Full May 2026 coverage: AI News May 2026.
xAI Opens Grok Build to SuperGrok and X Premium+ — How to Install
xAI expanded access to Grok Build today — its terminal-based coding agent with sub-agents, Gmail connectors, file handling, plugins, and inline image generation via Grok Imagine. Previously exclusive to SuperGrok Heavy at $300/month, Grok Build is now available to all SuperGrok ($30/month) and X Premium+ ($40/month) subscribers.
Install with one command:
npm install -g @xai/grok-build
Then authenticate: grok-build auth login. Version 0.1.218 ships alongside the access expansion, fixing Linux image paste, Windows shortcuts, and crash prevention during long sessions. Some users are reporting a 401 auth error on first login — fix: run grok-build auth logout, wait 2–3 minutes, and re-authenticate. Full install guide and feature breakdown →
Qwen3.7 Max Fully Live on OpenRouter and DashScope
Alibaba's Qwen3.7 Max — launched at the Alibaba Cloud Summit on May 20 — is now fully available via OpenRouter at $2.50/M input and $7.50/M output tokens, and through Alibaba Cloud DashScope. The model scored 56.6 on the Artificial Analysis Intelligence Index at launch, the highest ever recorded for a Chinese AI model, and leads all measured competitors on HMMT 2026 mathematics (97.1%), HLE with tools (41.4%), and the Apex benchmark (44.5). Its 35-hour autonomous coding run and 1M token context window make it the strongest choice for sustained long-horizon agentic tasks among Chinese-lab models. Note the verbosity caveat — add explicit output length constraints in your system prompt to avoid 4x output token inflation. Full Qwen3.7 Max review →
Meta Avocado — May Window Closed, June Confirmed by Sources
May ends in seven days. Meta Avocado has not been announced. Reuters sources from April said "May or June" — May is now effectively over for a model launch that needs marketing runway. Bloomberg sources this week confirmed June as the active target, with Meta's AI leadership under pressure to ship before the narrative hardens around GPT-5.5, Claude Opus 4.7, and Gemini 3.5 Flash as the established frontier trio. Internal tests reportedly showed Avocado performing between Gemini 2.5 and Gemini 3.0 — below what is needed to compete credibly on developer benchmarks. A June launch that benchmarks below the current frontier does reputational damage, but a further delay does more. Expect an announcement in the first two weeks of June.
OpenAI Codex Mobile — Fully Live for Pro Subscribers
OpenAI confirmed Codex mobile is fully live on iOS and Android for Codex Pro subscribers ($200/month). Developers can now queue coding tasks from their phone, monitor agent progress, approve or reject proposed changes, and launch new Codex sessions remotely. The mobile interface shows a real-time log of what Codex is doing — file reads, writes, test runs, and API calls — with a one-tap approve/reject for each proposed change before it commits to the codebase. Standard ChatGPT Plus and Pro subscribers get read-only monitoring of active Codex sessions; full autonomous control requires Codex Pro.
Microsoft Copilot Studio Computer Use — First Enterprise Feedback
Two days after Microsoft made Copilot Studio computer use agents generally available, enterprise teams are reporting initial results. The most common use case in early feedback: automating workflows in legacy ERP systems (SAP, Oracle) that lack modern APIs. Copilot Studio's visual reasoning approach — navigating software by reading the screen rather than calling an API — works particularly well on these systems precisely because they were never designed for programmatic integration. Failure cases reported: multi-step workflows involving CAPTCHAs or MFA prompts, which break the agent's visual navigation. Microsoft is aware of both and has flagged them as known limitations in the GA documentation.
Thinking Machines Previews Real-Time Voice, Video, and Text AI
Thinking Machines previewed Interaction Models for near-real-time voice, video, and text AI — built to listen, speak, see, and use tools simultaneously. The demo showed a model that can hold a video conversation, read documents shared on screen, and execute tool calls (web search, code execution, file management) without pausing the conversation. The architecture targets the gap between current voice AI (which handles one modality at a time sequentially) and true simultaneous multimodal interaction. No pricing or availability timeline was announced — Thinking Machines described it as a research preview.
Full May 2026 coverage: AI News May 2026 →