2026-05-22

1. xAI Grok 4.1 Fast + Agent Tools API — Entering the Agent Infrastructure War (score: 2.10)

• Grok 4.1 Fast: 2M context, xAI's "best tool-calling model," 50% reduction in hallucinations (vs Grok 4 Fast). $0.20/$0.50 per 1M tokens — the cheapest among May's frontier models.
• Agent Tools API: Developers can enable web search, X post search, code execution, and document search with just a few lines of code. Integrated execution on xAI infrastructure → no need to manage API keys, rate limits, or sandboxes.
• Real-time X post search is a unique xAI moat that no other frontier model has: specialized for real-time agents covering news and markets.
• Why it matters: xAI is directly countering Anthropic's MCP ecosystem and OpenAI's tool-call infrastructure with "managed agent infrastructure." The agent platform war is descending into the infrastructure layer.
• → Grok 4.1 Fast (xAI), xAI
• Sources: xAI blog, source

2. DeepMind Co-Scientist → Nature Paper + Researcher Rollout (score: 2.02)

• Co-Scientist (a multi-agent hypothesis generation system) was elevated from a research demo to a Nature paper (2026-05-21).
• Cambridge case: in infectious disease (sepsis) research, it discovered a protein candidate that human researchers had missed → shortening amino-acid characterization that would take 2-3 years to a 6-month target (~4× acceleration).
• Experimental rollout has begun via the Gemini for Science program based on individual researcher requests.
• Why it matters: Peer-reviewed evidence that AI generated a "hypothesis beyond the researcher's prior knowledge." Following AlphaFold (structure) and AlphaEvolve (algorithms), Co-Scientist (hypotheses) fills the third production layer of DeepMind's AI-for-science stack. A data point that could change wet-lab economics.
• → Co-Scientist (Google DeepMind), Google DeepMind
• Sources: DeepMind blog, source

3. OpenAI GPT-Realtime-2 — First Production GA for Voice Agents (score: 1.86,

• A GPT-5-class reasoning voice model: 128K context (4× increase), adjustable reasoning effort (minimal~xhigh), simultaneous tool calls.
• The Realtime API reaches GA for the first time — previously beta/experimental, now deployable in production.
• Launched alongside: GPT-Realtime-Translate (live interpretation in 70+ languages), GPT-Realtime-Whisper (streaming STT).
• Big Bench Audio +15.2%, Audio MultiChallenge +13.8% (vs GPT-Realtime-1.5).
• Why it matters: Voice agents that "speak while thinking" reach production readiness for the first time. GPT-5-level reasoning can now be brought to call centers, medical documentation, and multilingual services. The text reasoning-effort pattern extends to voice — progress toward unified cross-modality UX.
• → GPT-Realtime-2 (OpenAI), OpenAI
• Sources: OpenAI blog, source