Gemini 3.1 Flash-Lite: $0.25/M Token Pricing

At $0.25 per million input tokens, Gemini 3.1 Flash-Lite is the cheapest multimodal inference option Google has ever shipped at this quality tier — and if your production pipeline runs classification, summarization, or light retrieval-augmented generation at scale, the math on a model switch just changed.

Google released Gemini 3.1 Flash-Lite in developer preview on March 3, 2026, and moved it to general availability on May 8, 2026 ¹. The model sits at the efficiency end of the Gemini 3.1 generation, below Gemini 3.1 Flash and Gemini 3.1 Pro. The pricing and speed numbers are worth examining carefully, because the claims Google is making are specific enough to test against your own workload.

What Google Actually Announced

Google's launch post describes Flash-Lite as delivering a "2.5× faster response time and 45% faster output generation" compared to earlier Gemini versions ¹. The comparison baseline matters: Google benchmarks these numbers against Gemini 2.5 Flash (not 2.5 Flash-Lite), which is a heavier model in the prior generation.

Third-party measurements from Artificial Analysis put Flash-Lite's output throughput at approximately 363 tokens per second, compared to 249 tokens/second for Gemini 2.5 Flash — consistent with the ~45% throughput improvement Google cited. Time to first token (TTFT) measured independently sits around 5.3–5.7 seconds at median latency on Google's API, which is longer than TTFT on Claude Haiku 4.5 or GPT-5 mini at equivalent concurrency. Flash-Lite's advantage is in throughput once the stream starts, not in absolute time-to-first-byte.

For production workloads, Google's own case study with Gladly (a customer service platform) reports p95 latency of roughly 1.8 seconds for full reply generation under heavy concurrent load, with classifiers and tool calls completing below 1 second p95 ¹. These are real-world numbers under production concurrency, not lab conditions — treat them as directionally reliable but specific to Gladly's infrastructure configuration.

What "Matches Gemini 2.5 Flash" Means

Google claims Flash-Lite "approaches Gemini 2.5 Flash performance across key capabilities." The benchmark data is more specific: Flash-Lite scores 89.2% on MMLU and 85.6% on MATH-500. For context, Gemini 2.5 Flash scored approximately 87% on MMLU in its own launch announcement. The headline is that Flash-Lite in generation 3.1 exceeds what Flash (not Flash-Lite) achieved one generation back — which, if it holds on your evaluation data, is a meaningful quality gain at the lower price tier.

Treat any benchmark comparison critically. Google's published scores use standardized evaluation sets. Your production accuracy on domain-specific classification or extraction tasks may diverge significantly from MMLU or MATH-500 numbers — benchmark hacking (overfitting eval sets during training) is a documented problem across the industry, and independent evaluation on your own labeled data remains the only reliable signal.

The Comparison Table

The decision to switch depends on your specific price-per-output-quality tradeoff. Here is the current state of the "fast and cheap" tier across providers, using prices from official pricing pages:

Model	Input ($/M tokens)	Output ($/M tokens)	Context window	Reported throughput
Gemini 3.1 Flash-Lite	$0.25	$1.50	1M tokens	~363 tok/s (Artificial Analysis)
Gemini 3.0 Flash	$0.50	$3.00	1M tokens	~232 tok/s (Artificial Analysis)
Claude Haiku 4.5	$1.00	$5.00	200K tokens	2× Sonnet 4 speed (Anthropic)
GPT-5 mini	$0.25	$2.00	400K tokens	Not published by OpenAI
Llama 4 Scout (hosted, inference.net)	$0.08	$0.30	10M tokens (provider caps vary)	Not published

Sources: Gemini pricing from Google AI for Developers pricing page ²; Claude Haiku 4.5 from Anthropic's pricing page ³; GPT-5 mini from OpenAI's API pricing page ⁴; Llama 4 Scout hosted pricing from inference.net as representative of cheapest available hosted rate ⁵.

