Google announced three new Gemini models on July 21, 2026: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The headline number is a 17% reduction in output token usage over 3.5 Flash, paired with lower API pricing and meaningful benchmark gains across coding, knowledge work, and computer use. The company also confirmed pre-training has begun on Gemini 4.

Why This Release Matters

Google's Flash series has become the backbone of production AI deployments. While larger flagship models grab headlines, it's the efficient, cost-effective models that actually power most enterprise agent workflows. With Gemini 3.6 Flash, Google is doubling down on that bet — making the workhorse faster, cheaper, and more capable in the same release cycle.

The timing is notable too. Gemini 3.5 Flash launched at Google I/O in May and is already being deprecated in favor of this update, signaling an accelerating release cadence.

17% Fewer output tokens vs. 3.5 Flash
49% DeepSWE coding score (up from 37%)
350 tok/s Output speed of 3.5 Flash-Lite
$7.50 Output price per 1M tokens (3.6 Flash)

Gemini 3.6 Flash: The New Workhorse

Gemini 3.6 Flash is positioned as the primary model for developers building agentic systems. Built directly on 3.5 Flash, it incorporates feedback about code generation reliability and overall efficiency.

Key performance improvements over 3.5 Flash:

  • Coding: DeepSWE benchmark jumps from 37% to 49%, with higher precision, fewer unwanted edits, and fewer execution loops
  • ML Research: MLE Bench improves from 49.7% to 63.9%
  • Computer Use: OSWorld-Verified rises from 78.4% to 83.0% — and computer use is now a built-in API tool, not a preview feature
  • Knowledge work: GDPval-AA v2 increases from 1349 to 1421
  • Knowledge cutoff: Updated from January 2025 to March 2026

The efficiency story is just as compelling as the performance story. On the Artificial Analysis Index, 3.6 Flash uses 17% fewer output tokens than 3.5 Flash. On some agentic benchmarks like DeepSWE, token reduction reaches up to 65%. This matters enormously for developers running high-volume agent workflows where token costs compound rapidly.

Pricing: $1.50/1M input tokens, $7.50/1M output tokens — down from $9.00/1M output on 3.5 Flash, roughly a 17% price cut on the output side.

Developer Tip
For agentic workflows, the combination of 17% fewer tokens + ~17% lower output price means the effective cost per task could fall by 30% or more when both factors compound. Benchmark on your specific workload before assuming the savings — but the direction is clear.

Gemini 3.5 Flash-Lite: Speed for Scale

Gemini 3.5 Flash-Lite targets the highest-throughput, lowest-latency use cases: agentic search, document processing, and AI Overviews in Google Search.

At 350 output tokens per second (per Artificial Analysis), it's the fastest model in the 3.5 series by a significant margin. Pricing is $0.30/1M input, $2.50/1M output — very competitive for high-volume production traffic.

What makes Flash-Lite stand out isn't just speed: it actually outperforms the previous-generation 3 Flash on several agentic benchmarks:

  • SWE-Bench Pro: 54.2% vs. 49.6%
  • OSWorld-Verified: 74.0% vs. 65.1%
  • Terminal-Bench 2.1: 54% vs. 31% (vs. prior Flash-Lite)

The model supports configurable thinking levels, letting developers dial between minimal (fastest, cheapest) and higher (for complex multi-step subagent workloads) depending on the task.

Gemini 3.5 Flash Cyber: Google Enters the Security LLM Race

Gemini 3.5 Flash Cyber is Google's first model purpose-built for cybersecurity — specifically vulnerability detection, validation, and patching. It's integrated with CodeMender, Google's code security agent platform, where multiple Flash Cyber agents collaborate to produce a unified security report.

Google claims competitive frontier-level performance on the CyberGym benchmark. Given the dual-use risks inherent to a model trained on offensive security knowledge, access is initially restricted to governments and trusted partners through a limited-access pilot.

This positions Google directly alongside Anthropic's Claude Mythos (which has attracted significant government attention for cyber capabilities) and OpenAI's GPT-5.6 Sol in the cybersecurity AI space — a segment that's moved from theoretical to urgently practical following this week's OpenAI/Hugging Face incident.

What's Coming Next
Google confirmed that Gemini 3.5 Pro — originally slated for June — is still in testing with unnamed partners and will ship "as soon as it's ready." More significantly, Google announced it has started pre-training for Gemini 4, describing it as "the most ambitious pre-training run yet." No timeline was provided.

Model Comparison at a Glance

Model Input Price Output Price Speed Best For
Gemini 3.6 Flash $1.50/1M $7.50/1M Standard Coding, agents, knowledge work
Gemini 3.5 Flash-Lite $0.30/1M $2.50/1M 350 tok/s High-throughput, low-latency
Gemini 3.5 Flash Cyber TBD TBD — Security vulnerability work

Key Takeaways

  • Gemini 3.6 Flash cuts output token usage 17% and drops output pricing from $9 to $7.50/1M
  • Coding benchmark DeepSWE jumps from 37% to 49%; computer use is now a standard API tool
  • 3.5 Flash-Lite runs at 350 tok/s and beats older 3 Flash on key agentic benchmarks
  • 3.5 Flash Cyber is Google's first security-specialized LLM — restricted to vetted partners for now
  • Gemini 4 pre-training is underway; Gemini 3.5 Pro still in partner testing

All three models — 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — are available now through the Gemini API, Google AI Studio, and Android Studio. General users can access the new models in the Gemini app, with Flash-Lite also rolling out in Google Search.

Related Reading · Official Sources
· Google Blog — Gemini
· Gemini API Docs (Google AI for Developers)
· Google DeepMind — Gemini models