Google launches its cheapest Gemini model: Flash-Lite

Google's latest release, the Gemini 3.1 Flash-Lite, has one core selling point: affordability.

At just $0.25 per million input tokens and $1.50 per million output tokens, it is significantly cheaper than the Gemini 3.1 Pro, which costs $2.00 per million input tokens and $18.00 per million output tokens — roughly one-eighth the price.

This makes it the most cost-effective Gemini model Google has ever offered.

Speed and pricing changes

Flash-Lite was first introduced as a preview on March 3 and is now available on Google AI Studio and Vertex AI. It is designed for high-throughput, low-cost scenarios — the kind that process hundreds of thousands of requests daily.

Key performance metrics include:

  • Generation speed: 45% faster than Gemini 2.5 Flash
  • Time to first token: 2.5x improvement
  • Throughput: Approximately 207 tokens per second
  • GPQA Diamond (scientific reasoning benchmark): 86.9%, an improvement of about 14 percentage points over 2.5 Flash Lite

The GPQA score is particularly noteworthy. An 86.9% result outperforms many flagship models, making this a compelling offering at the given price point.

However, it is important to note that on the more demanding HLA reasoning benchmark, Flash-Lite scores only 16%, compared to 44.4% for the 3.1 Pro. High-end reasoning tasks remain a clear differentiator.

Technical foundation and architecture

Flash-Lite is built on a Mixture-of-Experts (MoE) architecture derived from the Gemini 3 Pro model. It supports:

  • Context window: 1 million tokens
  • Input modalities: Text, images, audio, and video
  • Output length: Up to 64K tokens

The MoE architecture is key to achieving this performance at such a low cost — fewer activated parameters mean lower per-inference costs. This approach aligns with models like DeepSeek V3 and Qwen3.

Ideal use cases

Google has outlined very practical applications for this model:

  • Content moderation: Batch processing for policy violations
  • Data extraction: Extracting fields from unstructured text
  • Intent routing: First-layer classification in multi-agent systems
  • Customer service automation: Responding to standard queries
  • Translation and transcription: Large-scale language processing

In short, it is designed for tasks that do not require deep reasoning but are high-volume, latency-sensitive, and budget-constrained.

This combination is common in enterprise settings, where a significant portion of AI spending goes toward routine but essential tasks.

Competitive pricing comparison

ModelInput ($/1M tokens)Output ($/1M tokens)
Gemini 3.1 Flash-Lite$0.25$1.50
Gemini 3.1 Pro$2.00$18.00
GPT-4o Mini~$0.15~$0.60
Claude 4.5 Haiku~$0.25~$1.25

Competition in this price range is intense. GPT-4o Mini is cheaper, and Claude 4.5 Haiku is in a similar price band. Flash-Lite's advantages are its speed (2.5x faster time to first token) and its 1M-token context window, which is rare among similarly priced models.

What this launch means

With Flash-Lite, Google is clearly targeting the high-volume market segment.

For many small developers and startups building product prototypes, the primary criteria are "good enough and cheap." The price of Gemini 3.1 Pro can be a real barrier, but Flash-Lite's pricing is unlikely to be an obstacle.

At the same time, Google is using this low-cost model to keep developers within the Vertex AI ecosystem. The strategy is to attract traffic with Flash-Lite and then migrate heavier production tasks to Pro and Ultra — a conversion path Google has already validated with many customers.

For cost-sensitive API scenarios, there is now another option on the table.

Sources: CocoLoop, Google launches speedy Gemini 3.1 Flash-Lite model in preview (SiliconANGLE); Gemini 3.1 Flash Lite Review 2026: Pricing, Benchmarks, Features & Best Use Cases (AI/ML API Blog); Gemini 3.1 Flash Lite: Our most cost-effective AI model yet (Google Blog)