A few notes on this table:

Gemini 3.0 Flash is the most direct predecessor for teams already on Gemini. At $0.50/$3.00, the switch to Flash-Lite cuts input cost by 50% and output cost by 50%. If your task is output-light (classification, routing, intent detection), the savings land almost entirely on the input side.
GPT-5 mini matches Flash-Lite on input price but charges $2.00 per million output tokens versus $1.50 — a 33% output cost premium. Benchmark comparisons on MMLU favour Flash-Lite (89.2% vs 78.1% per third-party LayerLens data), but GPT-5 mini has a 400K context window and different strengths on coding tasks. Neither set of benchmark numbers is independently verified to be free of evaluation contamination.
Claude Haiku 4.5 costs 4× Flash-Lite on input. The 200K context window is smaller. The model has real strengths on instruction-following precision and coding, but for high-throughput classification or summarization jobs, you are paying substantially more per token.
Llama 4 Scout is the outlier. At $0.08/$0.30 on inference.net, it is cheaper by a significant margin. The tradeoff: provider-level context caps (most hosts expose 128K–1M, not the native 10M), less predictable SLAs on smaller inference providers, and the operational overhead of evaluating an open-weight model on your data. For teams that have already done that evaluation, Scout is the cost floor.

Which Workloads Should Switch

Flash-Lite's combination of 1M context, multimodal input (text, image, video, audio, PDF), sub-$0.30 effective cost per 1M output tokens for typical classification outputs, and ~363 tokens/second throughput makes it well-suited for specific production patterns.

High-Volume Classification and Routing

Intent classification, ticket routing, content moderation flags, and similar short-input/short-output tasks are the clearest fit. These workloads call the model thousands to millions of times per day. At $0.25 input and $1.50 output per million tokens, a pipeline processing 10 million classification calls per month with an average of 200 input tokens and 20 output tokens per call costs approximately $520/month — versus $2,100/month on Claude Haiku 4.5 for the same volume. The quality bar for classification is also easier to verify: you can build a labeled eval set, measure accuracy at the 94% threshold Gladly reports, and decide based on data whether Flash-Lite's accuracy is sufficient for your routing logic.

Summarization at Scale

Document summarization pipelines — summarizing support tickets, legal documents, product descriptions, news articles — benefit from Flash-Lite's 1M token context window. The model can ingest a full long document in a single call without chunking. Output quality for summarization is harder to benchmark than classification, but Flash-Lite's MMLU and BigBench-Hard scores (87.7%) suggest solid comprehension on factual content. For summarization tasks where the output does not need to be inference-heavy or multi-step, Flash-Lite is a reasonable default starting point before benchmarking alternatives.

Embedding Backfill and Preprocessing

Flash-Lite is not an embedding model, but it functions well as a preprocessing step before embedding: cleaning, normalising, and chunking raw text before it goes into a vector store. At $0.25/M input, the cost to preprocess a large corpus is low. The 1M context window means you can pass in large documents for structured extraction without multiple API calls.

Light RAG Reranking and Answer Synthesis

In a two-stage RAG setup — retrieve N chunks, pass to model for relevance ranking or answer synthesis — Flash-Lite's throughput and cost make it viable for the synthesis stage. Google's announcement specifically calls out improved snippet ranking accuracy in RAG pipelines ¹. This claim is unverified by independent third-party evaluation as of publication, but it is specific enough to test on your own retrieval dataset with a few hours of evaluation work.

Where Flash-Lite Falls Short

Do not use Flash-Lite as a drop-in replacement for reasoning-heavy tasks. The model lacks the extended thinking capability present in Gemini 3.1 Pro and 3.1 Flash. Multi-step mathematical reasoning, complex code generation, and multi-hop logical inference are where the quality gap between the efficiency tier and the pro tier becomes measurable.

Long-Context Reasoning

Flash-Lite supports a 1M token context window, but supporting long context and performing well on long-context reasoning tasks are different. Processing a 500K-token legal document for structured extraction is feasible. Answering complex questions that require synthesising information across 500K tokens — where the model must hold and reason over many interdependent facts — is where a Flash-Lite class model will degrade. If your RAG chunks are short and retrieval is doing the heavy lifting, this is not a concern. If you are passing entire long documents and expecting reliable multi-hop reasoning, test thoroughly before committing.

