This week in AI: ChatGPT Work, Meta's API, and Grok 4.5

Good morning,

OpenAI released GPT-5.6 to everyone and wrapped it inside a new long-running agent called ChatGPT Work. Meta opened its first model API with Muse Spark 1.1, while SpaceXAI launched Grok 4.5 at pricing aggressive enough to keep the model market uncomfortable.

GPT-5.6 reaches general availability

After a limited preview, OpenAI released the full GPT-5.6 family across ChatGPT, Codex, and the API.

OpenAI says Sol, Terra, and Luna are durable capability tiers that will carry forward as model generations change: Sol for maximum capability, Terra for balanced everyday work, and Luna for speed and affordability.

• Sol is the flagship model, priced at $5 per million input tokens and $30 per million output tokens.
• Terra targets everyday production work at $2.50 input and $15 output.
• Luna is the fastest and cheapest tier at $1 input and $6 output.
• A new ultra setting coordinates four agents in parallel by default for harder tasks.
• The API adds programmatic tool calling and a multi-agent beta for running concurrent subagents inside one request.

Artificial Analysis scored GPT-5.6 Sol one point behind Claude Fable 5 on its Intelligence Index at roughly one-third of the cost. Sol also leads its Coding Agent Index.

ChatGPT Work takes the fight beyond chat

OpenAI also launched ChatGPT Work, a long-running agent that can research across connected apps and files, then produce finished spreadsheets, presentations, documents, reports, and web apps.

Work can stay with a project for hours, ask for input when needed, and run scheduled tasks after the user steps away. OpenAI also introduced ChatGPT Sites, which turns this work into interactive dashboards, project trackers, internal portals, and lightweight apps.

The desktop product now combines Chat, Work, and Codex in one app for Windows and macOS. It can use local files, desktop apps, a built-in browser, plugins, and computer use with permission.

OpenAI is retiring its standalone Atlas browser on August 9 and moving its browser-based agent capabilities into ChatGPT and Codex. Atlas bookmarks, tabs, and browser history do not transfer automatically.

ChatGPT Work is rolling out first to Pro, Enterprise, and Edu users on web and mobile, with Plus and Business following. The unified desktop app is available across all plans.

Meta opens its model API

Meta released Muse Spark 1.1, a multimodal reasoning model built for coding, tool use, computer use, and long-running agent workflows.

It can manage a one-million-token context window, work with images, video, and PDFs, connect to MCP servers and custom skills, and coordinate parallel subagents. Meta says the model can decide when to automate a task with scripts and when to operate an interface directly.

The new Meta Model API is now in public preview for U.S. developers. Every new account starts with $20 in credits, followed by pay-as-you-go pricing of $1.25 per million input tokens and $4.25 per million output tokens.

Artificial Analysis scored Muse Spark 1.1 at 51, level with GPT-5.6 Luna, and estimated a cost of about $0.26 per benchmark task.

This gives developers a first-party Meta endpoint for building with Muse Spark, instead of limiting the model to Meta's consumer apps.

Grok 4.5 enters the cheap frontier tier

SpaceXAI launched Grok 4.5, a coding and agent model trained alongside Cursor and available through Grok Build, Cursor, and the SpaceXAI API.

• It supports text and image input with a 500,000-token context window.
• Standard API pricing is $2 per million input tokens and $6 per million output tokens, with cached input at $0.50. Costs double when prompts exceed 200,000 tokens.
• SpaceXAI says it serves at roughly 80 tokens per second and uses far fewer tokens than larger competitors on coding tasks.
• The API console is not yet available to EU users, with access expected later this month.

Artificial Analysis ranked Grok 4.5 fourth on its Intelligence Index and found its coding-agent performance level with GPT-5.5 at a much lower task cost.

At $2/$6 for standard requests, Grok 4.5 gives builders another credible low-cost option for coding agents.

Quick hits 🗞️

• ChatGPT Voice can now listen and speak at the same time. GPT-Live uses a full-duplex architecture for more natural interruptions and can delegate harder work to a frontier model without ending the conversation. GPT-Live-1 is rolling out to Go, Plus, and Pro users, while Free users get GPT-Live-1 mini. Source
• Claude's agents moved beyond laptop-bound workflows. Cowork sessions now run remotely on web and mobile, continue after your computer closes, and support scheduled tasks with no device online. The rollout starts with Max, with more plans following over the next several weeks. Claude Code desktop also gained a sandboxed browser that can open external sites, interact with pages, and test web apps directly. Cowork source · Browser source
• Google's managed Gemini agents gained background execution and remote MCP. Developers can run agents asynchronously inside Google's cloud sandbox, connect external MCP servers, call custom functions, and refresh credentials without rebuilding the environment. Source
• GPT-5.6 will become Microsoft 365 Copilot's preferred model. It will power Word, Excel, PowerPoint, Copilot Chat, and Cowork, with Microsoft also accessing GPT-5.6 directly through OpenAI's API. Source
• Meituan released a 1.6-trillion-parameter open model trained on alternative AI hardware. LongCat 2.0 activates roughly 48 billion parameters per token, trains on one-million-token context data, and ships under an MIT license. Meituan says the full training run used AI ASIC superpods rather than the usual GPU stack. Source

See you next week!

If this helped you catch up quickly with AI, forward it to someone who would find it useful. They can subscribe here.