Gemini 3.1 Pro vs Gemini 3 Flash vs Gemini 3.1 Flash-Lite: Model Comparison (2026)

Google's Gemini 3 generation ships in three tiers you can call from the API today: Gemini 3.1 Pro, Gemini 3 Flash, and Gemini 3.1 Flash-Lite, all with a 1 million token input context window. In the Gemini 3.1 Pro vs 3 Flash matchup, the short answer is: 3.1 Pro is the top-tier reasoning and agentic-coding model, 3 Flash is the frontier-class workhorse that actually outperforms Pro on SWE-Bench Verified at roughly a quarter of the price, and 3.1 Flash-Lite is the lowest-cost high-volume tier. This guide compares them on context, pricing tiers, benchmarks, and practical use cases using verified data from ai.google.dev, Vertex AI docs, and Google DeepMind model cards.

Choosing between Gemini 3.1 Pro, Gemini 3 Flash, and Gemini 3.1 Flash-Lite is a decision about reasoning quality, agentic capability, latency, and cost per million tokens. The three tiers share the same 1M-token input context window and the same full multimodal input support (text, images, audio, video, PDF). They differ in model intelligence, price by a factor of roughly 8x from the cheapest tier to the most expensive on the low-prompt Standard tier, and in how they behave inside agentic, tool-using workflows. Getting the choice right is the single biggest lever for cost control on any Gemini-backed application in 2026.

The Gemini 3 Generation: Shared Foundations

All three Gemini 3 models share a 1 million token input context window and native multimodal input per Google's official model documentation at ai.google.dev. That baseline matters because it means choosing a cheaper tier does not automatically cost you the long-context capabilities that make the Gemini family distinctive.

One million input tokens is roughly 750,000 words. That maps to entire codebases, multi-hour audio transcripts, full book manuscripts, or hundreds of PDF pages in a single call. Google's long-context documentation covers the supported input types and performance characteristics. For any workload where the model needs to reason across a large corpus in a single shot, the Gemini 3 family is a natural fit, and you do not need to jump to Pro to get it.

Output budgets differ slightly but not meaningfully: Gemini 3.1 Pro allows up to 64K output tokens per call, Gemini 3 Flash allows 65,535, and Gemini 3.1 Flash-Lite allows 65,536. For any realistic workload that does not involve generating a novel in one call, this is not a differentiator.

Every model in the generation accepts text, images, audio, video, and PDF inputs natively, and a separate sibling model gemini-3.1-flash-live-preview offers real-time audio streaming. That unified multimodal baseline means you can build a document understanding workflow, a vision workflow, or a voice workflow on the same tier without stitching providers together.

Gemini 3.1 Pro: Top-Tier Reasoning and Agentic Coding

Gemini 3.1 Pro is Google's current top-tier reasoning model, positioned by Google as "advanced intelligence, complex problem-solving skills, and powerful agentic and vibe coding capabilities" on the DeepMind model page. It is the tier to reach for when output quality directly affects outcome: production code generation, complex multi-step reasoning, hard math and science problems, long-form drafting that needs coherent argument structure, and agentic workflows where a wrong step is expensive.

Gemini 3.1 Pro ships with measurable improvements in instruction following, tool use, and long-horizon task execution. Inside Google's Antigravity agentic environment, 3.1 Pro supports full-stack runtimes, server-side logic, secrets management, and npm packages, which makes it realistically usable for ship-quality agent workflows and not just demos. This is the practical difference between 3.1 Pro and 3 Flash on the agentic axis: both can plan and call tools, but 3.1 Pro holds up over longer sequences when a wrong step is costly.

Per Google's Gemini Developer API pricing page, Gemini 3.1 Pro Standard-tier pricing is $2.00 per 1M input tokens for prompts at or below 200K tokens and $4.00 per 1M input tokens for prompts above 200K. Output pricing is $12.00 per 1M tokens for completions at or below 200K input tokens and $18.00 per 1M tokens above. That two-tier structure reflects the real cost of attending across very long contexts.

The right workloads for 3.1 Pro are high-value reasoning tasks where the output value outstrips the per-token cost. A structured code review at 3.1 Pro pricing costs a fraction of the engineering time saved by a correct review. A complex multi-table SQL generation against a long schema prompt is also a good fit. Long-running agents that need to plan and execute tool calls across many steps are the canonical 3.1 Pro workload and the one Google explicitly positions the model for.

