Google Ships Three Gemini Flash Models — but Still No 3.5 Pro

Google shipped three new Gemini models on July 21, and the most revealing thing about the launch is the model that wasn't in it.

Instead of the long-delayed flagship Gemini 3.5 Pro, the company rolled out a trio aimed squarely at the cheap, fast, high-volume end of the market: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a security-tuned variant called Gemini 3.5 Flash Cyber that is restricted to governments and trusted partners. All three land in the tier developers actually run at scale. None of them is the top-end model Google promised — and repeatedly failed to deliver — over the past two months.

The headliner, Gemini 3.6 Flash, is a genuine upgrade dressed as a routine point release. It is cheaper than the model it replaces, priced at $1.50 per million input tokens and $7.50 per million output tokens, down from $9.00 on the prior Flash. It is also more frugal per answer: on the Artificial Analysis Index it consumes roughly 17% fewer output tokens than Gemini 3.5 Flash while matching its intelligence score. Because output tokens are where inference bills add up, that combination compounds. As one industry write-up put it, "the real cost drop is bigger than the sticker."

The capability gains are concentrated in the two areas Google has been losing on. Its DeepSWE coding score climbs from 37% to 49%, and its OSWorld-Verified computer-use score rises from 78.4% to 83.0% — the latter, notably, the best mark in Artificial Analysis's cross-model comparison, ahead of GPT-5.6 and Grok 4.5. The knowledge cutoff advances more than a year, from January 2025 to March 2026. The model keeps a 1-million-token input context window with 64K maximum output, and it shipped inside GitHub Copilot at launch, giving it immediate distribution to millions of developers.

Speed is the other story. Artificial Analysis clocked Gemini 3.6 Flash at an average of 1.3 minutes per task, down from 2.7 minutes for the previous Flash — more than a 50% reduction — running at roughly 304 output tokens per second in pre-launch testing. "Maintains the same intelligence as Gemini 3.5 Flash" while halving time per task, the benchmarking firm noted, framing the release as an efficiency play rather than a frontier leap.

Gemini 3.5 Flash-Lite slots in below the flagship Flash as an even cheaper, lower-latency option for the highest-volume, least demanding work — classification, routing, extraction and the kind of agentic sub-tasks that fire thousands of times per workflow. Gemini 3.5 Flash Cyber is the odd one out: a security-hardened build gated to governments and vetted partners, and Google's answer to a growing enterprise appetite for models tuned to defensive and threat-analysis use cases rather than general chat.

Why it matters

The absence at the center of this launch is Gemini 3.5 Pro. Google previewed the flagship during I/O in the spring, but it has now slipped past multiple internal deadlines. Reporting through mid-July described a model that missed coding targets, hallucinated too often to clear reliability bars, and trailed OpenAI's GPT-5.6 in head-to-head benchmarks even after Google reset its training data in late June. Bloomberg reported the launch was delayed because the technology fell short of internal goals; the uncertainty helped knock several percent off Alphabet's share price, and prediction markets had by launch day pushed their best guess for a Pro release into August. A Google spokesperson would only confirm that Gemini 3.5 Pro "and other AI systems are currently being tested with partners."

So Google did the pragmatic thing: it shipped the tier it could win. And that tier is increasingly where the real fight is. Frontier models generate the headlines, but Flash-class models generate the invoices — they are what powers coding agents, customer-support bots, document pipelines and the sprawling agentic systems enterprises are wiring together. In that market, capability-per-dollar and long-context reliability matter more than topping a leaderboard. One analysis captured Google's logic bluntly: the company is "compressing pricing where it captures volume and holding capability where it captures prestige." Cutting output pricing 17% while improving coding and computer-use is a move designed to lock in developers who run millions of tokens a day and count every one.

The risk is the mirror image. Competing on cost is a comfortable place to be only if you also own the ceiling. OpenAI's GPT-5.6 still leads on DeepSWE and Terminal-bench, Grok 4.5 edges ahead on SWE-Bench Pro, and Anthropic's Claude Sonnet 5 tops other agentic benchmarks — while a wave of senior DeepMind researchers has departed, several for Anthropic. A cheaper, faster Flash is a strong hand. But a company that keeps missing on its flagship risks conceding the "best model available" conversation entirely, and with it the prestige that pulls developers toward a platform in the first place.

What to watch

Two things. First, Gemini 4: Google confirmed it is already in pre-training and teased it alongside this launch, which raises the question of whether 3.5 Pro ever ships as a standalone flagship or gets quietly folded into the next generation. If Pro slips much further, the story stops being "delayed" and starts being "skipped."

Second, the open-weight squeeze from below. Flash's whole value proposition is cheap, capable, high-volume inference — precisely the ground that open models like DeepSeek and the anticipated Kimi K3 are marching onto, often at zero marginal license cost. Google's Flash tier has to stay meaningfully better than free to justify its price. The next few releases from that camp will tell us whether Google's two-speed strategy is a durable structural advantage or a temporary lead on borrowed time.