Monday, May 25, 2026

Andrej Karpathy Joins Anthropic; Managed Agents Gets MCP Tunnels and Self-Hosted Sandboxes

In a major talent signal, OpenAI founding team member and deep learning educator Andrej Karpathy has joined Anthropic as a research scientist, citing the company's frontier R&D momentum. Separately, Anthropic expanded Claude Managed Agents with two new May features: MCP tunnels (OAuth Client ID support for HTTP MCP servers including Slack, GitHub, and Figma) and self-hosted sandboxes now in public beta, giving enterprises more control over code execution environments.

The company also launched Claude for Small Business, a package of pre-built connectors and ready-to-run workflows embedding Claude inside everyday small-business tools.

Code with Claude Conference Previews the Future of AI-Assisted Development

Anthropic's annual developer conference in San Francisco drew engineers and founders for two days of keynotes and workshops. MIT Technology Review called it a glimpse of "coding's future — whether you like it or not." Highlights included Claude Code's Agent Teams (multiple Claude instances working on parallel git worktrees), the claude ultrareview command for cloud code review in CI, and expanded Opus API limits. Claude Code now supports a 1M-token context window with Opus 4.7 (87.6% on SWE-bench).

Google I/O 2026: Gemini 3.5 Flash, Gemini Spark Agent, and the End of Vertex AI

Google's biggest I/O in years (May 19) delivered a sweeping agentic pivot. Gemini 3.5 Flash is a new frontier-class model at roughly half the price of comparable offerings, designed for mass deployment across Google's consumer and enterprise products. Gemini Spark, now in beta for AI Ultra subscribers, is a 24/7 personal AI agent that can reason across Gmail, Calendar, Docs, and connected apps — taking action autonomously "under your direction."

On the infrastructure side, Google retired the Vertex AI brand entirely, replacing it with the Gemini Enterprise Agent Platform — a fully re-imagined agentic cloud for enterprise customers. As of May 21, Vertex AI no longer appears in the Google Cloud Console.

ChatGPT Business Gets Richer Codex Workflows and C2PA Image Provenance

OpenAI rolled out two notable updates this week. ChatGPT Business now integrates richer Codex workflows with "appshots" (reproducible workflow snapshots), generally available Goal Mode, improved browser annotations, remote locked computer use, and new admin analytics for enterprise adoption tracking.

On the trust and safety front, OpenAI strengthened AI content provenance with full C2PA conformance, SynthID watermarking for generated images, and a public preview of a verification tool that checks whether an image was made with OpenAI tools and surfaces its Content Credentials.

xAI Launches Speech-to-Text, Text-to-Speech, and Grok Voice Agent APIs

xAI expanded its API platform with three new general-availability launches. Speech-to-Text supports audio transcription in 25 languages with batch ($0.10/hr) and streaming ($0.20/hr) modes. Text-to-Speech enables natural-sounding speech synthesis from text via Grok. The flagship addition is grok-voice-think-fast-1.0, a voice agent API model optimized for complex multi-step customer support and enterprise workflows — emphasizing low latency, accurate tool use, and strong real-world performance. The Batch API was also updated to support image generation, image editing, and video generation alongside chat completions.

Cursor 3.4: Parallel Agent Dev Environments and Bugbot Autofix on PRs

Cursor 3.4 (shipped May 13) introduced dev environment configuration for parallel agent fleets, letting teams spin up isolated environments for concurrent AI agents. The standout feature is Bugbot Autofix: Bugbot now reviews open PRs, identifies issues, spins up a cloud agent to devise and test a fix, and proposes the fix directly on the pull request — with xAI reporting ~78% self-resolution on issues Bugbot claims to fix. JetBrains integration also arrived, bringing Cursor's agent capabilities to IntelliJ, PyCharm, and WebStorm via the Agent Client Protocol.

Meta Unveils Muse Spark, First Model from Alexandr Wang's Superintelligence Labs

Meta launched Muse Spark, its first flagship LLM built under Chief AI Officer Alexandr Wang's newly formed Superintelligence Labs. The model delivers competitive performance on multimodal perception, reasoning, health AI, and agentic tasks — at a fraction of the compute cost of the older Llama 4 mid-size variant. The launch marks a strategic pivot for Meta's AI unit, with Wang positioning Superintelligence Labs as Meta's answer to OpenAI's research-first culture.

Apple to Open Apple Intelligence to Third-Party AI Providers; Big Tech Capex Keeps Climbing

Apple is preparing a major platform shift internally codenamed "Extensions" that would let users choose third-party AI providers — including Google and Anthropic — to power Apple Intelligence features across iOS 27, iPadOS 27, and macOS 27. Providers would integrate via App Store applications. Full details are expected at WWDC in June.

Meanwhile, the industry's AI capex race shows no signs of slowing: Alphabet raised its 2026 guidance to $180–190B, Meta to $125–145B, and Microsoft guided Q4 capex above $40B with a $190B full-year target — signaling that frontier model training and inference infrastructure remain the defining investment of the decade.