2026-05-19

1. 🆕 Google Gemini 3.5 Flash GA — Flash surpasses the previous Pro across the board

• At Google I/O 2026, Gemini 3.5 Flash went GA immediately (2026-05-19). The first model in the Gemini 3.5 family (source)
• Benchmarks: GPQA Diamond 90.4% (PhD-level science), MCP Atlas 83.6% (agents/MCP), Terminal-Bench 2.1 76.2%, CharXiv Reasoning 84.2%, MMMU-Pro 81.2%
• Beats Gemini 3.1 Pro on every item — the Flash tier fully overtakes the previous Pro tier. At the same time, 4× the output speed versus other frontier models
• Pricing: $1.50/$9.00 per 1M tokens (cache $0.15). Context: 1M input / 65k output
• Dynamic Thinking enabled by default. Immediately available in Gemini API, AI Studio, Antigravity, and Search
• Why it matters: A Flash tier surpassing Pro is a generational turning point that the previous Gemini family never reached. In particular, MCP Atlas 83.6% is among the highest publicly reported scores on the MCP agent-integration benchmark. We have entered an era where the "fast and cheap" model is also the "strongest agent" model.
• → Gemini 3.5 Flash, Google DeepMind

2. 🆕 Google Gemini Spark — Google enters the personal agent war

• At Google I/O 2026, Gemini Spark was announced — a 24/7 cloud-based personal AI agent (source)
• Email it directly via a dedicated Gmail address → Spark executes the task. Chrome web browsing. Full Google Workspace integration (Gmail · Docs · Slides) plus third-party apps (Canva, OpenTable, Instacart)
• Keeps running even when you're offline — handles long-running tasks in the cloud. Testers this week, US Ultra beta next week
• Why it matters: Spark is a channel for delivering agentic AI to hundreds of millions of existing Gmail/Workspace users with zero friction. Unlike OpenAI Codex (dev-first), Claude Code (MCP/coding-first), and Grok Build (terminal-first), Spark is consumer-first. This is the moment Google's platform advantage may play a substantive role in the agent race for the first time.
• → Gemini Spark, Agents (LLM Agents)

3. 🆕 Anthropic acquires Stainless ($300M) — vertically integrating MCP tooling

• Anthropic acquired Stainless, a startup specializing in SDKs and MCP servers, for roughly $300M (2026-05-18) (source)
• Stainless has auto-generated every official Anthropic SDK since the early days of the Claude API. TypeScript, Python, Go, and Java SDKs + CLI + MCP server auto-generation
• Key point: Stainless has also served OpenAI, Google, and Cloudflare, acting as "common infrastructure" for the AI development ecosystem. External customer service is set to end after the acquisition → competitors will need alternative tooling
• Why it matters: Anthropic is vertically integrating the core infrastructure for Claude API access (SDK + MCP server auto-generation). This is the most aggressive move yet in the strategy to cement the MCP ecosystem as "Claude's connection standard." If Anthropic monopolizes a tool that OpenAI and Google used jointly, Anthropic's leverage over the MCP ecosystem rises significantly.
• → Anthropic, Agents (LLM Agents)

4. Jack Clark (Anthropic co-founder): "RSI 60% probability, by 2028"

• Anthropic co-founder Jack Clark (currently Head of Public Benefit), Import AI #455 (2026-05-07) (source)
• Claim: "After reviewing 100+ data sources — a 60% probability that recursive self-improvement (RSI) is realized by the end of 2028, 30% by 2027"
• RSI definition: when AI meaningfully improves its own training process and architecture, and those improvements compound. A state where human researchers are no longer the primary driver of capability.
• Key evidence: SWE-Bench success rate going from 2% (Claude 2, late 2023) → saturating at 93.9% — this trajectory read as an early signal of a self-reinforcing loop
• Clarification: "This is not a 2028 AGI claim. It's a prediction that the recursive loop is established." — narrower, but still a significant claim
• Why it matters: This is a concrete probabilistic forecast from a frontier-lab insider (an Anthropic co-founder). Not an abstract "AI risk" discussion, but a time-bound forecast for a specific capability threshold. The very framing that reads SWE-Bench saturation as an RSI signal is an important indicator in the alignment field.
• → AI Alignment, Anthropic