gemini 4 isn't out yet — here's what actually shipped.

So is Gemini 4 out? No — and here's why that matters

A pre-training run is the first and most expensive phase of building a model: feeding it a massive dataset so it learns general patterns, before any of the work that turns a raw model into a product. At frontier scale, that phase alone commonly runs for months on huge compute clusters. Then comes post-training — instruction-tuning, reinforcement learning from human feedback, evaluation, red-teaming — which historically adds more months.

So when Kilpatrick said pre-training had just started, the honest translation is: Gemini 4 doesn't have a finished base model yet, let alone a tuned, tested, shippable one. A realistic window, based on how this has played out before, is late 2026 into 2027 — not weeks, not “any day now.”

The signal isn't “Gemini 4 is here.” It's “Google is confident enough in the run to say it out loud.” Useful — but not a release date.

What actually shipped: Gemini 3.6 Flash & 3.5 Flash-Lite

The usable news is the cheap, fast end of the lineup. Gemini 3.6 Flash is the new default workhorse for code, multimodal tasks and agents — and it uses 17% fewer output tokens than 3.5 Flash (per the Artificial Analysis Index), with a 1M-token context window. Gemini 3.5 Flash-Lite is the fastest and cheapest, up to 350 tokens/second, and on some coding and computer-use benchmarks it beats the older, bigger Gemini 3 Flash (SWE-Bench Pro 54.2% vs 49.6%; OSWorld 74.0% vs 65.1%).

Prices tell the story better than adjectives — here's what landed on July 21:

Where Gemini 4 sits — and when it might ship

Everything usable today — Gemini 3.1 Pro, 3.5 Flash, 3.5 Flash-Lite, and the new 3.6 Flash — is still the Gemini 3.x line. July 21 was really two separate stories on one day: an incremental 3.x update you can use now, and confirmation that the next generation has only just begun training. Keep them apart — one is a product, the other isn't yet.

A side-by-side demo made the rounds showing Gemini 3.6 Flash building a richer interactive 3D simulation than the larger Gemini 3.1 Pro — one data point, not a benchmark, but a clear sign the cheap tier keeps closing the gap on the flagship tier.

The sleeper release: Gemini 3.5 Flash Cyber

The most striking launch got the least noise. Gemini 3.5 Flash Cyber is a specialized security model (deployed through Google's CodeMender platform) built to find and fix software vulnerabilities. On Google's own test against the V8 JavaScript engine it flagged 55 confirmed issues — versus 47 for mainline 3.5 Flash and 36 for Claude Opus 4.6 — and found 10 vulnerabilities no other model caught.

One caveat worth stating plainly: that's Google's own published benchmark, and access is a limited pilot for governments and vetted partners — you can't just try it. But the direction is real: AI that hunts security bugs at scale is arriving fast. Source: deepmind.google.

What this actually changes for your ads and content

Here's the part that matters if you make content or ads, not if you collect model releases. At $0.30–$1.50 per million input tokens, generating one polished variant and generating twenty stops being a cost decision — it becomes a workflow decision. You can have a Flash-tier model draft a dozen hooks, angles or product descriptions in the time it takes to make coffee, and let real performance data pick the winner instead of your gut.

The agentic, multimodal side matters just as much: a model that's fast and cheap and can chain its own steps (draft → check against brand voice → flag the weak ones) is what makes automated first-pass production realistic. It's the same “generate many, test many, keep what performs” logic that makes AI video and image work worth doing at all — cheap enough to test a direction before you spend real budget.

The muscle to build now, while the tools are this cheap, isn't “which model is smartest.” It's “how many honest tests can I run before I spend real money.”