Gemini 3 Flash: Frontier Class at a Fraction of the Cost

Gemini 3 Flash is the most surprising model in the generation because it actually outperforms Gemini 3 Pro on SWE-Bench Verified (78.0% vs 76.2%) while costing roughly a quarter as much on Standard-tier input. Google positions it as "frontier-class performance rivaling larger models at a fraction of the cost," and the benchmark data on vals.ai supports that claim.

Standard-tier pricing for Gemini 3 Flash is $0.50 per 1M input tokens and $3.00 per 1M output tokens. That is 4x cheaper than 3.1 Pro on input (low tier) and 4x cheaper on output. For any workload where 3 Flash meets your quality bar, switching from 3.1 Pro to 3 Flash is the single biggest cost reduction you can make without changing architecture. At high volume, the difference is the difference between a sustainable cost structure and one that forces rate limiting.

Workloads that fit 3 Flash well include chat applications, RAG answer synthesis, document question answering, summarization at scale, moderation, content classification with long context, and most agent sub-steps that are not themselves the bottleneck reasoning call. SWE-Bench Verified tells you that 3 Flash is also surprisingly strong on code tasks, which widens the set of workloads where it replaces Pro without measurable quality loss.

Google positions 3 Flash specifically for agentic workflows, multi-turn chat, and coding assistance latency wins. In practice, 3 Flash is the new default tier for most application traffic on the Gemini API, and the reason to jump to 3.1 Pro is because you have a reasoning-quality ceiling that 3 Flash cannot clear, not because you assume bigger is better.

Gemini 3.1 Flash-Lite: Cheapest and Fastest

Gemini 3.1 Flash-Lite is the entry tier, optimized for latency and cost at high volume. Google describes it as "frontier-class performance rivaling larger models at a fraction of the cost" on the model card, and VentureBeat reports that it comes in at approximately one eighth of the cost of Gemini 3 Pro.

Flash-Lite Standard-tier pricing is $0.25 per 1M input tokens and $1.50 per 1M output tokens. That is half the input price of 3 Flash and half the output price. At scale, the difference between 3 Flash and 3.1 Flash-Lite is material: a workflow that costs a few hundred dollars a day on 3 Flash might cost under a hundred on 3.1 Flash-Lite for a compatible workload.

Flash-Lite is also the best tier for latency-sensitive UI. Autocomplete, typeahead, short responses in voice interfaces, first-pass classification, intent routing, and any interaction where the user is waiting for a response all benefit from the smaller model's faster time-to-first-token. On reasoning-heavy benchmarks, Flash-Lite still clears useful thresholds: 86.9% on GPQA Diamond and 90.5% on HumanEval per the model card, which is strong for the price point.

Where Flash-Lite stops being the right answer is at the edge of complex multi-step reasoning and long-horizon agent planning. There you either escalate to 3 Flash for a middle-tier cost bump, or to 3.1 Pro when the reasoning depth genuinely cannot be compressed.

Gemini 3 Generation: Full API Pricing Table (Standard Tier)

All prices are Standard-tier, per 1 million tokens, verified against the Gemini Developer API pricing page and Vertex AI pricing as of April 11, 2026. Multimodal input (text, images, audio, video, PDF) is supported across all three models.

Flex, Batch, and Priority Tiers: Pricing Modifiers

Gemini 3 pricing is not a single number per model. Google publishes four service tiers that apply across the family: Standard, Batch, Flex, and Priority. The tier you pick changes the effective price and the latency and reliability guarantees you get back.

• Standard is the baseline, synchronous API tier with normal latency and reliability. The prices in the table above are Standard-tier.
• Batch API runs requests asynchronously at roughly a 50% discount versus Standard. Good for evaluation runs, bulk data processing, offline summarization, and any workload where you can accept delayed completion in exchange for cost savings. Google documents the Batch API separately on ai.google.dev.
• Flex is a best-effort tier with variable latency at roughly a 50% discount versus Standard. Good for background jobs where you want cheap synchronous access but can tolerate higher tail latency when the provider is under load.
• Priority is the premium tier for business-critical, lowest-latency workloads and is roughly 1.75x to 2x Standard pricing. Use it when a late response is worse than a slightly higher bill - interactive chat, voice interfaces, and latency-SLO workloads.

