Google Dropped Three Gemini Models at Once. Here's What Actually Changed.

No staggered rollout. No drip feed of announcements spread across a week. On July 21, Google's Gemini team lead Tulsee Doshi published a single post on The Keyword and pushed all three live: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber.

Each one does a different job. None of them is Gemini 3.5 Pro — which still hasn't shipped publicly. That's worth keeping in mind when reading the launch headlines.

Gemini 3.6 Flash: The Same Price Bracket, But Leaner

The headline number here is 17%. According to the Artificial Analysis Index, 3.6 Flash burns 17% fewer output tokens than 3.5 Flash on the same tasks. On DeepSWE — a coding benchmark by Datacurve — that figure climbs to 65%. For developers running high-volume pipelines, that's not a marketing stat. Fewer tokens per call means a smaller bill at the end of the month, which compounds fast at production scale.

The price drop backs it up: output tokens now cost $7.50 per million, down from $9 per million on 3.5 Flash. Input stays at $1.50 per million. (See our detailed Gemini 3.6 Flash vs 3.5 Flash Comparison for benchmark data).

On coding specifically, 3.6 Flash scores 49% on DeepSWE, up from 37% on 3.5 Flash. Computer use — where the model controls a desktop or browser to complete tasks — moved from 78.4% to 83% on OSWorld-Verified. Google also bumped the knowledge cutoff from January 2025 to March 2026, which was long overdue. If you're building agents with Gemini, check out Google's Agentic Cinema Hackathon.

One other addition for developers: computer use is now available as a built-in tool through the Gemini API, rather than something you have to wire up yourself.

3.5 Flash-Lite: Built for Speed and Nothing Else

Flash-Lite sits below 3.6 Flash in the lineup and is currently rolling out inside Google Search to power agentic search queries. The defining spec is throughput: 350 output tokens per second, per Artificial Analysis Index figures. That's the fastest 3.5-class model Google has shipped.

The pricing reflects its position. Input runs $0.30 per million tokens; output is $2.50 per million. It's cheap enough to call frequently, which is the point — Flash-Lite is meant for workloads where you need a fast, disposable response, not a careful one.

On benchmarks, it beats its predecessor (3.1 Flash-Lite) by a wide margin on agentic tasks. Terminal-Bench 2.1 went from 31% to 54%. Long-context recall on GDM-MRCR v2 went from 60.1% to 72.2%. It also edges out the older 3 Flash on SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%), which puts a "Lite" model above what was a standard model just one generation ago.

3.5 Flash Cyber: The Odd One Out

This is the most unusual of the three. Flash Cyber is fine-tuned on 3.5 Flash, but the training focused entirely on finding, validating, and patching software vulnerabilities. It's not available to the general public — access is restricted to government agencies and a short list of trusted partners through something Google calls the CodeMender pilot.

The architecture makes sense once you understand what it's doing. Finding deep security flaws means exploring enormous execution search spaces. Running one huge model once is slow and expensive. Flash Cyber is meant to be called many times in parallel. Inside CodeMender, Google's internal code-security agent, up to five Flash Cyber instances run simultaneously, each probing different paths, then merge their findings into one report.

On Google's own Big Sleep evaluation, Flash Cyber significantly outperformed standard 3.5 Flash and 3.6 Flash. On the V8 JavaScript engine, it found 55 unique confirmed issues in a fixed number of invocations. On CyberGym, it held its own against models that are much larger and more expensive to run.

Whether this approach scales into something with broad commercial availability is an open question. For now, it doesn't exist for most developers.

What Google Didn't Ship

3.5 Pro was not in this release. Google says it's in partner testing, with broad availability coming when internal checks clear. TechCrunch noted the timing: since Google's I/O 2026 release, OpenAI has shipped GPT-5.5 and begun rolling out GPT-5.6. Anthropic launched Claude Opus 4.8 and Claude Sonnet 5. Google is moving fast on the Flash tier while its premium model sits in testing.

Also buried at the end of the announcement: Gemini 4 pre-training has started. Google called it their "most ambitious" training run yet. No timelines, no specs — just the signal that the next major version is in motion.

Where These Models Are Available

3.6 Flash and 3.5 Flash-Lite are available now through the Gemini API, Google AI Studio, Android Studio, and the Gemini Enterprise Agent Platform. 3.6 Flash is also in the Gemini Enterprise app. Flash-Lite is rolling out in Google Search. Flash Cyber remains gated.

The practical takeaway for developers: if you're running anything on 3.5 Flash today, the upgrade path to 3.6 Flash is straightforward and will cost you less per task. Flash-Lite is worth testing if your use case tolerates some quality tradeoff for raw speed. Flash Cyber, unless you're in the pilot, is not yet your problem to solve.