Gemini 4 Argon: a WANDR SoTA in Perplexity, a Minecraft clear, and the safety report
From the AI digest for 2026-10-16
What happened: three datapoints on the Gemini 4 post-ship thread, all developments of "ai — Gemini 4: the Kavukcuoglu ship window" (Argon shipped September 30; developer/Pro access opened October 13). (1) Perplexity's Arav Srinivas (~11:05 AM ET October 16): "Gemini 4 Argon is a frontier model and has set a new SoTA on Wide-And-Deep-Research (WANDR) benchmark. Available now to all Perplexity users both on fast search and Computer (agent) modes" for Pro and Max subscribers — Argon landing inside a third-party harness with a harness-branded benchmark record. (2) Google DeepMind (~12:11 PM ET October 16): four Gemini 4 Argon agents beat Minecraft's Ender Dragon from scratch in 27 hours — "Operating in a sandbox using only code and text, they wrote 2,783 programs, divided up jobs over chat, shared notes, and debugged errors together - all with no human help" — the multi-agent claim sourced to Maxence Faldor's thread. (3) Back-filled in the window from October 13 (~2:22 PM ET): Anca Dragan's post relayed on the GoogleDeepMind account — "We did a lot to understand Gemini 4's capabilities and propensities, and to improve our mitigations for this launch" — pointing to the Gemini 4 model-safety blog and the gemini_4_safety_report.pdf.
Watch: the WANDR SoTA is vendor-adjacent on both sides (Argon scored by a Perplexity that now resells it); Minecraft is a sandbox demonstration, not an external eval, and the safety report is unread in the window — the thread's standing watch, independent replication of the board claims once access opens, is unmet. Separate watch: Perplexity's WANDR record is also new evidence for the Perplexity-thread question of how harness-bramed benchmarks travel as model evidence.
Links: Arav Srinivas — WANDR SoTA · Google DeepMind — Minecraft agents · Maxence Faldor · Anca Dragan via GoogleDeepMind