OpenAI's MCP Extensions make plugins feel native in ChatGPT
A new Apache-2.0 SDK adds sidebar, composer and file-type hooks over MCP; plus a 125B MoE now runs on one Strix Halo box and GPT-6.1 Sol hits the gateways.

Copy markdown
Plugins that act like ChatGPT features
OpenAI published mcp-extensions (Apache-2.0) with TypeScript (@openai/mcp-extensions) and Python (openai-mcp-extensions) SDKs that bolt sidebar entrypoints, composer @-mentions, custom file-type viewers and rich forms onto your MCP server. Your tool stops being a buried tool call and becomes a first-class surface inside ChatGPT.
A 125B MoE on a single Strix Halo box
Halogen 0.1.1 is a one-model, one-GPU server for Qwen3.8-Flash-Next on AMD Strix Halo (gfx1151): 115.6GB of 4-bit weights fit in 124GB unified memory on a Ryzen AI Max+ 395, decoding 34-42 tok/s with speculation and ~1,309 tok/s prefill at 32K across a 262K context, OpenAI-compatible on port 8731. Frontier-class agent work on a mini-PC instead of a rented GPU.
Update: GPT-6.1 Sol reaches the gateways
A day after launch, you can route to GPT-6.1 Sol without an OpenAI key: Vercel AI Gateway (openai/gpt-6.1-sol), OpenRouter and GitHub Copilot Pro+/Max/Business/Enterprise all carry it. Pricing holds at $2/$10 per Mtok ($0.10 cached reads) with a ~1.05M-token context, so you can drop it into your existing router today.