Overview
Google DeepMind on July 21, 2026 published a company blog post announcing general availability of three new models in the Gemini Flash family: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. The release also disclosed that pre-training has begun for Gemini 4.
Model Specifications and Pricing
Gemini 3.6 Flash is priced at $1.50 per million (M) input tokens and $7.50 per M output tokens — a reduction from the $9.00 per M output price charged for Gemini 3.5 Flash. The model retains the one-million-token context window, supports multimodal input (text, image, audio, video, and file), and carries a knowledge cutoff of March 2026, a 14-month advance over the January 2025 cutoff of its predecessor.
The entry-level offering in the batch, Gemini 3.5 Flash-Lite, is priced at $0.30 per M input tokens and $2.50 per M output tokens, positioning it as Google's lowest-cost Flash-series model for high-throughput and low-latency workloads such as agentic search and document processing.
A third variant, Gemini 3.5 Flash Cyber, is a security-tuned model fine-tuned specifically to find and patch software vulnerabilities. It is deployed inside Google's CodeMender system and is not available through a public self-serve API.
Coding and Agentic Benchmark Performance
On DeepSWE v1.1, a long-horizon software engineering evaluation, Gemini 3.6 Flash scores 49%, up from 37% for Gemini 3.5 Flash — a 12-percentage-point gain. Google also reports a 14-point improvement on MLE-Bench (machine learning engineering), rising from 49.7% to 63.9%. On SWE-Bench Pro, which tests end-to-end resolution of real GitHub issues, the model scores 58.7% against 55.1% for its predecessor. Google states the model uses 17% fewer output tokens than 3.5 Flash across the Artificial Analysis Index, with reductions of up to 65% on specific DeepSWE tasks — a compound cost saving beyond the sticker-price reduction alone.
On OSWorld-Verified, a computer-use benchmark, Gemini 3.6 Flash posts 83.0%, the highest score in Google's own comparison table, ahead of external models including GPT-5.6 Luna and Grok 4.5. External rivals retain leads in other areas: GPT-5.6 Luna and Grok 4.5 lead on SWE-Bench Pro and DeepSWE, and Claude Sonnet 5 leads on MLE-Bench and the GDPVal-AA knowledge-work index.
Gemini 3.5 Flash-Lite, despite its lower price, outperforms the older Gemini 3 Flash on SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%), according to Google's published figures.
Gemini 4 Pre-Training Disclosure
In the same blog post, authored by Tulsee Doshi, Senior Director of Product Management on the Gemini team, Google stated it has "started our most ambitious pre-training run yet, for Gemini 4." Pre-training is the foundational phase in which a large language model (LLM) is trained on large datasets to develop language, reasoning, and coding capabilities before instruction tuning and safety evaluation. Google disclosed no technical specifications, parameter counts, or timeline for Gemini 4's release. The announcement also noted that Gemini 3.5 Pro remains in partner testing and has not yet reached general availability.
Distribution and Context
Gemini 3.6 Flash and 3.5 Flash-Lite are available through the Gemini API, Google AI Studio, Android Studio, Google Antigravity, and Gemini Enterprise. Gemini is Google DeepMind's flagship model family, competing directly with Anthropic's Claude and OpenAI's ChatGPT in the foundation-model and enterprise-agent markets. Coding environments and agent platforms including Windsurf, Cursor/Anysphere, and Amazon Q Developer frequently benchmark and integrate such models as underlying engines for AI-assisted development workflows.