Google released Gemini 3.5 Flash-Lite on July 21, 2026, alongside Gemini 3.6 Flash. It is the fastest model in the 3.5 family at 350 tokens per second, priced at $0.30 per 1M input tokens and $2.50 per 1M output tokens.
Note: This is Gemini 3.5 Flash-Lite, a text model. Not to be confused with Nano Banana 2 Lite, which is Google’s image generation model. Flash-Lite handles text, images, video, audio, and PDF input, but outputs text only.
This is not a frontier model. It is a speed-and-cost-optimized model for high-volume workloads where you need fast responses at low cost. Think agentic search, document processing, receipt translation, and high-throughput data pipelines.
Here is everything about Flash-Lite: what it can do, what it cannot, and when to use it over the bigger models.
Key Specs
| Spec | Value |
|---|---|
| Release date | July 21, 2026 |
| Model family | Gemini 3.5 |
| Input modalities | Text, Image, Video, Audio, PDF |
| Output modalities | Text only |
| Context window | 1,000,000 tokens |
| Max output tokens | 65,536 |
| Output speed | 350 tokens/sec |
| Input pricing | $0.30 per 1M tokens |
| Output pricing | $2.50 per 1M tokens |
| Cached input pricing | $0.03 per 1M tokens |
| Thinking mode | Yes (configurable levels) |
| Computer use | Built-in |
At $0.30/$2.50, Flash-Lite is 5x cheaper on input and 3x cheaper on output than Gemini 3.6 Flash. It is also faster: 350 tok/s vs 304 tok/s.
How It Compares to Other Gemini Models
| Model | Input/1M | Output/1M | Speed | Best for |
|---|---|---|---|---|
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | 350 tok/s | High-volume, low-cost |
| Gemini 3.6 Flash | $1.50 | $7.50 | 304 tok/s | General purpose |
| Gemini 3.5 Flash | $1.50 | $9.00 | 289 tok/s | Previous gen |
| Gemini 3 Flash | $0.50 | $3.00 | ~200 tok/s | Older, less capable |
Flash-Lite is the cheapest and fastest option. It outperforms the older Gemini 3 Flash on agentic benchmarks while costing less on input ($0.30 vs $0.50).
Benchmarks
Google’s official numbers:
| Benchmark | 3.5 Flash-Lite | 3.1 Flash-Lite | 3 Flash |
|---|---|---|---|
| Terminal-Bench 2.1 | 54% | 31% | 58% |
| SWE-Bench Pro | 54.2% | N/A | 49.6% |
| OSWorld-Verified | 74.0% | N/A | 65.1% |
| GDM-MRCR v2 (long context) | 72.2% | 60.1% | N/A |
| GDPval-AA v2 | 1140 | 642 | N/A |
Flash-Lite beats Gemini 3 Flash on SWE-Bench Pro (54.2% vs 49.6%) and OSWorld-Verified (74.0% vs 65.1%). That is remarkable for a Lite-tier model beating the previous generation’s standard Flash model.
The Terminal-Bench 2.1 score of 54% is lower than Gemini 3 Flash (58%), but the tradeoff is worth it for the massive cost savings.
Thinking Levels
Flash-Lite supports configurable thinking levels, which let you balance speed vs quality:
- Minimal/Low: Fastest responses, best for high-volume tasks where latency matters most. Good for simple classification, extraction, and routing.
- Medium: Balanced. Use for standard coding and agent tasks.
- High: Full reasoning. Use for complex multi-step workflows and subagent orchestration.
This is a unique feature. You can tune the same model for different performance tiers depending on the task.
When to Use Flash-Lite
High-volume data processing. Extracting information from thousands of documents, receipts, or forms. The $0.30 input price makes bulk processing affordable.
Agentic search. Running many parallel search queries as part of an agent pipeline. Flash-Lite’s speed (350 tok/s) and low cost make it ideal for search-heavy workflows.
Subagent orchestration. When you have a master agent (like Gemini 3.6 Flash) delegating tasks to subagents, Flash-Lite handles the subagent workload cheaply and quickly.
Real-time applications. When you need sub-second responses for user-facing features like autocomplete, suggestions, or inline help.
Cost optimization. If you are spending too much on Gemini 3.5 or 3.6 Flash, routing simpler tasks to Flash-Lite can cut costs significantly without proportional quality loss.
When NOT to Use Flash-Lite
Complex coding. Flash-Lite scores 54% on Terminal-Bench 2.1 vs 78% for Gemini 3.6 Flash. For serious coding tasks, use the bigger model.
Architectural decisions. Flash-Lite is not the model for designing systems or making complex technical decisions. Use 3.6 Flash or Claude Sonnet 5 for that.
Long-context reasoning. While Flash-Lite supports 1M tokens, its long-context performance (72.2% on MRCR v2) is good but not leading. For critical long-context tasks, use 3.6 Flash.
API Example
import google.genai as genai
client = genai.Client(api_key="YOUR_API_KEY")
response = client.models.generate_content(
model="gemini-3.5-flash-lite",
contents="Extract the vendor name, date, and total from this receipt image.",
config={
"temperature": 0.3,
"max_output_tokens": 512,
}
)
print(response.text)With Thinking Level
response = client.models.generate_content(
model="gemini-3.5-flash-lite",
contents="Your prompt here",
config={
"thinking_config": {"thinking_budget": 2048}
}
)Pricing Comparison With Other Cheap Models
| Model | Input/1M | Output/1M | Speed |
|---|---|---|---|
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | 350 tok/s |
| Gemini 3 Flash | $0.50 | $3.00 | ~200 tok/s |
| GPT-5.6 Luna | $0.50 | $2.00 | ~100 tok/s |
| DeepSeek V4 Flash | $0.44 | $1.74 | ~180 tok/s |
| Claude Sonnet 5 | $2.00 | $10.00 | ~180 tok/s |
Flash-Lite is the cheapest Gemini model and competitive with DeepSeek V4 Flash on price. It is significantly faster than both.
Limitations
- Not a frontier model. Flash-Lite is optimized for speed and cost, not raw capability. It will not match 3.6 Flash or Claude Sonnet 5 on complex tasks.
- Text-only output. Same as all Gemini models: multimodal input, text-only output.
- Thinking mode tradeoff. Using thinking mode improves quality but increases latency and cost. For high-volume use, you may want to leave it off.
- No open weights. Flash-Lite is proprietary. You cannot self-host it.
FAQ
Is Gemini 3.5 Flash-Lite free?
Not directly. It is available through the Gemini API at $0.30/$2.50 per 1M tokens. Google AI subscriptions may include access depending on your tier.
How does Flash-Lite compare to Gemini 3 Flash?
Flash-Lite is cheaper on input ($0.30 vs $0.50), faster (350 tok/s vs ~200 tok/s), and beats it on some benchmarks (SWE-Bench Pro: 54.2% vs 49.6%, OSWorld: 74.0% vs 65.1%). It is a straight upgrade for most workloads.
Should I use Flash-Lite or 3.6 Flash?
Use Flash-Lite for high-volume, low-cost tasks where speed matters. Use 3.6 Flash for complex tasks that need more capability. Many developers use both: 3.6 Flash as the master agent, Flash-Lite as the subagent.
Does Flash-Lite support computer use?
Yes. Computer use is a built-in tool, just like in 3.6 Flash.
What is the context window?
1 million tokens, same as 3.6 Flash and 3.5 Flash.