2026-07-22

1. Google ships three Gemini models in one day — and confirms Gemini 4 pre-training has begun

Gemini 3.6 Flash reached general availability on July 21 as the efficiency-focused stopgap while Gemini 3.5 Pro finishes its fourth rebuild. Headline numbers: SWE-Bench Pro 58.7% (vs. 55.1% for 3.5 Flash), DeepSWE 49% (vs. 37%), Computer Use 83% — with time-per-task cut from 2.7 to 1.3 minutes. Price: $1.50/$7.50/M tokens (slight premium over 3.5 Flash's $0.75/$3.50). Critically, the AA Intelligence Index is unchanged at 50 — this is a speed/efficiency upgrade, not a capability jump.

Alongside 3.6 Flash: Gemini 3.5 Flash-Lite launched at $0.30/$2.50 (350 t/s output), claiming the cheapest slot in Google's production lineup. And Gemini 3.5 Flash Cyber, a cybersecurity-specialized variant available only to governments and trusted partners via CodeMender, found 55 Chrome V8 vulnerabilities in a benchmark run (vs. 47 for base Flash, 36 for Claude Opus 4.6).

The most significant signal came in a throw-away line in the 3.6 Flash announcement: Gemini 4 pre-training has begun. No timeline, no specs — but the confirmation that Google is already training its next-generation architecture reframes the 3.5 Pro delays as a known pause in a longer roadmap, not a sign of strategic drift.

Why it matters: 3.6 Flash makes Google's coding lineup practical again while 3.5 Pro remains unreleased (now past three deadlines; prediction markets 81% July 31). The Cyber variant signals a new category — capability-gated security models — that we'll likely see from other labs soon. The Gemini 4 confirmation buys Google credibility time.

→ Gemini 3.6 Flash · Gemini 3.5 Flash-Lite · Gemini 3.5 Flash Cyber · Google DeepMind · Source

2. Jason Wei joins Meta Superintelligence Labs from OpenAI

Jason Wei — co-creator of chain-of-thought prompting (Wei et al. 2022) and the "Emergent Abilities of Large Language Models" paper — has joined Meta Superintelligence Labs (Meta SI), departing OpenAI where he led scaling and reasoning research. Wei's chain-of-thought work is foundational to all current reasoning models; his arrival at Meta SI gives the new lab meaningful bench strength in reasoning and alignment alongside last week's Noam Shazeer news.

Why it matters: Meta SI has now acquired three of the field's most-cited researchers in the reasoning/scaling space within two months (Shazeer, Wei, and the Dreamer acquihire earlier this year). This is a structural talent consolidation at Meta SI at exactly the moment the "Watermelon" model is in training. The meta-pattern: the two labs that most need credibility for their reasoning roadmaps (Meta SI with closed-model pivot, OpenAI post-Shazeer) are on opposite ends of this move.

→ Jason Wei · Meta AI · OpenAI · Source