# Gemini 3.5 Flash Lite

Gemini 3.5 Flash Lite brings agentic capability to the Flash-Lite tier, which makes it a good fit for subagents that handle scoped parts of a larger task.

- **Model ID:** `google/gemini-3.5-flash-lite`
- **Type:** chat
- **Providers:** google, vertex
- **Context window:** 1,000,000
- **Maximum output tokens:** 65,000
- **Pricing:** $0.3/1M input tokens, $2.50/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/gemini-3.5-flash-lite

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'google/gemini-3.5-flash-lite',
  prompt: 'Why is the sky blue?'
})
```

## About

Gemini 3.5 Flash Lite is the efficiency tier of the Gemini 3.5 family, released on July 21, 2026. Google built it for scaling agentic systems rather than for peak single-turn quality. Across thinking levels, Gemini 3.5 Flash Lite outperforms Gemini 3.1 Flash-Lite, and the practical effect is that a Flash-Lite model can now sit inside an agent graph instead of serving only one-shot classification.

The coding and agentic gains are wide. Gemini 3.5 Flash Lite scores 54% on Terminal-Bench 2.1 against 31% for Gemini 3.1 Flash-Lite, 72.2% on GDM-MRCR v2 for long context against 60.1%, and 1140 on GDPval-AA v2 for real-world task execution against 642. On several agentic and coding evaluations it also passes Gemini 3 Flash, including 54.2% on SWE-Bench Pro against 49.6% and 74.0% on OSWorld-Verified against 65.1%. Teams running workloads on the 2.5 or 3 Flash tiers have a cheaper option that scores higher on those tasks.

Subagent execution is the pattern Google highlights. A master agent on a heavier model, such as `google/gemini-3.6-flash`, decomposes a task and hands scoped pieces to Gemini 3.5 Flash Lite: read this document set, extract these fields, translate this batch, parse this JSON into a schema. Each step stays cheap and fast, so you can fan out widely without aggregate token cost dictating the architecture. Computer use ships as a built-in tool, so a scoped step can include driving a browser rather than only reading text.

Gemini 3.5 Flash Lite accepts multimodal input within a context window of 1M tokens and returns text, up to 65K tokens. Check the Specs table on this page for the current modality list. Running Gemini 3.5 Flash Lite through AI Gateway adds usage and cost tracking, automatic retries, and provider failover on one API surface, which is what keeps a wide fan-out of subagent calls observable.

## What to consider

Gemini 3.5 Flash Lite defaults to the `minimal` thinking level. Keep it there for high-volume extraction, classification, and translation, where latency and per-token cost decide whether the workload is viable at all. Move to `low` or a higher level when a subagent has to plan across several steps and stopping early would cost you a retry. Thinking tokens count toward output tokens, so the level you set changes what each request costs. Measure total spend under the thinking levels you actually plan to run.

## When to use

### Best For

- **Scoped subagent execution** where a heavier master agent hands off bounded extraction, translation, or parsing steps
- **High-volume document processing** where per-token cost decides how many documents you can afford to process
- **Agentic search pipelines** running query planning and result triage across many parallel calls
- **Structured JSON parsing** that needs schema-conformant output at the cheapest tier in the Gemini 3.5 family
- **Long-context extraction** where GDM-MRCR v2 results improved over Gemini 3.1 Flash-Lite

### Consider Alternatives When

- A master agent has to plan and write hard code. `google/gemini-3.6-flash` handles those steps in a multi-agent setup
- The problem needs sustained multi-step analysis. `google/gemini-3.1-pro-preview` carries more reasoning depth
- You need image output. `google/gemini-3.1-flash-lite-image` generates and edits images at a comparable tier
- The task is semantic retrieval or clustering. A dedicated embedding model like `google/gemini-embedding-2` fits better

## Best for

- **Scoped Subagent Execution:** A heavier master agent hands off bounded extraction, translation, or parsing steps
- **High-Volume Document Processing:** Per-token cost decides how many documents a pipeline can afford to process
- **Agentic Search Pipelines:** Query planning and result triage run across many parallel calls
- **Structured JSON Parsing:** Schema-conformant output arrives from the cheapest tier in the Gemini 3.5 family
- **Long-Context Extraction:** GDM-MRCR v2 results improved substantially over Gemini 3.1 Flash-Lite

## Consider alternatives

- **Master Agent Reasoning:** `google/gemini-3.6-flash` handles the planning and hard coding steps in a multi-agent setup
- **Sustained Multi-Step Analysis:** `google/gemini-3.1-pro-preview` carries more reasoning depth for difficult problems
- **Native Image Output:** `google/gemini-3.1-flash-lite-image` generates and edits images at a comparable tier
- **Semantic Retrieval Workloads:** A dedicated embedding model like `google/gemini-embedding-2` fits search and clustering better

## Frequently asked questions

### What changed in Gemini 3.5 Flash Lite compared to Gemini 3.1 Flash-Lite?

Gemini 3.5 Flash Lite outperforms Gemini 3.1 Flash-Lite across thinking levels, most visibly on agentic work. It scores 54% on Terminal-Bench 2.1 against 31%, 72.2% on GDM-MRCR v2 against 60.1%, and 1140 on GDPval-AA v2 against 642. Computer use also ships as a built-in tool.

### Why is Gemini 3.5 Flash Lite described as a subagent model?

The upgraded agentic capability lets it handle a scoped part of a larger task reliably. A master agent on a heavier model decomposes the work, then dispatches bounded steps such as field extraction, batch translation, or JSON parsing to Gemini 3.5 Flash Lite, where per-step cost stays low enough to fan out widely.

### How do thinking levels affect cost on Gemini 3.5 Flash Lite?

Gemini 3.5 Flash Lite defaults to `minimal`. Higher levels add reasoning compute that counts toward output tokens, so each step up raises per-request cost. Use `minimal` and `low` for high-volume, latency-sensitive execution, and higher levels for multi-step subagent workloads.

### Does Gemini 3.5 Flash Lite beat Gemini 3 Flash on any tasks?

Yes, on several agentic and coding evaluations. Gemini 3.5 Flash Lite scores 54.2% on SWE-Bench Pro against 49.6% for Gemini 3 Flash, and 74.0% on OSWorld-Verified against 65.1%. Compare on your own workload before you move production traffic between tiers.

### Does Gemini 3.5 Flash Lite support computer use?

Yes, as a built-in tool. That lets a subagent step drive a browser or an application surface instead of only processing text, which is what makes scoped delegation practical at this tier.

### Which inputs does Gemini 3.5 Flash Lite accept?

Gemini 3.5 Flash Lite accepts multimodal input within a context window of 1M tokens and returns text, up to 65K tokens. See the Specs table on this page for the current modality list.

### How much does Gemini 3.5 Flash Lite cost on AI Gateway?

Current rates are $0.3 per million input tokens and $2.50 per million output tokens, listed on this page and subject to provider updates. AI Gateway reflects provider pricing with no markup and charges no platform fee on inference.

### How do I call Gemini 3.5 Flash Lite on AI Gateway?

Set the model to `google/gemini-3.5-flash-lite` using the AI SDK, the OpenAI-compatible Chat Completions endpoint, the Responses API, or another supported interface. AI Gateway handles provider routing, retries, and failover automatically.

### When should I use `google/gemini-3.6-flash` instead?

Use `google/gemini-3.6-flash` for the planning layer, hard coding tasks, and multimodal knowledge work. Use Gemini 3.5 Flash Lite for the high-volume steps underneath it, where the number of calls rather than the difficulty of any one call drives your bill.

### How does Zero Data Retention work with Gemini 3.5 Flash Lite through AI Gateway?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/gemini-3.5-flash-lite)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://ai.google.dev/gemini-api/docs/models)