Foreword¶

On July 21, 2026, Google officially launched three models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. For developers running AI in production environments, this update does not focus on “being the top of the leaderboard”, but on more practical metrics in Agent workflows: whether output tokens are fewer, latency is lower, and whether the bill for the same task is cheaper.

The core data released in the official blog shows that on the Artificial Analysis Index, 3.6 Flash reduces output tokens by 17% compared to 3.5 Flash; on coding benchmarks such as DeepSWE, the reduction can reach up to 65%. The output pricing is 7.5 USD per million tokens (input pricing is 1.5 USD per million tokens), which is lower than the previous generation of Flash models. The DeepSWE score has increased from 37% to 49%. One week later, on July 28, Gemini API’s Managed Agents switched the default model to 3.6 Flash, and launched Environment Hooks. This release received 760 upvotes and 577 comments on Hacker News, indicating that the developer community has a high level of interest in “Agent cost-effective models”.

This article is based on Google’s official blog and public documentation, sorting out the capability changes of 3.6 Flash, the two sister models released in the same batch, as well as the engineering updates to Managed Agents.

3.6 Flash: Do More with Fewer Tokens¶

Google positions 3.6 Flash as the “workhorse model” of the Flash series, directly taking over the role of 3.5 Flash in coding, knowledge work, and multimodal scenarios, while emphasizing the simultaneous improvement of Token efficiency and Agent reliability.

The official directions of change include:
- Less redundant talk: On the Artificial Analysis Index, output tokens are 17% less than 3.5 Flash; the reduction is even greater in some tasks such as DeepSWE.
- Fewer steps: In multi-step workflows, the number of reasoning steps and tool calls has decreased.
- Lower unit price: Input pricing is 1.5 USD per million tokens, and output pricing is 7.5 USD per million tokens. According to official statements, this reduces the overall cost of a single Agent task.

For teams that run long-term Agent loops, sub-Agent orchestration, and batch document processing, “fewer output tokens + lower unit price” is a superimposed effect, which is more practically meaningful than simply looking at single-inference latency.

Benchmark Tests: Coding, Computer Use, and Knowledge Work¶

Google cited multiple public or in-house benchmarks in the launch article. The following data are all from the official blog for readers’ reference:

A few points worth noting:
1. DeepSWE measures the performance of software engineering Agents on real codebases. The improvement from 37% to 49%, combined with improved Token efficiency, indicates that the model has made progress in “making fewer coding mistakes and getting stuck in execution loops less often”.
2. OSWorld-Verified is directly related to Computer Use. 3.6 Flash reached 83.0% on OSWorld-Verified, and Computer Use is now available as a built-in client tool via Gemini API and Gemini Enterprise, meaning developers do not need to separately assemble a GUI operation layer to call it in Agent workflows.
3. GDPval-AA v2 is targeted at knowledge work scenarios. Customer cases from Hebbia, Harvey and others mention that 3.6 Flash performs better in multimodal document parsing, chart analysis, and report drafting.

It should be objectively stated that some developers on Hacker News cited third-party evaluations from Artificial Analysis, believing that 3.6 Flash is at an “upper-middle level” in comprehensive intelligence metrics, but the speed advantage of the Flash lineup is more obvious in the dimensions of intelligence vs. time per task and intelligence vs. output speed. If your business is high-frequency frontend iteration or rapid prototyping, these “fast and sufficient” models are often more cost-effective than the flagship Pro models.

Simultaneous Releases: 3.5 Flash-Lite and 3.5 Flash Cyber¶

This update did not only include 3.6 Flash. Google also released two other models with clear divisions of labor.

3.5 Flash-Lite: Throughput and Latency¶

3.5 Flash-Lite is targeted at high-throughput, low-latency scenarios, such as Agent search and large-scale batch document processing. Official data:
- Output speed is approximately 350 tokens per second (measured by Artificial Analysis)
- Pricing: 0.3 USD per million input tokens, 2.5 USD per million output tokens
- Terminal-Bench 2.1: 54% (compared to 31% for 3.1 Flash-Lite)
- Also supports the Computer Use built-in tool

In Managed Agents, 3.5 Flash-Lite can be used as a combination of main Agent (3.6 Flash) + sub-Agent (Flash-Lite): the main Agent is responsible for reasoning and orchestration, while Flash-Lite is responsible for high-concurrency, low-cost subtasks, such as generating 25 webpage design drafts at one time.

