Google released three new Gemini AI models on July 22, 2026, that slash the cost of running AI agents for Windows developers and enterprises—but the company's anticipated flagship model, Gemini 3.5 Pro, is still in partner testing with no public launch date. The new models, headlined by Gemini 3.6 Flash, arrive with lower output token prices and notable coding improvements, even as Google confirms that training for the next-generation Gemini 4 has already begun.
What Google Actually Announced
The launch includes three distinct models, each with a specific operational role. Gemini 3.6 Flash takes over as the general-purpose workhorse, replacing Gemini 3.5 Flash for coding, multimodal analysis, and complex agent workflows. Gemini 3.5 Flash-Lite targets high-throughput, low-cost document processing and structured extraction. Gemini 3.5 Flash Cyber is a security-focused model restricted to a limited-access pilot for vulnerability detection and patching.
Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens—the input price matches the previous Flash generation, but the output price is lower. In agent-driven scenarios where models generate multiple rounds of plans, code diffs, and reports, that reduction adds up. Google also says the model uses 17% fewer output tokens on an independent evaluation index and requires fewer reasoning steps and tool calls to complete multi-step tasks.
Coding is where Gemini 3.6 Flash makes its most practical case. Google reports a jump from 37% to 49% on the DeepSWE benchmark and a leap from 49.7% to 63.9% on MLE Bench. For Windows developers, that translates into a model more likely to inspect project files, check dependencies, and reason about code changes before editing—a deliberate style that suits large .NET refactors or PowerShell automation overhauls better than quick front-end fixes.
Multimodal support is baked in from the start. Gemini 3.6 Flash handles text, images, video, audio, and PDFs with a 1 million-token context window and up to 64,000 output tokens. That matters in Windows-centric IT shops where operational knowledge is scattered across PDF incident reports, Excel exports, screenshots, and application logs rather than clean text databases.
Gemini 3.5 Flash-Lite is the economy option, priced at $0.30 per million input tokens and $2.50 per million output tokens with throughput of 350 output tokens per second. It is not a low-quality fallback but a rapid executor intended for subagent work: parsing thousands of forms, tagging metadata, routing support tickets, or generating structured JSON. Developers can tune its “thinking level” to match complexity, and benchmarks show it beating the first Flash-Lite generation by wide margins—74% versus 65% on OSWorld-Verified, for instance.
The security model, Gemini 3.5 Flash Cyber, is fine-tuned for vulnerability discovery and remediation through Google’s CodeMender agent system. It uses multiple agents to validate exploitable weaknesses and propose safe patches. However, access is restricted to governments and trusted partners—a wise precaution given the dual-use nature of such a tool. For most Windows developers, Flash Cyber is not yet available.
What These Models Mean for You
For Windows developers, Gemini 3.6 Flash should become the default model to evaluate for complex coding tasks. It is relevant for:
- Updating .NET applications to newer frameworks
- Refactoring PowerShell and TypeScript projects
- Generating test suites and analyzing crash dumps
- Building AI-powered dashboards that connect to Azure services
- Converting internal documentation into deployment scripts
Gemini 3.5 Flash-Lite is better for throughput-heavy apps where cost dominates. If you need to process 50,000 invoices or extract structured data from millions of log entries, Flash-Lite’s economics make it a building block for the labor-intensive part of an agent pipeline.
The documentation note about computer-use support is shaky. The main Gemini docs list computer-use as a tool, but the Flash-Lite model card currently says unsupported. Before you design UI automation on Flash-Lite, verify its capabilities in your specific API region.
For IT administrators, the new models can inform your AI automation roadmap but shouldn’t replace disciplined scripting overnight. An AI agent might help interpret Event Viewer errors, compare registry states, or draft a remediation summary. But granting it broad administrative privileges without guardrails is reckless. At minimum, any AI-connected tool in a Windows domain should have:
- Least-privilege access
- Human approval gates for production changes
- Full audit logs of prompts, tool calls, and outputs
- Sandboxed testing before touching Group Policy or Intune configurations
- Independent review of security- or compliance-sensitive conclusions
For everyday Windows users, the immediate effect is subtle. Gemini 3.6 Flash is rolling into the Gemini web app and related Google experiences, so search summaries and productivity tools may feel snappier. The bigger impact will come as third-party Windows apps adopt these cheaper models for features like smart document tagging, email triage, or in-app coding assistants.
The Road So Far: How Gemini Got Here
Google’s Gemini lineup has been in a sprint. Gemini 3.5 Flash arrived earlier this year as the default fast model, and the February update to Gemini 3.5 Pro raised expectations for a true flagship to compete with OpenAI’s GPT-5.5 and GPT-5.6 or Anthropic’s Claude Sonnet 5. Instead, that flagship has been stuck in testing while the competition shipped multiple premium models.
The current release reframes the narrative. Rather than chasing the biggest general-purpose model, Google is optimizing for the economics of agents—the kind that must run thousands of repetitions without burning through budgets. Logan Kilpatrick, product lead at Google DeepMind, confirmed on X that the 3.5 Pro testing is ongoing and that Gemini 4 pre-training has begun. That acknowledgment both signals ambition and underlines that the Pro tier is not ready.
A separate report from The Information, citing unnamed sources, claims Alphabet is designing a new server chip, codenamed “Frozen v2,” that could be six to ten times more efficient than current Google AI chips and launch around 2028. Google neither confirmed nor denied the chip rumor, saying only that it co-designs hardware and software for optimization. If accurate, such a chip would eventually run the Gemini models that Windows customers adopt, but it won’t affect the models available today.
What You Should Do Right Now
- Start evaluating Gemini 3.6 Flash through the Gemini API or the Gemini app. If you have an existing project that used 3.5 Flash, switch a test instance to compare quality, token usage, and turn counts on your real data—not just benchmarks.
- Audit your AI agent costs. Track output tokens as closely as input tokens, especially in agent loops. Multi-turn coding or research agents can generate massive output; the 17% reduction could matter more than headline price cuts.
- Map your multimodal needs. If your organization’s knowledge lives in PDFs, screenshots, or voice transcripts, design a prototype that feeds those directly into 3.6 Flash’s context window instead of relying on separate OCR or speech-to-text tools.
- Test Flash-Lite for high-volume tasks. For classification, extraction, or structured generation at scale, benchmark Flash-Lite against your current solution. Its 350 tokens/second throughput may speed up batch jobs significantly.
- Don’t wait for 3.5 Pro. Build your production systems on stable, generally available models. When Pro finally ships, treat it as a potential upgrade, not a dependency. Architecting agent workflows so that model substitution is painless will protect you from roadmap surprises.
- Establish agent governance now. Even if you aren’t deploying AI agents broadly, draft the policies you’ll need: what privileges an agent can hold, how its actions are logged, and who approves high-impact operations. The model is not the hardest part; the operational wrapping is.
Looking Ahead: The 3.5 Pro Mystery and Beyond
Google’s admission that Gemini 3.5 Pro is still in testing—and that Gemini 4 pre-training has already started—creates both excitement and uncertainty. The overlapping tracks may mean a smoother path to the next generation, or they may signal that 3.5 Pro will have a short reign once it finally arrives. For now, the pragmatic move is to focus on the tools that exist. Gemini 3.6 Flash and Flash-Lite represent genuine progress on the economics of AI agents, and for Windows developers and admins building real systems today, that’s the news that counts.