Google Launches Gemini 3.6 Flash and Two New AI Models

Google has expanded its Gemini lineup with three efficiency-focused AI models, although the company’s most anticipated release, Gemini 3.5 Pro, remains unavailable several weeks after its expected launch.

Announced on July 21, the new lineup includes Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and the cybersecurity-focused Gemini 3.5 Flash Cyber. The first two are available through Google’s consumer, developer and enterprise platforms, while Flash Cyber will initially be restricted to governments and trusted security partners.

The release at a glance

The pricing and performance figures come from Google’s launch announcement and Gemini API documentation. Both generally available models support input contexts of up to 1,048,576 tokens and maximum text outputs of 65,536 tokens.

Gemini 3.6 Flash targets the cost of running AI agents

Gemini 3.6 Flash is positioned as the main workhorse of the updated lineup. Rather than chasing the highest possible benchmark score, it is designed to reduce the number of tokens, tool calls and reasoning steps required to complete complex tasks.

Google says the model consumes 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. On some software-engineering workloads, the reduction reportedly reaches 65%. Lower token use matters because businesses pay not only for the final response but also for the intermediate reasoning and repeated model calls generated by autonomous agents.

The model also produced higher results than 3.5 Flash on several Google-reported evaluations. Its DeepSWE coding score increased from 37% to 49%, while its MLE Bench result rose from 49.7% to 63.9%. On OSWorld-Verified, which measures an AI system’s ability to complete computer-based tasks, the score moved from 78.4% to 83%.

These gains suggest that 3.6 Flash is intended for practical workflows such as software migration, document analysis, financial research and multi-step automation rather than simple chatbot conversations.

Flash-Lite is built for volume rather than maximum intelligence

Gemini 3.5 Flash-Lite addresses a different part of the market. At $0.30 per million input tokens and $2.50 per million output tokens, it costs one-fifth as much as 3.6 Flash for input and one-third as much for generated output.

That price difference becomes significant when an application processes millions of receipts, product records, search results or customer queries every day.

Google says Flash-Lite reaches approximately 350 output tokens per second and is the fastest model in the Gemini 3.5 family. It is aimed at document parsing, agentic search, translation, classification and subagents that complete smaller tasks under the direction of a more capable model.

Despite its lower price, the model recorded notable improvements over Gemini 3.1 Flash-Lite. Its Terminal-Bench 2.1 score increased from 31% to 54%, while its long-context result on GDM-MRCR v2 rose from 60.1% to 72.2%. It also outperformed the older Gemini 3 Flash on SWE-Bench Pro and OSWorld-Verified, according to Google’s evaluations.

Flash Cyber will not receive a normal public release

The third model, Gemini 3.5 Flash Cyber, is a specialised version of 3.5 Flash trained to discover, confirm and repair software vulnerabilities.

It will operate through CodeMender, Google DeepMind’s security agent. The system can call the model repeatedly, allowing different agents to explore additional code paths before combining their findings into one report.

During testing on Google’s V8 JavaScript engine, Flash Cyber identified 55 confirmed issues. Gemini 3.5 Flash found 47, while Anthropic’s Opus 4.6 identified 36 under the reported comparison. Ten of the issues discovered by Flash Cyber were not found by any of the other evaluated models.

However, the model’s ability to locate exploitable weaknesses also creates misuse risks. Google will therefore limit early access to governments and selected partners instead of making it openly available through the Gemini API.

This distinction is important: Google introduced three models, but only 3.6 Flash and 3.5 Flash-Lite are broadly available at launch. Flash Cyber is scheduled to enter a restricted pilot.

Gemini 3.5 Pro has missed its original launch window

The absence of Gemini 3.5 Pro is more notable because Google previously gave a relatively clear timeline.

When Gemini 3.5 Flash was announced at Google I/O on May 19, the company said 3.5 Pro was already being used internally and would roll out “next month.” That indicated a June release, but the model was still unavailable by July 21.

Google now says the model is being tested with partners and will be released broadly when it is ready. It has not provided another date. The company has also confirmed that its most ambitious pre-training run so far has begun for Gemini 4.

Reuters reported that the Pro model had fallen short of internal targets, particularly for coding tasks, although Google has not publicly confirmed the specific reason for the delay. Coding has become a commercially important AI category because businesses increasingly use models to write, inspect, migrate and secure software.

The real competition is shifting toward cost per completed task

The release shows that the AI market is no longer focused only on which company owns the highest-scoring flagship model. Developers must also consider speed, token consumption and the total cost of completing a workflow.

A model that produces a slightly weaker individual response may still be more valuable if it can process hundreds of thousands of tasks quickly and at a fraction of the cost. This is especially relevant for AI agents, which may make several model calls before completing one user request.

The strategy also gives Google a way to strengthen its commercial Gemini offering while 3.5 Pro remains unfinished. Instead of waiting for the flagship release, the company can compete for high-volume enterprise workloads through cheaper models that already integrate with the Gemini API, Google AI Studio, Android Studio, Gemini Enterprise and the Gemini app.

The Pro delay still matters for Google

The lighter models may help businesses reduce operating costs, but they do not completely replace the need for a frontier-level Pro system. Complex software development, advanced scientific analysis and long-running autonomous tasks often depend on the strongest reasoning model available.

The delay is also receiving attention because Alphabet continues to commit heavily to AI infrastructure. The company raised its expected 2026 capital expenditure to between $180 billion and $190 billion. Analysts surveyed by LSEG expected Alphabet’s second-quarter revenue to increase 21.3% to approximately $116.93 billion, with cloud revenue projected to grow by about 64%.

For now, Google has delivered faster and cheaper Gemini options, including a specialised security model with promising early results. The larger test will come when Gemini 3.5 Pro arrives and shows whether the company can match its improvements in efficiency with a comparable advance in top-tier reasoning and coding performance.