Google invented the Transformer.

Most authors of that 2017 “Attention Is All You Need” paper were at Google Brain. But on July 2026 third-party leaderboards, not a single Google model places in the global top ten. On July 21, Google DeepMind shipped three models in one day — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, Gemini 3.5 Flash Cyber — yet the promised flagship Gemini 3.5 Pro was conspicuously absent, with the announcement noting only that it's “currently testing with partners and we plan to make it broadly available as soon as it's ready.”

A Triple Flash Release: Cheap, Fast, Good — But Not Flagship

Let's give the new products their due. Objectively, these three Flash models have real highlights.

Gemini 3.6 Flash is the workhorse, priced at $1.50/1M input tokens and $7.50/1M output tokens — cheaper than 3.5 Flash. On the Artificial Analysis Index, 3.6 Flash consumes 17% fewer output tokens than 3.5 Flash; on some benchmarks like Datacurve's DeepSWE, efficiency gains of up to 65% were observed. Coding jumped from 37% to 49% on DeepSWE; MLE Bench rose from 49.7% to 63.9%; OSWorld-Verified improved from 78.4% to 83.0%.

3.5 Flash-Lite is the value killer: just $0.30/1M input and $2.50/1M output, running at 350 tokens/s, and on SWE-Bench Pro (54.2% vs 49.6%) and OSWorld-Verified (74.0% vs 65.1%) it even outperforms Gemini 3 Flash. 3.5 Flash Cyber is a security-specialized model delivered through the CodeMender agent, available exclusively to governments and trusted partners.

The numbers look good. But here's the thing: these are all “Flash” tier, built for efficiency and cost-effectiveness, not maximum capability. In the frontier model race, the competition is over Pro/Ultra-class flagships — whose reasoning is strongest, whose multimodal is sharpest, whose agents are most reliable. On those dimensions, Google hasn't delivered a convincing answer in quite some time.

Behind the Flagship Delay: Google AI's Structural Dilemma

Gemini 3.5 Pro was originally promised for June 2026. It's now late July, and the official language is “still testing with partners.” This isn't the first time a Google flagship has slipped — Gemini Ultra also went through extended delays and capability controversies after launch.

Why does a company that invented the Transformer, houses a research organization as elite as DeepMind, and spends tens of billions annually on AI keep falling behind in the flagship race? Several structural factors stand out.

First, tension between Google's “research culture” and “product culture” persists. DeepMind scientists want to publish papers and do fundamental research; Google Cloud and product teams need stable, reliable models to serve customers. Coordination costs between these two cultures have been magnified dramatically in the LLM era. OpenAI is a “product-driven research” company — research's goal is productization. Google operates more like “research-driven products” — package research breakthroughs into products after they happen, a rhythm that naturally lags half a beat behind.

Second, the Flash-first strategy reflects a pragmatic bet on AI commercialization. In most enterprise scenarios, the most-used model isn't the smartest one — it's the one with the best cost-performance ratio. What Google Cloud needs in the enterprise market are models that can deploy at scale with predictable costs and low latency; the Flash line precisely serves this need. But relying solely on Flash over the long run concedes the “frontier capability benchmark” mindshare continuously.

Third, Google is already training Gemini 4. DeepMind explicitly noted in the announcement that “our most ambitious pre-training run yet, for Gemini 4, has started.” That means some resources originally slated for 3.5 Pro may have been redirected to 4.0. If Gemini 4 ships before year-end with generational improvements, the wait might prove worthwhile; if it slips again, Google's position in the AI race will become even more awkward.

See you tomorrow.