Exact per-model Flex, Batch, and Priority line items for the Gemini 3 generation are not published as a separate table on ai.google.dev. Google's public documentation describes the multiplier policy at the tier level and applies it across models. For budgeting, treat Flex and Batch as approximately half of Standard and treat Priority as approximately 1.75x to 2x Standard until you run a live test. Google's Flex and Priority inference post and the priority inference docs are the authoritative references.

Benchmark Highlights

Gemini 3.1 Pro leads the Gemini lineup on reasoning benchmarks, but Gemini 3 Flash quietly wins on coding agents. The benchmark table below uses data verified against Vellum's Gemini 3 benchmarks, the DeepMind model cards, and third-party leaderboards.

SWE-Bench Verified is the one most people care about in 2026. Gemini 3 Flash at 78.0% is within two points of 3.1 Pro at 80.6% and scores higher than Gemini 3 Pro's earlier 76.2% result. For reference, Claude Opus 4.6 leads SWE-Bench Verified at 80.9% and GPT-5.2 / 5.4 sit at roughly 80%. The three top-tier reasoning and coding models are within a point of each other on this metric, which reinforces that the Gemini decision is mostly about price and context, not raw capability, unless you are at the absolute frontier.

API Model IDs and Migration Notes

Use these model IDs in your API calls:

• gemini-3.1-pro-preview (Gemini 3.1 Pro)
• gemini-3-flash-preview (Gemini 3 Flash)
• gemini-3.1-flash-lite-preview (Gemini 3.1 Flash-Lite)
• gemini-3.1-flash-live-preview (sibling real-time audio model)

Google still carries the -preview label on the Gemini 3 API IDs as of April 11, 2026, which reflects the internal release track rather than product maturity - these are the current frontier models Google is shipping for active production use. The critical migration note: teams using the older gemini-3-pro-preview string were required to move to gemini-3.1-pro-preview by March 9, 2026 per a Google developer forum announcement. Any codebase still referencing the old ID will fail by default and should be updated.

Gemini 2.5 Deprecation Timeline

Gemini 2.5 Pro, 2.5 Flash, and 2.5 Flash-Lite are on a published retirement path. Per Google's Gemini API deprecations page, the AI Studio and Gemini API deprecation date for Gemini 2.5 Pro is June 17, 2026. Per Vertex AI documentation, the Vertex AI retirement date is October 16, 2026, with supervised-fine-tuned 2.5 Pro endpoints shutting down the following day. Gemini 2.5 Flash and 2.5 Flash-Lite follow the same timeline.

The practical consequence is simple: new projects should default to Gemini 3, and existing 2.5-based projects should plan migration before mid-2026. The Gemini 3 family is not a breaking change at the API level - model ID is the main thing that changes - but prompts that were tuned to 2.5 behavior can drift on 3, so expect to re-evaluate your eval set when you cut over.

There Is No Gemini Ultra API Model

One important clarification: "Gemini Ultra" is not an API model in 2026. The name "Ultra" refers only to the Google AI Ultra consumer subscription tier on one.google.com and gemini.google.com. That tier bundles access to Gemini 3.1 Pro, Veo 3.1, Nano Banana Pro, and Project Mariner at the highest quotas, plus 30TB storage and YouTube Premium, for $249.99 per month.

When selecting a Gemini API model for a production application, the three production tiers are 3.1 Pro, 3 Flash, and 3.1 Flash-Lite. If you see "Gemini Ultra" referenced as an API model in a third-party article, it is outdated. Verify against the current model documentation before making a commitment.

When to Pick 3.1 Pro vs 3 Flash vs 3.1 Flash-Lite

Pick Gemini 3.1 Pro when reasoning quality or agent reliability dominates cost. Long-horizon agentic workflows, complex code generation, hard math and science problems, long-form structured writing, and any workload where a wrong step is expensive. This is the tier for agentic coding, code review, and production-grade planning over tool calls.

Pick Gemini 3 Flash for most production traffic. Chat applications, RAG answer synthesis, document question answering, summarization at scale, content moderation, classification with long context, and most agent sub-steps. The surprising SWE-Bench Verified result means 3 Flash is also a strong default for non-critical coding workloads. This is the new default tier in 2026, and the reason to escalate to 3.1 Pro is a concrete quality ceiling, not an assumption that bigger is better.

Pick Gemini 3.1 Flash-Lite when throughput, latency, and unit economics dominate. High-volume intent classification, short structured outputs, autocomplete, typeahead, voice interface first-pass responses, lightweight extraction, and anywhere a smaller model's quality is verified good enough through evaluation on your own data.

