🤚 The Open-Palm Inventory

Google DeepMind released three new models on Tuesday, and if you’re thinking “finally, the 3.5 Pro they promised five months ago,” prepare to be efficiently disappointed.

What shipped:

• Gemini 3.6 Flash — Google’s self-described “workhorse model,” featuring improvements in coding, knowledge work, and multimodal performance, plus a 17% reduction in token usage. The workhorse is now cheaper to feed.
• Gemini 3.5 Flash-Lite — The most cost-effective model in its class, for customers who looked at Flash and said “but what if it was less?”
• Gemini 3.5 Flash Cyber — A cybersecurity-specific model fine-tuned for identifying and fixing vulnerabilities, available exclusively to governments and “trusted partners” through a limited access pilot program. Your startup is neither of those things.

Google’s stated priority: “efficiency, latency, and reliability to customers that are building AI agents at scale.” Three words that would be more convincing if the company had also shipped the model everyone is building agents with.

👐 The Two-Handed Deflection

The elephant in the model registry is Gemini 3.5 Pro, last updated in February 2026. Google explicitly promised its release in May, noting the model was “already being used internally, and we look forward to rolling it out next month.”

That was two months and zero rollouts ago.

Bloomberg reported that the delay stems from the model struggling to meet internal performance objectives, which is the diplomatic way of saying it’s not good enough to put Google’s name on, in a market where Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol have spent the summer trading benchmark records like Pokémon cards.

DeepMind product lead Logan Kilpatrick told reporters the team is “currently testing Gemini 3.5 Pro with partners” and hopes to “land soon.” In product management, “land soon” means somewhere between “next week” and “when we figure out why it keeps hallucinating the wrong capital of Australia.” The model is in the quantum state of being both “already used internally” and “not ready for you.”

But here’s the real magic trick. In the same breath that Kilpatrick declined to commit to a 3.5 Pro timeline, he announced that DeepMind has begun “its most ambitious pre-training run yet for Gemini 4.” This is the corporate equivalent of your mechanic saying your car isn’t ready, but would you like to hear about the next car?

🌿 The Gentle Awakening

There is something deeply familiar about Google’s relationship with its own product roadmap. The company that invented the Transformer architecture — the foundation of every large language model that has made every competitor billions of dollars — continues to operate as though being first to invent something and first to ship it are two entirely different departments. Which, at Google, they are.

Three Flash models in a single day is not laziness. It’s a company leaning into what it’s good at: infrastructure. The workhorse, the budget tier, the government special. Google is not trying to win the frontier model race today. It’s trying to be the cheapest, fastest, most reliable option for enterprise customers who need their AI agent to process 40 million invoices, not compose a sonnet about regulatory capture.

The problem is that the market has decided frontier models are the scoreboard, and Google’s scoreboard entry is five months late with no ETA.

👑 The Gold-Leaf Reckoning

Let us appreciate the strategic choreography. Google simultaneously:

• Shipped three models nobody was waiting for
• Did not ship the one model everybody was waiting for
• Pre-announced the generation after the one it hasn’t shipped yet
• Created an entire government-exclusive cybersecurity model to remind Washington that Google is a serious defense partner, thank you very much

This is not a product launch. This is a press release designed to change the subject. And it’s working — we are now discussing Gemini 4’s ambitions instead of Gemini 3.5 Pro’s absence, which is exactly the point.

Meanwhile, Anthropic is rumored to be in acquisition talks with robotics startup Physical Intelligence, OpenAI is racing toward an IPO with Sol as its flagship, and Moonshot AI just dropped 2.8 trillion parameters at one-third the price of anything Google charges. The frontier is moving, and Google’s most anticipated model is still in the testing phase with unnamed partners who are presumably sworn to secrecy and equipped with industrial-grade patience.

The good news for Google is that it has more compute, more data, and more infrastructure than any AI company on earth. The bad news is that having the best kitchen in the world doesn’t matter if the chef keeps promising the soufflé is coming next month.

“We are excited to announce three models you didn’t ask for and look forward to shipping the one you did at a date we are unable to specify but which we assure you is ambitious.” — The Slap of Wisdom Product Launch Desk, previewing the preview of the roadmap to the roadmap