The AI Catchup - August 12, 2026

Welcome back. Last issue was about headroom: Cursor doubled usage limits, Claude Code got a live iOS simulator pane, and Google priced Gemini for agent work. This issue is about coordination and access. SpaceXAI put Grok 4.6 in the Cursor picker with a one-week 2x usage window. Claude Code sessions can now discover and message one another. And OpenAI spent the week deciding who gets which model: a portable plugin format, Luna for Free and Go, and a gated cyber model that does not belong in a public picker.

Let us get into it.

At a Glance

This Week in AI

The model drop of the week is Grok 4.6. SpaceXAI released it on August 12, and Cursor shipped it the same day. The pitch is not a raw intelligence jump. It is persistence: stay with a complex task across many steps, research an unfamiliar domain, work through a codebase, or turn a product idea into a working first version. SpaceXAI says first passes on visual and interactive projects are stronger than with Grok 4.5, and that longer trajectories showed more self-testing before the model moved on.

The number the launch post wants you to remember is the composite: SpaceXAI says Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index. The same table labels that column GPT-5.6 Sol Max, at 61 vs 61. Fable 5 Max still leads that index at 62. The rest of the table is a split decision. Grok 4.6 High posts 69.9% on CursorBench v3.2 (ahead of Sol Max at 67.2%, behind Fable 5 Max at 70.5%) and 65.9% on DeepSWE v1.1 (up from 54% on Grok 4.5 High, still behind Sol Max at 73%). Terminal-Bench v3.0 is the caution row: 26% vs 34.6% for Sol Max. If your work looks like a long terminal agent loop, that is the bake-off we would run before changing a default.

Pricing is the part that needs no interpretation. The API starts at $2 per million input tokens and $6 per million output tokens, with cached input at $0.50. Cross 200,000 prompt tokens and those rates double for the whole request ($4 / $1 / $12). Context window is 500,000 tokens. The model name is grok-4.6. Reasoning effort goes low, medium, high (default), or xhigh, and SpaceXAI says reasoning cannot be turned off.

The reason to try it this week rather than next: SpaceXAI and Cursor are offering 2x included usage inside Cursor and Grok Build for the first week. That is a launch window, not the same thing as Cursor's permanent doubled first-party pool from July. The posts do not say whether the two stack, so do not budget as if they do. Pick Grok 4.6 in Cursor, rerun one real task, and spend the promo on evidence.

Claude Code Sessions Can Now Message Each Other

Independent Claude Code sessions can now discover and message one another. This is not resuming one conversation, and it is not an agent team. It is a plain-text ping between sessions that stay separate: a migration terminal can tell a docs terminal the schema landed, or you can ask "the session in my other terminal" whether the tests finished. The receiving session gets a sender name, a reply address when one exists, and the message text. It does not get the sender's files or transcript.

On the same Mac or Linux box, including WSL 2, sessions can start new exchanges over a per-user Unix-domain socket. Messages do not go through Anthropic servers. Across machines, or with Claude Code on the web, Remote Control carries replies only. The remote session has to send first. Native Windows and the listed cloud platforms are out.

The safety boundary is the part we would actually configure. A peer message is not user consent. It cannot approve a permission prompt, change CLAUDE.md, or run a command written inside the message. Inbound traffic can be set to accept, hold, or refuse. Cross-machine replies can require explicit approval even in bypass-permissions mode. Requires Claude Code 2.1.224 or later. Accepted messages waiting for Claude are capped at 50 per session.

Agent Plugins Packages Skills and MCP Once

OpenAI Developers introduced Agent Plugins, an open package format for Agent Skills and MCP server configs, developed with AWS, Cursor, GitHub, Code, and Vercel. Version 1.0.0 is a directory with a required plugin.json manifest, optional skills/ and mcp.json, and reverse-domain folders for client-specific extras. Spec: agent-plugins.org.

What it solves is duplication. Authors stop rearranging the same skill for every host. What it does not solve is everything else. Each client still controls installation, permissions, transports, and UX. A plugin that loads in one product is not automatically authorized in another. Check the client's compatibility list before you ship.