A practical 2026 architecture often uses all three: Flash-Lite for fast pre-processing and routing, 3 Flash for the main application calls, and 3.1 Pro for escalation when confidence is low or when the user explicitly requests deep reasoning. That hybrid architecture minimizes total cost while preserving quality on the calls that matter most.

Whichever Gemini model you use, your conversations are worth keeping. Gemini Toolbox adds full-text search, date / role / exact-match filters, and four-format export (TXT, Markdown, JSON, PDF) on top of gemini.google.com. From the makers of ChatGPT Toolbox (18,000+ users, 4.5/5 Chrome Web Store rating). Install Gemini Toolbox free on Chrome.

Multimodal Support Across the Gemini 3 Family

All three Gemini 3 models accept text, images, audio, video, and PDF inputs natively. That baseline is one of the most cost-effective reasons to pick the Gemini family over single-modality alternatives. A document workflow can accept PDFs with embedded images, a support workflow can accept screenshots, a media workflow can accept audio and video clips directly, all without stitching multiple model providers together.

If you need real-time audio streaming specifically (voice agents, real-time transcription with follow-up responses), Google ships a dedicated sibling model gemini-3.1-flash-live-preview. The Live model is not a replacement for 3 Flash on general-purpose workloads, but it is the right choice when you need low-latency, continuous audio handling rather than batch audio input to a standard text model.

Consumer Subscription Tiers (Not API Models)

The consumer Gemini product at gemini.google.com is offered through subscription tiers that are separate from the API pricing above. As of April 2026, per gemini.google/subscriptions, the consumer tiers are: free Gemini, Google AI Pro at $19.99 per month, and Google AI Ultra at $249.99 per month. Google AI Plus is also available at a lower price in some regions.

Google AI Pro includes Gemini 3.1 Pro access, Deep Research, Nano Banana Pro image generation, Veo 3.1 Fast video, 1,000 AI credits per month, and 2TB storage. Google AI Ultra adds Veo 3.1 full, Project Mariner (up to 10 parallel agents), 30TB storage, and YouTube Premium. The "Gemini Advanced" and "Google One AI Premium" brand names from earlier cycles have been retired - if you see them in documentation, treat them as historical references.

Consumer subscription tiers govern access to the gemini.google.com web app and mobile apps. API pricing is independent and billed separately through Google Cloud or Google AI Studio. For a direct comparison of the Google AI Pro consumer tier against ChatGPT Plus, see the Google AI Pro vs ChatGPT Plus comparison.

Use Gemini Toolbox to Manage Conversations Across All Three Models

Gemini Toolbox works at the gemini.google.com interface layer, which means it integrates with whichever Gemini model you have selected for a given conversation. Free Gemini users default to 3 Flash. Google AI Pro subscribers get Gemini 3.1 Pro. The extension does not care which model produced a message; it indexes every synced conversation into IndexedDB and surfaces full-text search across all of them.

Practical consequence: a power user with a mix of 3.1 Pro and 3 Flash conversations can search across both in a single query. Filter to "Gemini responses" to recover a specific answer regardless of which model produced it. Filter by date to track model selection over time. Export any conversation as TXT, Markdown, JSON, or PDF regardless of underlying model.

This is the right pairing for anyone serious about Gemini in 2026: choose the model per task on cost, quality, and latency grounds, then use Gemini Toolbox to keep the resulting conversation history searchable and exportable over time.

Frequently Asked Questions

What is the difference between Gemini 3.1 Pro and Gemini 3 Flash?

Gemini 3.1 Pro is Google's top-tier reasoning and agentic coding model, priced on the Standard tier at $2.00 / $4.00 per 1M input tokens (at or below 200K / above 200K prompts) and $12.00 / $18.00 per 1M output tokens. Gemini 3 Flash is the cost-efficient workhorse at $0.50 per 1M input tokens and $3.00 per 1M output tokens, with a 1M input context window like Pro. 3 Flash actually outperforms Gemini 3 Pro's earlier version on SWE-Bench Verified (78.0% vs 76.2%) while costing about a quarter as much on input.

Which Gemini 3 model has the longest context window?

All three Gemini 3 models share a 1 million token input context window per Google's official documentation at ai.google.dev. Context window is not a differentiator within the generation. Output budgets are also similar: Gemini 3.1 Pro allows 64K output tokens, 3 Flash allows 65,535, and 3.1 Flash-Lite allows 65,536.

