Gemini 3.1 Flash-Lite: Google's Cheapest Gemini 3 Model Just Landed

Google DeepMind dropped Gemini 3.1 Flash-Lite today, March 3, 2026. It’s the first Flash-Lite model in the Gemini 3 series, and it’s positioned squarely at developers running high-volume workloads who care about cost per token above almost everything else.

Open post on X

The model is available now in preview through the Gemini API in Google AI Studio and via Vertex AI for enterprise customers.

Pricing: $0.25 Per Million Input Tokens

The headline number is the price. Gemini 3.1 Flash-Lite costs $0.25 per million input tokens and $1.50 per million output tokens. Batch pricing cuts that in half: $0.125 input, $0.75 output.

To put that in context against Google’s own lineup:

Model	Input (per 1M tokens)	Output (per 1M tokens)
Gemini 3.1 Flash-Lite	$0.25	$1.50
Gemini 3 Flash	$0.50	$3.00
Gemini 2.5 Flash	$0.30	$2.50
Gemini 2.5 Flash-Lite	$0.10	$0.40
Gemini 3.1 Pro	$2.00	$12.00

So it’s half the price of Gemini 3 Flash on input, and half on output too. Compared to 2.5 Flash, it’s slightly cheaper on input ($0.25 vs $0.30) and significantly cheaper on output ($1.50 vs $2.50). The tradeoff: it’s more expensive than last-gen Flash-Lite models (2.5 Flash-Lite was $0.10/$0.40), but Google is betting the quality jump justifies the premium.

Against competitors in the lightweight model tier: OpenAI’s GPT-4o-mini runs $0.15/$0.60, and Anthropic’s Claude Haiku 4.5 costs $1.00/$5.00. Flash-Lite sits between them on input pricing but closer to GPT-4o-mini on output. The real question is whether the quality gap to those models justifies the price difference in either direction.

Speed: 363 Tokens Per Second

According to Artificial Analysis benchmarks, Gemini 3.1 Flash-Lite outputs at 363 tokens per second. That’s a 45% increase over Gemini 2.5 Flash’s 249 tokens per second. Google also claims a 2.5x improvement in Time to First Answer Token compared to 2.5 Flash.

For latency-sensitive applications like real-time translation, content moderation pipelines, or agent routing layers, those speed gains compound fast when you’re processing millions of requests.

Benchmarks: Punching Above Its Weight Class

Google published benchmark numbers that show Flash-Lite competing well above what you’d expect from a budget model:

Benchmark	Score	What It Measures
GPQA Diamond	86.9%	Graduate-level science reasoning
MMMU Pro	76.8%	Multimodal understanding
Video-MMMU	84.8%	Video comprehension
LiveCodeBench	72.0%	Code generation
MMMLU	88.9%	Multilingual question answering
SimpleQA	43.3%	Parametric knowledge
Arena.ai Elo	1432	Overall model ranking

The GPQA Diamond score of 86.9% is striking for a model at this price point. For reference, Gemini 3.1 Pro scores 94.3% on that same benchmark. Getting 86.9% from a model that costs one-eighth as much is the kind of ratio that makes production architects reconsider their model routing strategies.

The Arena.ai Elo of 1432 places it ahead of several larger models from prior Gemini generations, including some 2.5-series variants.

One weak spot: MRCR v2 at 1M tokens came in at just 12.3%, suggesting that while the model accepts a 1M token context window, its ability to retrieve specific information from very long contexts is limited. This is a common tradeoff in lightweight models, and it’s worth knowing before you build a system that depends on long-context retrieval.

What’s Under the Hood

Gemini 3.1 Flash-Lite is based on the Gemini 3 Pro architecture, distilled down for efficiency. It was trained on Google’s TPUs using JAX and ML Pathways.

Key specs:

Context window: 1M tokens input, 64K tokens output
Input modalities: Text, images, audio, video, PDF
Output: Text only
Thinking levels: Built-in, adjustable in AI Studio and Vertex AI

The thinking levels feature is worth flagging. Like larger Gemini models, Flash-Lite gives developers control over how much the model reasons before responding. For simple classification tasks you can turn thinking down to minimize latency. For more complex generation, like building UI components or running simulations, you can dial it up. This flexibility is unusual at this price tier.

The Gemini 3 Series Lineup Gap

There’s an odd hole in Google’s current model lineup. With today’s release, the Gemini 3 series now has:

3.1 Pro (flagship, $2.00/$12.00)
3.1 Flash Image (image generation)
3 Flash (mid-tier, $0.50/$3.00)
3.1 Flash-Lite (budget, $0.25/$1.50)

What’s missing: a Gemini 3.1 Flash text model. Google jumped from Pro to Flash-Lite in the 3.1 generation, skipping the middle. Whether that’s coming soon or Google decided the 3 Flash preview covers that tier well enough, we don’t know yet.

Gemini 3 Pro Is Being Sunset (This Week)

Buried in the same API changelog: Gemini 3 Pro shuts down March 9, 2026. That’s six days from today. The gemini-pro-latest alias switches to Gemini 3.1 Pro on March 6, and Vertex AI follows with shutdown on March 23.

Google’s reasoning: they need to consolidate compute resources. Developers have flagged this as a very short deprecation window, and some have pointed out it appears to violate Google’s own stated policy of providing at least two weeks notice.

If you’re running production workloads on Gemini 3 Pro, this is your heads up. Migrate to 3.1 Pro or evaluate whether 3.1 Flash-Lite can handle your use case at a fraction of the cost.

Who Should Care About This Model

Flash-Lite is built for specific patterns:

High-volume classification and routing. If you’re building an agent system that needs a fast, cheap model to triage requests before sending them to a more capable (and expensive) model, Flash-Lite’s speed and price make it a natural fit.

Translation at scale. The MMMLU score of 88.9% and the emphasis on multilingual capabilities suggest Google optimized for this use case specifically.

Content moderation pipelines. Fast inference, low cost, multimodal input support. Content moderation is exactly the kind of high-throughput, moderate-complexity task that Flash-Lite is designed for.

Prototyping and development. At $0.25 per million input tokens, the cost of iterating on prompts and testing agent architectures drops to near-zero. You can burn through millions of tokens during development without worrying about the bill.

Where it probably won’t work: deep reasoning tasks that require following long chains of logic, long-context retrieval over massive documents (given that 12.3% MRCR score), or any task where you need output quality comparable to Pro-tier models.

The Competitive Picture

The lightweight model tier is getting crowded. GPT-4o-mini is cheaper on raw token pricing. Claude Haiku 4.5 arguably has better reasoning for its tier. Gemini 2.5 Flash-Lite from last gen is far cheaper if you don’t need the quality bump.

What 3.1 Flash-Lite brings to the table is a combination: Gemini 3 series intelligence at a price point that makes high-volume deployment practical, with multimodal input support and adjustable thinking levels. No single number tells the whole story here. The model that wins for your workload depends on what you’re actually building.

Google is clearly betting that developers will pay a small premium over rock-bottom pricing for a meaningful quality improvement. Whether that bet pays off depends on whether the benchmark gains translate to real-world tasks at production scale.

Sources: