by Google DeepMind

Gemini 3.1 Flash-Lite review, pricing and verdict

Google's most cost-efficient Gemini 3 model, built for high-volume classification, translation, and lightweight agent tasks rather than frontier reasoning.

Best for
High-volume classification and tagging
Not for
Autonomous coding agents
  • Available
  • Proprietary

Should you use Gemini 3.1 Flash-Lite?

Flash-Lite is Google's pick for high-volume, latency-sensitive work: classification, translation, and RAG retrieval, not deep reasoning or coding agents. Its 1432 LMArena Elo score reflects solid real-world chat quality despite weaker academic-reasoning benchmarks, so save Gemini 3.1 Pro or Gemini 3.5 Flash for tasks that need more careful reasoning.

Best for

  • High-volume classification and tagging $0.25/$1.50 per 1M tokens and 301 t/s output make per-request cost negligible at scale.
  • Translation and RAG retrieval over long documents 1,048,576-token context window handles large documents in a single call.
  • Lightweight tool-using agents and support bots Native function calling and structured output with adjustable thinking levels for cost control.

Not for

  • Autonomous coding agents No published SWE-bench Verified score; Gemini 3.1 Pro and Claude Opus 4.5 publish far stronger agentic-coding numbers.
  • Competition math or research-grade reasoning 16.0% on Humanity's Last Exam trails Gemini 3.1 Pro and Gemini 3.5 Flash by a wide margin; Google has not published an AIME 2025 score for this model.
  • Real-time voice assistants ~5.2s time-to-first-token is too slow; use Gemini 3.1 Flash Live instead.

Gemini 3.1 Flash-Lite pricing

Input, per 1M tokens
$0.25/1M
Output, per 1M tokens
$1.50/1M
Cached input, per 1M tokens
$0.05/1M
Pricing model
per-token
Free tier
No

What a real job costs

JobInputOutputTotal
Summarise a 20-page PDF$0.0075$0.0015$0.0090
Support reply$0.0005$0.0004$0.0009
One coding agent run$0.050$0.030$0.080

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Cost examples

1,000 leads/day classification, ~400-token avg response
$0.62
Summarize a 100K-token document with a 1K-token output
$0.03
High-volume tagging: 1M requests/day at 50 tokens in / 10 tokens out
$27.50

$0.25 per 1M input tokens and $1.50 per 1M output tokens via the Gemini API and Vertex AI, with thinking tokens billed at the output rate. Cached input is priced at $0.05 per 1M tokens. Context caching can cut repeated-prompt input costs by up to 90% for workloads that reuse the same system prompt or reference document.

What Gemini 3.1 Flash-Lite can do

Gemini 3.1 Flash-Lite scores 86.9% on GPQA Diamond, the headline benchmark Google published at its March 2026 launch. It is Google DeepMind's most cost-efficient Gemini 3 model, a sparse mixture-of-experts transformer with a 1M-token context window built for high-volume, latency-sensitive workloads rather than frontier reasoning.

What goes in and what comes out

Reads

  • text
  • image
  • audio
  • video
  • pdf

Writes

  • text
  • tool-calls

Capabilities

  • Audio inputSupported
  • Code executionSupported
  • Function callingSupported
  • Structured outputSupported
  • Tool useSupported
  • Video inputSupported
  • VisionSupported
  • Audio outputNot supported
  • Web browsingNot supported
Reasoning modes
minimal, low, medium, high
Long context recall
good

Key features

  • 1M-Token Context Window

    A 1,048,576-token input window paired with a 65,536-token output ceiling, well beyond what most similarly priced models offer.

  • Four Thinking Levels

    Four selectable reasoning depths, from a fast minimal setting to a deep high setting, put latency and spend under direct developer control per request.

  • Multimodal Input

    Reads text, images, audio, video, and PDF documents in one call and returns parseable JSON or native tool calls for agent pipelines.

  • Context Caching

    Caching a reused system prompt or reference document cuts its input-token cost by up to 90% versus resending it on every call.

