The AI Catchup - August 28, 2026

Welcome back. Last issue was about coordination and access. This one has four headlines: Cursor Cloud Agents can subscribe to a pull request or a Slack thread and hold a goal until it is met, Claude Code can start a session from your phone instead of only continuing one, ChatGPT Voice can launch and steer Codex, and Claude Code picks itself back up after a usage-limit reset. Where a feature existed before in another form, we say exactly what changed.

Two deadlines land Monday. Details in Ship It.

Let us get into it.

At a Glance

What changed	What you can do now that you could not before	Can you use it today?
Cursor Cloud Agents subscribe to events and hold a `/goal`	Stop starting an agent per request. An agent watches a PR, a Slack thread, or a schedule, and works toward one objective until it is met	Yes for Cloud Agents. Subscriptions are cloud-only for now
Claude Code Remote Control starts sessions from your phone	Before, you exposed a session you had already started at the machine. Now you pick a directory from the Claude app and the session starts on that computer	Yes on Pro, Max, Team, and Enterprise. Team and Enterprise need an Owner toggle
ChatGPT Voice starts and steers Codex	Voice was a conversation. Now it can launch a Codex task, check the thread, and send follow-ups while you are away from the keyboard	Yes on Plus, Pro, Business, Edu, and Enterprise in the desktop app; iPhone via Remote
Claude Code auto-continues after a usage-limit reset	A session that hit the limit used to sit there until you came back. It now resumes on its own when the limit resets	Yes, on by default from 2.1.234. Opt out in `/config`
Codex 1M-token context for ChatGPT accounts	GPT-5.6 Sol's 1M-token window in Codex worked only with API keys. Now it works when you sign in with ChatGPT	Yes, as an option. The default context stays the default
Ox Alpha was Z.ai's GLM-5.3-Flash	The anonymous free model has a name, a price, and a retention answer. You can now choose it on purpose instead of gambling on it	Yes, paid. $0.075 / $0.25 per 1M until September 9, then $0.15 / $0.50
GPT-Image-2 transparent backgrounds	Generate an asset without an opaque background in one call, for PNG or WebP	Yes, in preview, in the Images API and Responses API

This Week in AI

The biggest workflow change of the fortnight is Cursor's Cloud Agents update. Until now the loop was one request, one agent run. As of August 19 a Cloud Agent can subscribe to an event source and wake when something happens: a pull request gets a comment, a Slack thread moves, a schedule fires. Cloud Agents automatically subscribe to the pull requests they create and drive them toward completion by fixing CI failures and addressing bot comments. Cursor's own framing is the right one: it moves from starting an agent for every discrete request to giving the agent a place to keep watching.

The second half of the change is /goal, which gives an agent a long-lived objective it works toward until it is fully complete. Cursor's example is /goal fix all flaky tests and make CI green, used in a new chat, and it pairs with a Custom Mode for following a playbook and /loop for recurring check-ins. Cursor does not promise every goal finishes without review, so read it as persistent task direction rather than an autonomy guarantee.

Two supporting pieces make the persistent loop safer to run. Subagents now get their own virtual machines, each with an isolated copy of the project, clean context, and its own cloud environment, so you can have them test the parent agent's changes in a fresh environment or swarm independent fixes without collisions. And steering no longer interrupts: a follow-up message waits for the next tool call instead of cutting the agent off mid-action. The published boundary is that subscriptions are cloud-only for now, and there is no new pricing tier in the changelog.

Claude Code Remote Control Can Now Start a Session, Not Just Continue One

Remote Control existed before this fortnight, so the change is worth stating precisely. The documented flow used to be: start Claude Code at the machine, enable Remote Control with --remote-control or /rc, and the session's URL or QR code opens it in the Claude app or claude.ai/code. Your phone continued something you had already begun.

Now a machine running claude remote-control appears as a device card in the Claude app's Code tab. Tap the machine, choose a directory, and a new session starts on that computer. You no longer need to have been sitting at the machine first. The execution boundary has not moved: code, filesystem, MCP servers, tools, and project configuration stay on the machine, and the phone is a control surface that can also show and stop subagents and workflows running in the background.

The same update tightened sync and reconnection. Resuming a session on the laptop keeps the phone on the live session instead of archiving it, the phone shows the session offline shortly after Claude Code exits, and if the laptop sleeps or the network drops, Claude Code reconnects automatically when the machine returns online and delivers the queued status updates from subagents and workflows. Plans: Pro, Max, Team, and Enterprise, with Team and Enterprise needing an Owner to enable the toggle. It requires a claude.ai login, does not work with an API key, and is unavailable on Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry.

ChatGPT Voice Can Now Launch and Steer Codex

The normal Codex workflow assumes you are looking at the desktop app, terminal, or editor while the agent works. ChatGPT Voice changes that loop: you can now describe a task aloud and have Voice start a Codex thread, check an existing thread, send follow-up instructions, and bring progress, blockers, and results back into the spoken conversation while the work continues. Voice was a conversation surface; it is now a supervisory layer over an agent.

The documented setup is specific. Open a new, empty chat or task in the ChatGPT desktop app, select Start new voice chat before sending anything, then ask Voice to start a Codex task. A chat started in another mode gets voice dictation, not a full Voice session. On mobile, Voice reaches Codex through Remote on iOS after pairing an iPhone with a desktop host, and the host is the machine that runs the work. Voice-driven tasks draw from the Codex usage budget, only one voice chat can be active across the desktop app at a time, and the documented plans are Plus, Pro, Business, Edu, and Enterprise.

