Gemini 3.5 Features: Flash, Flash-Lite, Computer Use, and Pro Status

Gemini 3.5 is no longer a rumored model. Google officially launched Gemini 3.5 Flash on May 19, 2026, added native Computer Use in June, and introduced Gemini 3.5 Flash-Lite and the defensive Gemini 3.5 Flash Cyber in July.

That does not mean Gemini 3.5 Pro is available. Google’s July 21 announcement still described 3.5 Pro as being tested with partners, with broad availability planned only when it is ready. Early claims about a Cappuccino codename, anonymous benchmarks, or a fixed launch date should not be treated as product facts.

Current status

Gemini 3.6 Flash became the newer Flash workhorse in July 2026. New projects should compare it with 3.5 Flash instead of assuming that every 3.5 model is still the default choice.

Feature 1: Agentic coding and long-horizon work

Gemini 3.5 Flash was designed to complete multi-step work more reliably, not merely improve chat scores. Google emphasizes coding, tool use, task decomposition, and workflows that continue across longer time spans.

Useful patterns include inspecting a repository before editing and testing it; combining search, code execution, and custom functions; analyzing multiple documents while preserving context; and delegating extraction or validation to subagents.

Agent capability is not permission control. Applications still need tool allowlists, confirmations, and reversible boundaries for file writes, messages, payments, deletions, and production operations.

Feature 2: Multimodal input as part of a workflow

Gemini 3.5 Flash continues Google’s native multimodal approach. Text, images, video, audio, and documents can participate in the same task, while official demos emphasize richer graphics and interactive interfaces produced from complex visual material.

A practical workflow can read PDFs and chart screenshots, extract fields, detect inconsistent data, run calculations through code execution, and return structured output or an interactive presentation. Production teams should still measure OCR, chart reasoning, long-video retrieval, and schema accuracy on their own files.

Feature 3: Computer Use in the main model

On June 24, Google introduced Computer Use as a built-in tool for Gemini 3.5 Flash. The model can interpret screenshots and propose mouse, keyboard, scrolling, and navigation actions across browser, mobile, and desktop environments.

The normal loop is: send a screenshot and goal; receive a proposed action; execute it in an isolated client; return the new screenshot and result; pause for human confirmation before sensitive operations.

The model proposes actions, but the developer-controlled client executes them. Restrict domains, files, and accounts; isolate test sessions; and defend against prompt injection embedded in web pages. Login, purchase, publish, delete, and permission changes should not be approved automatically.

Feature 4: Flash-Lite for low-cost scale

The stable model ID is gemini-3.5-flash-lite. Its model page lists a 1,048,576-token input limit, a 65,536-token output limit, and text, image, video, audio, and PDF input with text output.

Flash-Lite is intended for high-volume parsing, extraction, classification, routing, structured JSON, and lightweight subagent tasks. Official July 2026 pricing is $0.30 per million input tokens and $2.50 per million output tokens. The default thinking_level is minimal; multi-step tool work can test medium or high, with new latency and cost measurements.

Google’s current Computer Use documentation lists Flash-Lite as supported, although capability tables and regional rollout may update at different times. Verify the tool in the actual project and region before production use.

Feature 5: Flash Cyber is a defensive specialization

Gemini 3.5 Flash Cyber is a lightweight security model built on 3.5 Flash to help defenders find, validate, and patch vulnerabilities. Google presented it together with the CodeMender security agent, not as a replacement for a general chat or coding model.

Do not assume it is a universal Gemini API model ID. Vulnerability validation must also remain within authorized repositories, test systems, and asset boundaries.

Gemini Spark is an app agent, not a Pro codename

Google announced Gemini Spark as a persistent personal-task agent in the Gemini app. With user authorization, it can work on research, planning, and other longer-running activities.

Spark is a product-layer feature. Model selection should separately identify the underlying Gemini model and whether users interact through the Gemini app, AI Mode, Antigravity, the Gemini API, or an enterprise agent platform.

Gemini 3.5 Pro is still not generally available

Google said it was developing 3.5 Pro when Flash launched in May. The July 21 update still described partner testing and promised broad availability when ready. The Google DeepMind model page also continues to show “3.5 Pro coming soon.”

The confirmed facts are that 3.5 Pro is in development, but it has no broadly available stable model ID or GA specification. A guide should not present Cappuccino, leaked benchmark numbers, or a fixed launch date as current product information.

Choosing a Gemini model now

Record model ID, API surface, thinking level, tools, input length, output tokens, success rate, P95 latency, and cost per completed task. One benchmark score is not a deployment decision.

Developer migration and API notes

1
2
3
4
5
6
7
8

from google import genai

client = genai.Client()
response = client.models.generate_content(
    model="gemini-3.5-flash-lite",
    contents="Extract the invoice fields and return valid JSON.",
)
print(response.text)

Set thinking_level by workload; remove deprecated temperature, top_p, and top_k configuration for the latest models; do not end a conversation with a prefilled model turn; test Function Calling, Computer Use, and code execution separately; and pin an explicit stable model ID instead of following a latest alias without review.

Before migration, preserve outputs from a fixed evaluation set. Compare field accuracy, tool-call count, output length, latency, and total cost. An HTTP 200 response does not prove that a migration succeeded.

Operational boundaries

Long-running tasks can drift, Computer Use can select the wrong control, and multimodal parsing can miss details. Production use needs human approval for external writes, tool logs, isolated accounts, prompt-injection and authorization checks, a fixed regression suite, and a rollback plan for model upgrades.

The meaningful change in Gemini 3.5 is the Flash family’s move toward agent execution, multimodal workflows, and UI operation—not a still-unreleased Pro codename.

Official sources

• Gemini 3.5 launch
• Gemini 3.5 Flash model card
• Computer Use in Gemini 3.5 Flash
• Gemini API Computer Use documentation
• Gemini 3.6 Flash, 3.5 Flash-Lite, and Flash Cyber
• Gemini 3.5 Flash-Lite model page
• Latest Gemini model migration guide
• Gemini app and Spark update