Google has launched Gemini 3.1 Flash-Lite, probably the most cost-efficient entry within the Gemini 3 mannequin sequence. Designed for ‘intelligence at scale,’ this mannequin is optimized for high-volume duties the place low latency and cost-per-token are the first engineering constraints. It’s presently accessible in Public Preview through the Gemini API (Google AI Studio) and Vertex AI.
Core Characteristic: Variable ‘Pondering Ranges’
A big architectural replace within the 3.1 sequence is the introduction of Pondering Ranges. This function permits builders to programmatically modify the mannequin’s reasoning depth primarily based on the precise complexity of a request.
By deciding on between Minimal, Low, Medium, or Excessive pondering ranges, you may optimize the trade-off between latency and logical accuracy.
- Minimal/Low: Supreme for high-throughput, low-latency duties similar to classification, primary sentiment evaluation, or easy information extraction.
- Medium/Excessive: Makes use of Deep Assume Mini logic to deal with complicated instruction-following, multi-step reasoning, and structured information era.
Efficiency and Effectivity Benchmarks
Gemini 3.1 Flash-Lite is designed to exchange Gemini 2.5 Flash for manufacturing workloads that require quicker inference with out sacrificing output high quality. The mannequin achieves a 2.5x quicker Time to First Token (TTFT) and a 45% enhance in general output velocity in comparison with its predecessor.
On the GPQA Diamond benchmark—a measure of expert-level reasoning—Gemini 3.1 Flash-Lite scored 86.9%, matching or exceeding the standard of bigger fashions within the earlier era whereas working at a considerably decrease computational price.
Comparability Desk: Gemini 3.1 Flash-Lite vs. Gemini 2.5 Flash
| Metric | Gemini 2.5 Flash | Gemini 3.1 Flash-Lite |
| Enter Value (per 1M tokens) | Larger | $0.25 |
| Output Value (per 1M tokens) | Larger | $1.50 |
| TTFT Velocity | Baseline | 2.5x Quicker |
| Output Throughput | Baseline | 45% Quicker |
| Reasoning (GPQA Diamond) | Aggressive | 86.9% |
Technical Use Circumstances for Manufacturing
The three.1 Flash-Lite mannequin is particularly tuned for workloads that contain complicated constructions and long-sequence logic:
- UI and Dashboard Technology: The mannequin is optimized for producing hierarchical code (HTML/CSS, React elements) and structured JSON required to render complicated information visualizations.
- System Simulations: It maintains logical consistency over lengthy contexts, making it appropriate for creating setting simulations or agentic workflows that require state-tracking.
- Artificial Knowledge Technology: Because of the low enter price ($0.25/1M tokens), it serves as an environment friendly engine for distilling information from bigger fashions like Gemini 3.1 Extremely into smaller, domain-specific datasets.
Key Takeaways
- Superior Worth-to-Efficiency Ratio: Gemini 3.1 Flash-Lite is probably the most cost-efficient mannequin within the Gemini 3 sequence, priced at $0.25 per 1M enter tokens and $1.50 per 1M output tokens. It outperforms Gemini 2.5 Flash with a 2.5x quicker Time to First Token (TTFT) and 45% greater output velocity.
- Introduction of ‘Pondering Ranges’: A brand new architectural function permits builders to programmatically toggle between Minimal, Low, Medium, and Excessive reasoning intensities. This supplies granular management to stability latency in opposition to reasoning depth relying on the duty’s complexity.
- Excessive Reasoning Benchmark: Regardless of its ‘Lite’ designation, the mannequin maintains high-tier logic, scoring 86.9% on the GPQA Diamond benchmark. This makes it appropriate for expert-level reasoning duties that beforehand required bigger, dearer fashions.
- Optimized for Structured Workloads: The mannequin is particularly tuned for ‘intelligence at scale,’ excelling at producing complicated UI/dashboards, creating system simulations, and sustaining logical consistency throughout long-sequence code era.
- Seamless API Integration: Presently accessible in Public Preview, the mannequin makes use of the
gemini-3.1-flash-lite-previewendpoint through the Gemini API and Vertex AI. It helps multimodal inputs (textual content, picture, video) whereas sustaining a typical 128k context window.
Take a look at the Public Preview through the Gemini API (Google AI Studio) and Vertex AI. Additionally, be at liberty to observe us on Twitter and don’t overlook to hitch our 120k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