  • Cost Efficiency

    Priced below most same-tier rivals on a blended per-token basis, with caching further cutting repeat-workload costs; exact per-token rates are in the pricing FAQ.

Known limitations

  • Time-to-first-token of about 5.2 seconds is slower than the roughly 2-second median for similarly priced models.

Who it is for

  • Teams running high-volume classification, tagging, or translation pipelines
  • RAG and retrieval-augmented search applications needing a 1M-token context at low cost
  • Customer support and lightweight agent products with tool-calling but shallow planning
  • Cost-sensitive teams migrating off Gemini 2.5 Flash-Lite before its July 22, 2026 shutdown

Supported languages

  • English
  • Spanish
  • French
  • German
  • Japanese
  • Korean
  • Chinese
  • Portuguese
  • Arabic
  • Hindi

Gemini 3.1 Flash-Lite benchmarks

Gemini 3.1 Flash-Lite against its nearest rivals

GPQA Diamond % correct, higher is better

Output speed tokens per second, higher is better

Standout result

86.9%

GPQA Diamond score, the headline benchmark Google published for Flash-Lite

Source: Artificial Analysis / Google DeepMind, March 2026 launch

Where it sits

Cheaper than 67% of the 106 GA models with a published price, mid-pack on GPQA Diamond (rank 34 of 63), and one of 90 whose vendor states it does not train on customer data.

MeasureThis modelRankMedian
Blended price (3:1)$ per 1M tokens$0.563/M#36 / 106peer median $1.73/M
Output speedtokens/s301 tok/s#5 / 59peer median 94 tok/s
GPQA Diamond% correct86.9%#34 / 63peer median 87.8%

Ranks are against GA models on Hokai that publish the same figure; ties share a rank.

Benchmark scores

  • GPQA Diamond vendor-reported · 08 Sep 202686.9%PhD-level science questions that are hard to search for, % correct.
  • MMMU-Pro vendor-reported · 08 Sep 202676.8%College-level questions that need reading images and diagrams, % correct.
  • Humanity's Last Exam vendor-reported · 08 Sep 202616%Expert-written questions across many fields, % correct.
  • AA Intelligence Index cited: Artificial Analysis · 08 Sep 202616Composite of the evaluations Artificial Analysis runs for its index, 0 to 100.

Measured figures

LMArena Elo
1432
Output speed
301 tok/s

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Strengths and limits

Strengths

  • Blended per-token cost runs about 3.4x below Claude Haiku 4.5's, per this record's price comparison.
  • The 1M-token context window comfortably handles full support tickets, transcripts, or medium-length codebases in a single call, more headroom than most models in this price tier offer.
  • Beats GPT-5 mini and Claude Haiku 4.5 on 6 of 11 benchmarks Google published at launch.
  • Runs at roughly 301 tokens per second per Artificial Analysis, among the faster models in its price tier.
Show 2 more
  • 1,048,576-token context window is 5x larger than Claude Haiku 4.5's 200K limit, with 65,536 max output tokens.
  • At $0.25/$1.50 per 1M tokens it is roughly 3.4x cheaper than Claude Haiku 4.5 and 40% cheaper on output than its predecessor Gemini 2.5 Flash.

Limits

  • Humanity's Last Exam (16.0%) is Flash-Lite's weakest published result, trailing the rest of the Gemini 3 lineup; Google has not published an AIME 2025 or MMLU-Pro score for this model to compare.
  • Time-to-first-token runs about 5.2 seconds, noticeably slower to start responding than most same-tier models, likely because thinking is enabled by default.
  • No native audio or image output; voice and image generation require separate Gemini 3.1 Flash Audio and Flash Image models.
  • Not on AWS Bedrock or Azure AI Foundry, limiting multi-cloud deployment options.

Gemini 3.1 Flash-Lite alternatives

ModelLabContextIn / out per 1MReads and writes
Gemini 3.1 Flash-Lite This pageGoogle1M$0.25 / $1.50text, image, audio, video, pdf → text
Gemini 3.1 ProPick Flash-Lite for high-volume, low-cost workloads; pick 3.1 Pro when you need its 94.3% GPQA Diamond score.Google1M$2 / $12text, image, audio, video, pdf → text
Gemini 3.5 Flash-LitePick 3.1 Flash-Lite when price is the deciding factor; pick 3.5 Flash-Lite for Google's newer generation if the higher per-token rate is acceptable.Google1M$0.30 / $2.50text, image, video, audio, pdf → text
GPT-4o miniPick Flash-Lite for its much larger context window; pick GPT-4o mini if you're already standardized on OpenAI's ecosystem and want a lower per-token price.OpenAI128K$0.15 / $0.60text, image → text

Also compared

  • Ministral 3 8BPick Flash-Lite for stronger benchmark scores and a larger context window; pick Ministral 3 8B for its flat input/output token rate.

Gemini 3.1 Flash-Lite specs

Context window
1.0M tokens
Max output
66K tokens
Latency (P50)
5180 ms
Tokens per second
300.9 tok/s
Mixture of experts
Yes
Parameters
Not disclosed
Architecture
Transformer (sparse MoE, Gemini 3 architecture)

Training data

Trained on a curated mix of public web text, licensed datasets, and synthetic data as part of the Gemini 3 generation; knowledge cutoff January 2025. Aligned with Google's standard safety training and red-teaming process described in the model card.

API endpoints

NameEndpointAuth
Gemini APIhttps://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-liteAPI Key
Vertex AIhttps://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-1-flash-liteGCP IAM

SDKs

  • Python
  • JavaScript
  • TypeScript
  • Go
  • Java

Regions

  • us-central1
  • europe-west4
  • asia-southeast1

Prompting patterns

Show 2 prompting patterns
  • Set the thinking level explicitly

    generationConfig: { thinkingConfig: { thinkingLevel: 'minimal' } }

    Default thinking behavior can burn output-rate tokens on simple classification prompts; pinning to 'minimal' or 'low' keeps cost predictable.

  • Cache repeated system prompts and documents

    Use the Gemini context caching API for any system prompt or reference document reused across requests

    Cached reads can cut input costs by up to 90% on repeat-context workloads.

Known quirks

Show 4 known quirks
  • Slow time-to-first-token for a 'lite' model

    Artificial Analysis measures Flash-Lite's time-to-first-token at roughly 5.2 seconds on Google's API, versus a roughly 2-second median for similarly priced models, likely due to default thinking behavior.

    Workaround: Set thinkingLevel to 'minimal' for latency-sensitive endpoints, stream responses, and show a loading state immediately rather than waiting for the first token.

  • Thinking tokens are billed at the output rate

    Leaving the thinking level at 'medium' or 'high' on simple extraction or classification prompts silently increases output token usage and cost.

    Workaround: Set thinkingLevel to 'minimal' or 'low' for classification, extraction, and translation tasks, and reserve 'medium'/'high' for prompts that genuinely need multi-step reasoning.

  • Preview model ID was retired mid-2026

    gemini-3.1-flash-lite-preview launched March 3, 2026, was deprecated May 11, 2026, and was shut down May 25, 2026 in favor of the GA gemini-3.1-flash-lite ID.

    Workaround: Pin production traffic to gemini-3.1-flash-lite (no -preview suffix) and monitor Google's Gemini API release notes for future renames.

  • Weak on expert-level reasoning relative to siblings

    16.0% on Humanity's Last Exam is far below Gemini 3.1 Pro and Gemini 3.5 Flash, even though Flash-Lite leads on cost and speed; Google has not published an AIME 2025 score for this model.

    Workaround: Route math-heavy or PhD-level reasoning prompts to Gemini 3.1 Pro or Gemini 3.5 Flash, and keep Flash-Lite for retrieval, classification, and short-form generation.