How much does Gemini 3.1 Pro cost per 1 million tokens?

Gemini 3.1 Pro Standard-tier pricing is $2.00 per 1M input tokens for prompts at or below 200K and $4.00 per 1M input tokens for prompts above 200K; output is $12.00 per 1M tokens (at or below 200K) or $18.00 per 1M tokens (above 200K). Batch and Flex tiers apply a roughly 50% discount. Priority tier is roughly 1.75x to 2x Standard for latency-critical workloads.

Is there a Gemini Ultra API model in 2026?

No. "Ultra" refers only to the Google AI Ultra consumer subscription tier at $249.99 per month, not an API model. The three Gemini 3 API tiers in April 2026 are Gemini 3.1 Pro, Gemini 3 Flash, and Gemini 3.1 Flash-Lite. Any reference to "Gemini Ultra" as an API model is outdated.

When does Gemini 2.5 get retired?

Per Google's Gemini API deprecations page, Gemini 2.5 Pro is scheduled for AI Studio and Gemini API deprecation on June 17, 2026, and Vertex AI retirement on October 16, 2026. Gemini 2.5 Flash and 2.5 Flash-Lite follow the same timeline. Projects that currently target 2.5 should plan migration to Gemini 3 before mid-2026 and re-run their evaluation set on the new model before cutting traffic over.

Can I use Gemini 3.1 Flash-Lite for production chatbots?

Yes, if reasoning quality requirements are modest and you have verified performance on your own evaluation set. Flash-Lite at $0.25 per 1M input tokens and $1.50 per 1M output tokens is the cheapest tier in the generation and shines at classification, routing, and short structured responses. For nuanced conversational chat, most teams use Gemini 3 Flash as a middle ground that still scores within a few points of 3.1 Pro on SWE-Bench Verified.

Do all three Gemini 3 models support vision and audio?

Yes. All three Gemini 3 models accept text, images, audio, video, and PDF inputs natively. For real-time audio streaming specifically, Google ships a sibling gemini-3.1-flash-live-preview model that handles continuous audio instead of batch audio inputs.

How do Flex, Batch, and Priority tiers change Gemini 3 pricing?

Batch API runs requests asynchronously at roughly a 50% discount versus Standard. Flex is a best-effort synchronous tier at roughly a 50% discount. Priority is the latency-critical tier at roughly 1.75x to 2x Standard. Google's Flex and Priority inference announcement and the priority inference docs are the canonical sources. The multiplier is applied across the Gemini 3 generation.

How does Gemini 3.1 Pro compare to Claude Opus 4.6 and GPT-5 for coding?

On SWE-Bench Verified, the three frontier models are within roughly one percentage point of each other. Claude Opus 4.6 leads at 80.9%, Gemini 3.1 Pro is at 80.6%, and GPT-5.2 / 5.4 sit at roughly 80.0%. Gemini 3 Flash at 78.0% is surprisingly close at a fraction of the price. Practical model selection at the top tier is driven by ecosystem, tooling, pricing, and long-context capability more than raw SWE-Bench. For a cross-model view of the Claude side, see the Claude Toolbox home page.

Bottom Line

The Gemini 3 family in 2026 is a three-tier production lineup with a shared 1M-token input context window, full multimodal input support, and a published retirement path for Gemini 2.5 (June 17, 2026 in AI Studio, October 16, 2026 in Vertex AI). Gemini 3.1 Pro is the top-tier reasoning and agentic coding model. Gemini 3 Flash is the new default for most production workloads and is surprisingly strong on SWE-Bench Verified. Gemini 3.1 Flash-Lite is the cheapest tier for high-volume, latency-sensitive use. There is no "Gemini Ultra" API model; Ultra is only a consumer subscription tier name.

For most applications, the right architecture is a hybrid: 3.1 Flash-Lite for pre-processing and routing, 3 Flash for the main calls, 3.1 Pro for escalation. That structure minimizes cost while preserving quality on the calls that matter. Whichever model you end up on, Gemini Toolbox layers full-text search, filters, and four-format export on top of gemini.google.com so your model-agnostic conversation history stays findable and exportable.

Install Gemini Toolbox from the Chrome Web Store. For related reading, see the filter guide, the developer use cases, the researcher use cases, the Google AI Pro vs ChatGPT Plus comparison, the export formats guide, the ChatGPT Toolbox home page, and the Claude Toolbox home page.

Last updated: April 11, 2026