ChatGPT Adds a Reasoning Slider and Luna for Free and Go

OpenAI is folding reasoning effort into everyday ChatGPT instead of making us pick a second model. Plus and Pro users get updated GPT-5.6 Sol with a slider on web, mobile, and desktop, live since August 6. Keep it low for a quick answer; move it up for planning, research, or a coding explanation that actually needs the extra pass.

Free and Go users are moving to GPT-5.6 Luna as the default. OpenAI said unlimited text chats and a Think button would follow the default switch, subject to abuse guardrails. Unlimited means text. File uploads, images, and other tools keep their own limits. The GPT-5.6 Sol version that powers Work and Codex is unchanged.

OpenAI Expands Daybreak With GPT-5.6-Cyber

This one is not a public model drop. OpenAI split Daybreak into Blue and Red and shipped GPT-5.6-Cyber through Red for approved defenders. Blue pairs GPT-5.6 Sol with safeguards for vuln discovery, code review, malware analysis, and incident response. Red is purpose-trained models for authorized exploit research, validation, and red teaming.

Access requires identity verification, monitoring, legal attestations, and hardware security keys for individual accounts beginning September 1, 2026. OpenAI reports GPT-5.6-Cyber completed 95.0% of requests on its internal Advanced Cybersecurity Completion Rate eval, versus 2.0% for Sol on Blue and 57.3% for GPT-5.5-Cyber. Those are OpenAI's numbers, not an independent ranking. Teams with an authorized mandate apply at openai.com/daybreak/partners. For Codex, OpenAI is pushing auto-review over full-access.

Ship It This Week

Four direct wins, in order.

Spend the Grok 4.6 window on one real task. In Cursor, pick Grok 4.6 and rerun a multi-step feature or a visual first pass you already know. Compare it against your current default on the same prompt. The 2x included-usage week is the reason to measure now.

Update Claude Code to 2.1.224+ and ping another session. Leave one terminal on a migration or test suite. Ask the other to request a concise result instead of reopening the transcript. Set inbound to hold or refuse if you do not want unattended workers acting on peer text.

On Plus or Pro, use the ChatGPT reasoning slider on one hard prompt. Compare a quick setting with a higher-effort setting and watch both quality and wait time. On Free or Go, treat Luna as the new default and save Think for questions that actually need a second pass.

If you do not want Claude Code auto mode tomorrow, pin a default today. New Pro, Max, and Team sessions start in auto mode on August 14 unless you already pinned one. Shift+Tab still switches. Admins can set disableAutoMode.

Quick Hits

• Cursor Router is now a documented product, not a hidden Auto toggle. Auto Intelligence spends more budget on frontier models. Auto Balance keeps more traffic on the cheap path. Cursor's own production numbers: Auto Intelligence above Fable-level satisfaction at 68% lower cost; Auto Balance ahead of Opus 4.8 at 41% lower cost. Those are Cursor's measurements. You pick a cost-performance preference; the turn-level router chooses among Grok, Sol, Opus 5, and Fable 5.
• The ChatGPT desktop app can import from Claude Code, Claude Cowork, and Cursor. Codex CLI imports from Claude Code and Cursor. Instruction files, skills, plugins, MCP config, hooks, and recent chats move over. Importing is non-destructive. Review permissions before you run the imported agent. What actually transfers.
• ChatGPT Desktop is in Linux preview: Ubuntu 24.04 and 26.04, Debian 13, Fedora 43 and 44, .deb and .rpm, x64 and ARM64. Distro list.
• OpenAI reset ChatGPT Work and Codex limits for paid users on August 8. It is a one-time replenishment, not a new quota schedule, and it does not change API billing. Work and Codex still share an agentic pool.
• Codex Security Review is in research preview on GitHub pull requests. It reads the diff, repo context, and an optional threat model, then posts findings on the PR. Enterprise, Business, Edu, and Pro are listed; Plus is not. Comment @codex security review to run one manually.

---

That is it for this issue. The through-line this week is coordination and access: a new model you can try in Cursor without rebuilding your stack, Claude Code sessions that can ping each other instead of sharing a transcript, and OpenAI deciding who gets Luna, a portable plugin, or a gated cyber model. If a teammate is still waiting for a "real" reason to try Grok this week, or to update Claude Code, forward this -- the Grok window is time-limited.

Until next week, stay caught up.