Gemini 3.1 Flash-Lite facts

Deployment
API · Google Vertex AI
Trains on your data
No
Safety posture
Balanced
Released
03 Mar 2026
Current version
gemini-3.1-flash-lite
Knowledge cutoff
01 Jan 2025
Family
Gemini 3.1
Variant
Multimodal
Licence
Proprietary (Google Generative AI / Gemini API Additional Terms of Service)
Data retention
Standard Gemini API data handling per Google's Generative AI Additional Terms of Service; Vertex AI offers enterprise-grade retention and data residency controls. See the official model card for full details.

Content restrictions

  • Standard Google content policy refusals apply (CSAM, weapons of mass destruction, malware generation).
  • Model card reports low unjustified refusal rates relative to Gemini 2.5 Flash-Lite.

Safety controls

Rlhf
Yes
Red teaming
documented in Google DeepMind model card
System prompt
Yes
Constitutional AI
No
Content filter API
Yes
Moderation endpoint
No

Data governance

Trains on inputs
No
Data residency options
us, eu, asia

Version history

  • gemini-3.1-flash-lite · 25 May 2026

    General availability release. Replaced gemini-3.1-flash-lite-preview, which was shut down the same day.

  • gemini-3.1-flash-lite-preview · 03 Mar 2026

    Initial preview launch. 86.9% GPQA Diamond, 1432 LMArena Elo, Artificial Analysis Intelligence Index 34. Deprecated 2026-05-11, shut down 2026-05-25.

What changed

  • Deprecation Gemini 2.5 Flash-Lite scheduled for shutdown source
  • Release Gemini 3.1 Flash-Lite reaches general availability source
  • Release Gemini 3.1 Flash-Lite preview launch source

Questions about Gemini 3.1 Flash-Lite

How much do you pay for Gemini 3.1 Flash-Lite?

It runs $0.25 for every 1M input tokens and $1.50 for every 1M output tokens through the Gemini API or Vertex AI, with thinking tokens counted as output and cached input billed separately at $0.05 per 1M tokens. There is a single Standard pay-as-you-go tier; Google has not published a discounted committed-use rate for this model. A workload classifying leads at roughly 400 tokens per response costs a fraction of a cent per request.

Does Gemini 3.1 Flash-Lite have a free plan?

No, Gemini 3.1 Flash-Lite has no free tier. Every request through the Gemini API or Vertex AI is billed per token from the start, so evaluating it means budgeting a small amount for test calls rather than testing for free.

What should you use instead of Gemini 3.1 Flash-Lite?

GPT-4o mini is worth checking when a 128,000-token context window is enough and a lower per-token price matters more than benchmark score. Ministral 3 8B bills a flat rate for input and output tokens, which suits teams that want predictable pricing over a stronger GPQA Diamond result. Gemini 3.5 Flash-Lite is Google's newer sibling in the same family, worth it if you want the latest generation and can accept a higher per-token rate.

Is Gemini 3.1 Flash-Lite better than GPT-4o mini?

The two publish different benchmark suites, but the gap is clear on context and price: Flash-Lite's context window is far larger than GPT-4o mini's 128,000 tokens, while GPT-4o mini undercuts Flash-Lite on input price. Flash-Lite also publishes a GPQA Diamond score that GPT-4o mini has not reported. Pick Flash-Lite for long-document work or when a documented reasoning benchmark matters; pick GPT-4o mini when requests are short and the lowest per-token price wins.

How do you get started with Gemini 3.1 Flash-Lite?

Get a Gemini API key from Google AI Studio, or enable the model in an existing Google Cloud project through Vertex AI. Send a request specifying gemini-3.1-flash-lite as the model ID, since the older preview ID was retired. Set the thinking level to minimal for simple classification or extraction tasks to keep latency and cost predictable, and raise it only for prompts that need multi-step reasoning.

About Gemini 3.1 Flash-Lite

Gemini 3.1 Flash-Lite is Google DeepMind's most cost-efficient model in the Gemini 3 family, first previewed on March 3, 2026 and promoted to general availability on May 25, 2026, when the preview model ID, gemini-3.1-flash-lite-preview, was retired. It sits below Gemini 3.1 Flash and Gemini 3.1 Pro in Google's lineup, built as a sparse mixture-of-experts transformer optimized for high-volume, latency-sensitive traffic such as classification, translation, retrieval, and lightweight agent steps. It replaces Gemini 2.5 Flash-Lite, which Google is shutting down on July 22, 2026.