3.5 Flash Cyber + CodeMender: Security Specialization¶

3.5 Flash Cyber is fine-tuned based on 3.5 Flash, focusing on the discovery and repair of cybersecurity vulnerabilities. It uses CodeMender multi-Agent collaboration to generate unified reports, reaching the cutting-edge level on the CyberGym benchmark.

Due to dual-use risks, Google has adopted a restricted deployment approach: this model will be opened via CodeMender through limited pilots for governments and trusted partners, rather than a full public API. For ordinary developers, just knowing that this product line exists is sufficient; if you are working on security compliance or government and enterprise projects, you can pay attention to subsequent pilot access methods.

Managed Agents: Switching to 3.6 Flash by Default and Environment Hooks¶

On July 28, the Google DeepMind team released the Managed Agents update, which is the direct channel for 3.6 Flash to land in Agent production environments.

Default Model Switch¶

The antigravity-preview-05-2026 Agent will use Gemini 3.6 Flash by default without modifying code. You can also explicitly specify it via agent_config.model:
- gemini-3.6-flash (default): Balanced reasoning, coding, and tool calling
- gemini-3.5-flash: The previous generation general-purpose Agent model
- gemini-3.5-flash-lite: Lowest latency and cost

Example (TypeScript, @google/genai SDK):

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({});

const interaction = await client.interactions.create({
  agent: "antigravity-preview-05-2026",
  input: "Audit all dependencies in package.json, upgrade outdated packages, and verify the build by running npm test.",
  environment: "remote",
  agent_config: {
    type: "antigravity",
    model: "gemini-3.5-flash-lite",
  },
});

console.log(interaction.output_text);

Environment Hooks: Intercept and Audit Tool Calls in the Sandbox¶

Environment Hooks allow you to execute custom scripts before and after each tool call by an Agent. Place a .agents/hooks.json file in the project root directory to register handlers for pre_tool_execution or post_tool_execution events.

Typical use cases:
- Security gate: Run gate.py before code_execution or write_file, and return {"decision": "deny"} to block dangerous operations
- Auto-formatting: Run a linter after each tool execution to maintain consistent code style
- HTTP callback: POST tool call events to an external audit system

The matcher field supports regular expressions, such as code_execution|write_file or a wildcard *. Teams like Offdeal have already used the post_tool_execution Hook to perform image quality verification in a remote sandbox — previously, remote sandboxes could not run client-side verification logic, and Hooks have filled this gap.

Cost and Automation¶

Several engineering-oriented capabilities were also added during the same period:
- Free Tier: Managed Agents is now open to free-tier projects, and you can get an API key to test even without a billing project
- Budget Controls: agent_config.max_total_tokens can cap total tokens (including input, output, and thinking tokens), and will pause safely when the limit is exceeded, preserving the environment state for resumption
- Scheduled Triggers: Use Cron to trigger Agent tasks on a schedule, and files in the same sandbox can be retained across executions
- Environments API: List, inspect, and delete sandbox sessions via code, avoiding waiting for the 7-day TTL automatic expiration

For concerns about “Agents running out of control and bills spiraling”, Budget Controls is the most direct targeted feature.

Security and Availability Channels¶

3.6 Flash, along with its model card, notes enhanced Frontier Safety coverage for CBRN (chemical, biological, radiological, nuclear) and cyberattack abuse scenarios, with the goal of improving jailbreak resistance while reducing false rejections of legitimate requests.

Availability channels (official listed entry points):
- Developers: Gemini API (Google AI Studio, Android Studio), Google Antigravity
- Enterprises: Gemini Enterprise Agent Platform, Gemini Enterprise applications
- Consumers: Gemini App (3.5 Flash-Lite is also gradually rolling out on Google Search)

Summary: The Flash Lineup Bets on Agent Scaling¶

The launch logic of Gemini 3.6 Flash is very clear: in the context of Agents becoming the default interaction form, Google has used the three-pronged approach of Token efficiency, price cuts, and Managed Agents switching to default to push the Flash series to the position of “the default foundation for production Agents”. 3.5 Flash-Lite is responsible for throughput, 3.5 Flash Cyber is responsible for security vertical scenarios, and the Hooks and budget controls of Managed Agents have completed the final engineering landing link.

If you are already using Gemini API’s Managed Agents or Antigravity, your next interaction will automatically benefit from 3.6 Flash; if you are building your own Agent orchestration, it is recommended to compare the number of output tokens and number of task completion steps of the same prompt under 3.5 Flash and 3.6 Flash, which is closer to real billing than just looking at single benchmark scores.

Reference sources:
- Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
- Gemini API Managed Agents: 3.6 Flash, hooks, and more
- Hacker News Discussion