Google released three new Gemini models this week — and buried the biggest news in the fine print.
On July 21, 2026, Google DeepMind quietly rolled out Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a specialized Gemini 3.5 Flash Cyber security model. The release was notable both for what it delivered and what it conspicuously omitted: Gemini 3.5 Pro, the flagship model that has now missed its release deadline multiple times, remained absent. But tucked into the announcement was a revelation that eclipsed everything else — Google has already begun pre-training Gemini 4, which it describes as “our most ambitious pre-training run yet.”
Three New Models, One Big Headline
The centerpiece of this week’s release is Gemini 3.6 Flash, an upgrade to Google’s “workhorse” Flash line designed for developers and enterprises who need capable AI at scale. Google positioned the model around efficiency: despite meaningful performance gains, Gemini 3.6 Flash produces roughly 17% fewer output tokens than its predecessor, which translates directly to lower costs for API users.
Pricing lands at $1.50 per million input tokens and $7.50 per million output tokens — a significant reduction from the $9.00 per million output tokens charged for Gemini 3.5 Flash. For high-volume applications, that’s not a trivial difference.
On the performance side, the improvements are tangible. The model’s computer use capabilities jumped from 78.4% to 83% on OSWorld-Verified benchmarks, a standard test for how well AI can navigate graphical interfaces and perform real-world desktop tasks. Coding precision also improved, with Google highlighting fewer unwanted code edits and reduced execution loops — a pain point that developers had flagged with earlier Flash models.
Perhaps most practically significant for enterprise users: the knowledge cutoff date advanced from January 2025 all the way to March 2026, meaning the model now has substantially more current training data to draw on.
Early adopters include Harvey, the AI legal research platform, and Hebbia, which builds AI for institutional knowledge work. Google also announced that CodeMender, an AI-assisted code modernization tool, entered preview with Salesforce, Robinhood, and Palo Alto Networks as launch partners.
The Cheaper and the Specialized
Alongside Gemini 3.6 Flash, Google released Gemini 3.5 Flash-Lite, billed as the most cost-effective model in the Gemini family. Priced at $0.30 per million input tokens and $2.50 per million output tokens, it targets use cases where cost efficiency matters more than top-tier capability — think document classification, summarization pipelines, or high-throughput inference tasks.
The third model, Gemini 3.5 Flash Cyber, takes a different approach entirely. Rather than a general-purpose model, it’s a fine-tuned variant specialized for identifying and patching cybersecurity vulnerabilities. It won’t be widely available — Google is limiting it to governments and “trusted partners” through a restricted pilot program, citing the sensitive nature of security-focused AI capabilities.
The Absence That Speaks Loudest
What didn’t ship is at least as significant as what did. Gemini 3.5 Pro, Google’s next flagship model, has now missed three separate target dates. It was initially expected in June following its announcement at Google I/O 2026, then pushed to July 17 after Google disclosed it was conducting a full architectural rebuild to address coding performance issues. As of this week, the model remains in partner testing, with Google product lead Logan Kilpatrick saying the company hopes to “land soon.”
The pattern is unusual for Google, which has historically shipped models on relatively tight cadences. Multiple rebuilds and missed deadlines for a flagship suggest that internal performance benchmarks — particularly around coding and reasoning tasks — are setting a higher bar than earlier generations required.
The Real Surprise: Gemini 4 Has Already Started
The most consequential sentence in Google’s announcement was also the easiest to miss: the company has “already started our most ambitious pre-training run yet, for Gemini 4.”
That disclosure, buried near the end of the announcement post, reframes the entire week’s news. Google isn’t just iterating on Gemini 3.x — it’s already working on the generation after the one it hasn’t fully shipped yet. The company said it is “excited by the progress,” without providing details on timelines, capabilities, or what makes this run more ambitious than prior ones.
Pre-training is the computationally intensive foundational phase of building a large language model, where a model learns from massive amounts of data before any specialized fine-tuning occurs. Starting a new pre-training run is a significant investment — it typically involves months of GPU time across enormous compute clusters and represents a long-term bet on where AI capabilities are headed.
The announcement signals that Google is playing a longer game: even as it irons out Gemini 3.5 Pro and ships incremental Flash improvements, the next architectural leap is already underway.
Competitive Context
The timing matters. Google’s releases come as the broader AI landscape is compressing. DeepSeek is expected to drop its V4 model on July 24, and Moonshot AI’s Kimi K3 — a large open-weight model — is set to release weights on July 27. The pace of frontier model releases has accelerated to a point where a two-week lag in a flagship launch has measurable competitive consequences.
Google’s strategy appears to be holding the Flash line competitive at the lower end while preserving the Pro slot for something genuinely differentiated. Whether Gemini 3.5 Pro, when it finally arrives, justifies the wait remains to be seen. But the early signal on Gemini 4 suggests that Google is betting its next generation is where it will reassert itself at the frontier.
For now, developers get three capable models at lower price points and better performance. Enterprises get expanded access to more current knowledge. And everyone gets a hint that the model race in 2027 is already being shaped by decisions being made right now.
*Sources: Google DeepMind blog, TechCrunch, 9to5Google, Android Authority, Gizbot*