Read the full background

On Google's published evaluation suite, Flash-Lite scores 86.9% on GPQA Diamond and 76.8% on MMMU Pro for multimodal understanding, with a Chatbot Arena Elo of 1432 and an Artificial Analysis Intelligence Index of 16. Reasoning-heavy benchmarks lag the rest of the lineup: it manages 16.0% on Humanity's Last Exam, well below its Gemini 3.1 and 3.5 siblings. Google has not published an MMLU-Pro, AIME 2025, or SWE-bench Verified score for Flash-Lite, a signal the company isn't positioning this tier for broad academic reasoning or autonomous coding.

The model accepts up to 1,048,576 tokens of input and can generate up to 65,536 output tokens in a single response, a window 5x larger than Claude Haiku 4.5's 200K limit. Beyond the wider Gemini 3 family's general long-context claims, Google has not broken out a recall benchmark for this smaller variant specifically, so treat performance past roughly 500K tokens as good rather than independently confirmed.

Flash-Lite accepts text, image, audio, video, and PDF input and returns text plus tool-call output. It supports function calling, structured JSON output, and four thinking levels, minimal, low, medium, and high, that let developers trade reasoning depth for latency and spend on a per-request basis. It does not produce native audio or image output; Google ships those separately as Gemini 3.1 Flash Audio and Gemini 3.1 Flash Image.

Pricing is well below Gemini 2.5 Flash's $0.30/$2.50 per-1M-token rate on output, despite Flash-Lite scoring higher across reasoning and multimodal benchmarks, and its blended per-token cost undercuts Claude Haiku 4.5 too. Exact current pricing and a sample workload cost are in the pricing FAQ below.

Flash-Lite is available through the direct Gemini API with API key authentication, through Vertex AI with GCP IAM, and through the Gemini Enterprise Agent Platform. Google has not extended distribution to AWS Bedrock or Azure AI Foundry as of mid-2026. As a closed, proprietary model there are no downloadable weights, quantization options, or self-hosting paths.

The model's training data has a knowledge cutoff of January 2025. Google's model card states that, based on Gemini 3.1 Pro's capability assessments, Flash-Lite is unlikely to reach any Critical Capability Level, and that it improves on Gemini 2.5 Flash-Lite for both safety and tone while keeping unjustified refusals low. Output speed measured by Artificial Analysis is about 301 tokens per second, though time-to-first-token runs around 5.2 seconds, slower than the roughly 2-second median for similarly priced models, likely reflecting the model's default thinking behavior.

On the release timeline, gemini-3.1-flash-lite-preview launched March 3, 2026, was marked deprecated May 11, 2026, and was shut down May 25, 2026 in favor of the GA gemini-3.1-flash-lite model ID that teams should pin to going forward. Data handling follows Google's standard Gemini API terms, with enterprise residency and retention controls available through Vertex AI.

Sources and guides

More AI Models on Hokai

  • Hy4 flashTencent's Hy4 flash, released 16 November 2026, is a lower-cost multimodal model that reads text and images across a 1M-token window.
  • Nemotron-Labs-SigmaNVIDIA's Nemotron-Labs-Sigma, released in November 2026 in 8B and 3B sizes, writes text by denoising whole blocks of tokens in parallel, not one at a time.
  • Qwen4 MaxAlibaba's Qwen4 Max, released 16 November 2026, is a sparse mixture-of-experts model with a 1M-token context and text, image and video input.
  • Gemini 4 HydrogenGoogle DeepMind's November 2026 Gemini 4 model for coding agents, with four thinking levels from low to xhigh and a free tier in AI Studio.
  • Qwen4-Max-PreviewAlibaba's November 2026 Qwen 4 flagship preview, open only to Token Plan subscribers, with three reasoning levels from low to xhigh.
  • Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context.