Production SLA Considerations

Flash-Lite's p95 latency of ~1.8 seconds for full reply generation under heavy load is fast for a 1M-context multimodal model. But it is not uniformly fast. Under burst concurrency, tail latencies can stretch further. If your application has tight synchronous latency SLAs (sub-500ms end-to-end), evaluate carefully. The sub-1-second p95 numbers apply specifically to classifier calls, not full generation.

Benchmark Caveats

Flash-Lite's MMLU score of 89.2% and MATH-500 score of 85.6% come from Google's published benchmarks. These figures position Flash-Lite ahead of GPT-5 mini on standard academic evaluations. However, MMLU and MATH-500 are well-known to the training pipelines of all major labs, and independent contamination-controlled benchmarking consistently shows smaller gaps than headline numbers suggest. Do not use benchmark-to-benchmark comparisons as the sole basis for a model selection decision. Run your own evals on your own domain data.

Cost Modelling for Real Workloads

To make the pricing concrete: assume a summarisation pipeline that processes 50,000 documents per day, each with an average of 1,500 input tokens (after chunking) and 300 output tokens per call.

Daily token volume: 75M input tokens, 15M output tokens.

Model	Daily input cost	Daily output cost	Daily total
Gemini 3.1 Flash-Lite	$18.75	$22.50	$41.25
Gemini 3.0 Flash	$37.50	$45.00	$82.50
Claude Haiku 4.5	$75.00	$75.00	$150.00
GPT-5 mini	$18.75	$30.00	$48.75

Monthly projections: Flash-Lite at ~$1,237/month versus Haiku 4.5 at ~$4,500/month for the same volume. The savings are real, but they only materialise if Flash-Lite's quality meets your accuracy threshold. The evaluation cost to determine that is a few days of engineering time — worth spending before committing.

Google also offers a 50% batch API discount for asynchronous workloads, bringing Flash-Lite to $0.125/M input and $0.75/M output for non-latency-sensitive jobs like overnight backfill. At batch pricing, the daily cost for the pipeline above drops to approximately $20.62.

Takeaways

Gemini 3.1 Flash-Lite went GA on May 8, 2026 at $0.25/M input and $1.50/M output tokens, with a 1M token context window and multimodal input support (text, image, video, audio, PDF) ¹².
Google's claimed 2.5× faster response and 45% throughput improvement are measured against Gemini 2.5 Flash (not Flash-Lite). Third-party throughput measurements (~363 tokens/second) are consistent with the throughput claim; TTFT is less impressive at 5–6 seconds median.
The workloads worth switching first: high-volume intent classification, document summarization pipelines, RAG preprocessing, and structured extraction where input volume dominates cost. Run your own labeled evaluation before switching production traffic.
Do not use Flash-Lite for multi-step reasoning, complex code generation, or applications where benchmark-level quality on reasoning tasks is required. The model does not include extended thinking.
GPT-5 mini matches on input price but costs 33% more per output token. Claude Haiku 4.5 is 4× more expensive on input. Llama 4 Scout hosted on commodity inference is cheaper still, but comes with provider SLA tradeoffs and evaluation overhead.
The Batch API brings effective cost to $0.125/M input for asynchronous workloads — the lowest-cost option for overnight backfill tasks among cloud-hosted multimodal models as of publication.
Benchmark comparisons (Flash-Lite 89.2% MMLU vs GPT-5 mini 78.1%) should inform — not replace — evaluation on your own production data.

Google Cloud Blog, "Gemini 3.1 Flash-Lite is now generally available," May 8, 2026. https://cloud.google.com/blog/products/ai-machine-learning/gemini-3-1-flash-lite-is-now-generally-available ↩↩↩↩↩
Google AI for Developers, Gemini Developer API Pricing. https://ai.google.dev/gemini-api/docs/pricing ↩↩
Anthropic, Claude pricing. https://www.anthropic.com/claude/haiku ↩
OpenAI, API Pricing. https://openai.com/api/pricing/ ↩
inference.net, LLM API Pricing Comparison 2026. https://inference.net/content/llm-api-pricing-comparison/ ↩

Gemini 3.1 Flash-Lite at $0.25/M Tokens: What Builders Should Switch