Claude Code Resumes Itself After a Usage-Limit Reset

Until 2.1.234, a Claude Code session that reached a claude.ai usage limit stopped, and the documented options were to wait for the reset, upgrade, or buy usage credits; waiting meant remembering to come back. Claude Code now continues the session automatically when the limit resets, and it is on by default. The opt-out is in /config, under "Continue automatically at usage limit".

It is a scheduling convenience, not more allowance: the limit is unchanged. The line to note is the default's direction. A session left open at a limit will resume by itself unless you turned the setting off, which matters if you run unattended or usage-billed work. Anthropic's changelog gives the version, the default, and the setting name, and nothing about confirmation prompts or multi-session behavior, so do not infer those.

Three More Capability Gains

Codex's 1M-token GPT-5.6 Sol context now works through ChatGPT accounts. Per OpenAI's Tibo Sottiaux, the option previously worked only for API keys; the switch has now been flipped for ChatGPT sign-in as well. The default context length stays the default because it was tuned for performance and cost, so treat the larger window as a workload-specific setting and measure it on your repository before leaving it on.

Ox Alpha was Z.ai's GLM-5.3-Flash, and the free week is over. Two weeks ago it was an anonymous free listing; now it has a name, a price, and a retention answer. OpenRouter names ZAI as developer and operator, the stealth listing serves no providers, and OpenCode dropped Ox Alpha Free. The named model lists at $0.15 per 1M input and $0.50 output, halved to $0.075 and $0.25 through September 9, 2026, and it is a 320B-parameter model with 18B active and a 1M-token context. OpenRouter says prompts and completions were retained by the provider, which settles the contradiction between the two hosts against the comfortable reading. What you can do now that you could not before: pick it deliberately, from Z.ai directly, from twelve OpenRouter providers, or on OpenCode Go at $10 a month.

GPT-Image-2 can return transparent backgrounds. In preview since August 20, gpt-image-2 and gpt-image-2-2026-04-21 accept background: "transparent" in the Images API and the Responses API image-generation tool, so a product shot, icon, or illustration comes back ready to drop onto any layout without an opaque background. Request PNG or WebP; JPEG does not support it.

Ship It This Week

Two deadlines and three upgrades to your own workflow.

Move any ChatGPT scheduled task off gpt-5.4 before August 31. OpenAI's automations doc says a task pinned to gpt-5.4 or gpt-5.4-mini under ChatGPT sign-in must be updated before those models retire on August 31, 2026: replace gpt-5.4 with gpt-5.6-terra and gpt-5.4-mini with gpt-5.6-luna. A scheduled task fails silently, not loudly. Our scheduler comparison has the note.

Plan for Claude Code weekly limits to drop back after August 31. The +50% weekly-limit promotion has been extended three times and currently runs through August 31, 2026 at 11:59 PM PT. Treat clau.de/cc-50-promo as the source of truth if you are reading this near the date.

Give one Cloud Agent a /goal in a new chat and let it subscribe to the PR it opens. Pin your team's playbook as a Custom Mode first.

Start a Claude Code session from your phone. Run claude remote-control in a repo, pick the directory from the Claude app, and keep the first session visible on the machine so you can confirm permissions and MCP connections before going phone-only.

Decide which way you want auto-continue to point. If you run unattended or usage-billed Claude Code sessions, open /config and either leave "Continue automatically at usage limit" on deliberately or turn it off.

Quick Hits

Cursor is now part of SpaceX. The acquisition has closed and the Cursor team joins SpaceXAI to work on Grok Build, Grok Bot, Grok API, and Cursor. No new pricing, migration date, or rebrand was announced with the close.
Claude's computer use takes several actions per turn now, and its browser tool targets elements, not pixels. Computer use, browser use, the Skills API, and the Files API are generally available on the Claude Platform. Multi-action turns replace one action per model call (Anthropic reports one customer workflow going from 32 minutes to 13), the browser tool reads page structure alongside the screenshot, skills are versioned and run in Claude's sandbox, and files upload once and are referenced by file_id with 1 TB per organization.
Claude can send Gmail and manage Drive files. Send, reply, forward, share, move, trash, upload, with approval before each outbound email by default. Attachment content is not accessible, only its metadata.
GPT-5.6 Sol, Terra, and Luna are in Kiro. OpenAI reports roughly an 82% cost reduction per successful Terminal-Bench 2.1 task for Terra in Kiro's spec-driven environment. Vendor-reported, harness-specific.

For Admins and Finance

Account, billing, and workspace-management changes. Skip this section if you do not run a team.

OpenAI API-key spend dashboards and hard limits. Usage and Costs now group by API key, and an organization or project monthly spend limit can stop traffic with a documented 429 (organization_spend_limit_exceeded or project_spend_limit_exceeded in error.code). Enforcement is not instantaneous, so recorded spend can slightly exceed the cap.
Enterprise-managed auth for MCP connectors is GA. Team and Enterprise admins provision a connector once through the identity provider; users get it without a per-connector OAuth grant. Datadog, Notion, and Slack join the GA set.
Claude Security scans now run on Claude Mythos 5 for Claude Enterprise customers. Findings carry a CWE, confidence, severity, and a suggested fix that opens in Claude Code on the web; every patch needs human approval; billed as standard token usage. Scans do not grant direct Mythos 5 access.
Claude Managed Agents self-hosted sandboxes can attach memory stores, and web_search and web_fetch accept allowed_domains or blocked_domains.

That is it for this issue. The through-line is capability gained on tools you already use: an agent that keeps watching, a phone that can start the work, a voice that can run Codex, and a session that comes back on its own. If a teammate has a ChatGPT scheduled task pinned to gpt-5.4, forward this -- it stops working Monday.

Until next week, stay caught up.