace adoption. Track Codex CLI releases for additional proxy and token budget controls. Check region rollout status for Computer Use features in the coming days.

Harshs take

The week shows heavy overlap in agent management features across three vendors, with many near-duplicate announcements on background execution and permission defaults. This suggests rapid feature parity chasing rather than differentiated capabilities. Second-order effect is increased user choice in runtime but higher switching costs as each tool locks teams into its own MCP and skill formats.

Builders should pick one primary agent runtime this week and benchmark it against their existing stack before the next wave of updates lands.

by Harsh Desai

More AI news

Everything AI. One email.
Every Monday.

New tools. Model launches. Plugins. Repos. Tactics. The moves the sharpest builders are making right now, before everyone else.

No spam. Unsubscribe anytime.