Gemini 3.1 Pro is the model Google reaches for when it wants to demonstrate the strongest artificial intelligence in its stable, and six months after launch it remains the company’s reigning flagship. Released in preview on 19 February 2026, it was framed in the official launch announcement as “a smarter model for your most complex tasks” — the option you pick when reasoning depth matters more than raw speed or cost.
That flagship status carries an unusual 2026 twist. Google’s fast-cycling Flash line has shipped three new models since May, while the Gemini 3.5 Pro successor promised at I/O 2026 is still nowhere to be seen. The result is a line-up in which Gemini 3.1 Pro keeps the crown even as cheaper “workhorse” siblings overtake it on selected agentic benchmarks.
This guide gives you the full, sourced picture of Gemini 3.1 Pro: its specifications, benchmark scores, API pricing, every access route, and honest advice on when a Flash model serves you better. It lives in our AI models and tools hub, alongside our wider AI comparison of the leading assistants.
Every price, date and score below comes from a referenced source — Google’s own blog, the DeepMind model cards, the Gemini API documentation and independent reporting. Where Google has not published a number, we say so rather than guessing. And if you only take one thing away, make it this: the model-picking question in 2026 is no longer “which vendor?” but “which tier of which vendor’s range?”. Let us start with what actually sits behind the name.
Table of contents
- What Is Gemini 3.1 Pro?
- Gemini 3.1 Pro Specifications and Capabilities
- Gemini 3.1 Pro Benchmark Performance
- Gemini 3.1 Pro API Pricing Explained
- How to Access Google’s Flagship Model
- The Flash Line Racing Past the Flagship
- Gemini 3.1 Pro vs Gemini 3.7 Flash
- The Missing Gemini 3.5 Pro and the Road to Gemini 4
- Consumer Plans: Where the Flagship Fits
- Choosing Between Pro and Flash for Real Workloads
- Business Adoption and Agentic Workflows
- The Wider Gemini Ecosystem in 2026
- Frequently Asked Questions About Gemini 3.1 Pro
- References
What Is Gemini 3.1 Pro?
Gemini 3.1 Pro is Google’s flagship large language model as of August 2026 — the most capable general-purpose model the company currently sells. A common misconception is worth clearing up immediately: there is no single model called “Gemini 3”. The Gemini 3 generation is a family, and the 3.1 Pro model sits at the top of its reasoning tier, above the fast, inexpensive Flash models.
The February 2026 Launch of Gemini 3.1 Pro
Google released Gemini 3.1 Pro in preview on 19 February 2026, describing it as “a smarter model for your most complex tasks”. The launch was emphatic: Google’s own materials showed the model leading on 13 of 16 headline benchmarks at release, and the company highlighted that its 77.1% ARC-AGI-2 score was more than double that of its predecessor, Gemini 3 Pro. Independent benchmark analysis at the time echoed the framing — this was less an incremental point release than a reset of the range’s ceiling.
Half a year on, no Google model has replaced it at the top of the range. That longevity — an eternity in the 2026 release cycle — is precisely why this complete guide still matters in August.
Why Gemini 3.1 Pro Still Says “Preview”
Curiously, the flagship has never dropped its preview label. As of August 2026 the Gemini API still serves it under the model ID gemini-3.1-pro-preview, per Google’s own changelog. There is also a second endpoint, gemini-3.1-pro-preview-customtools, which prioritises custom tools for developers building agentic systems.
In practice the preview label has not stopped enterprise adoption — the model ships in Vertex AI and Gemini Enterprise — but it signals that Google always intended a successor to arrive sooner than it has.
From Gemini 3 Pro to Gemini 3.1 Pro
The predecessor, Gemini 3 Pro, launched on 18 November 2025 and was available across the Gemini app, Google Search, AI Studio and Vertex AI on day one. Its API life was short: Google shut down the gemini-3-pro-preview endpoint on 9 March 2026, and requests now redirect to gemini-3.1-pro-preview.
That redirect tells you how Google views the relationship: the newer model is not a side branch but the direct, drop-in upgrade of the November 2025 flagship — same family, sharper reasoning.
Gemini 3.1 Pro Specifications and Capabilities
Under the bonnet, Gemini 3.1 Pro pairs a very large context window with genuinely broad multimodality. The DeepMind model card is the authoritative source here, and its headline figures explain why the model remains the default choice for long-document and multi-format work.
Model ID: gemini-3.1-pro-preview (plus a customtools variant)
Input context: 1 million tokens · Output limit: 64K tokens
Inputs: text, images, audio, video and code repositories
Released: 19 February 2026 · Status: preview, flagship of the range
Gemini 3.1 Pro Context Window and Output Limits
The model accepts up to 1 million tokens of input context and can generate up to 64K tokens of output, according to the DeepMind model card. A million tokens comfortably covers entire codebases, lengthy contract bundles or hours of transcribed audio in a single request.
The 64K output ceiling matters just as much for practical work: long structured reports, full test suites or multi-file refactors can come back in one pass rather than being stitched together.
Multimodal Inputs Gemini 3.1 Pro Accepts
The model card lists text, images, audio, video and code repositories as supported inputs. That last item is notable — the flagship treats a repository as a first-class input type, which is a large part of why Google positions it for complex software engineering rather than quick completions.
For mixed-media workloads — say, analysing a recorded meeting against a slide deck and a spec document — this breadth means one model call instead of a pipeline of converters.
The Custom Tools Endpoint
The gemini-3.1-pro-preview-customtools endpoint deserves its own mention. Google’s changelog describes it as a variant that prioritises custom tools, aimed at developers whose agents depend on reliably invoking their own functions rather than Google’s built-ins.
If you are building tool-heavy agents on Gemini 3.1 Pro, this endpoint is the one to benchmark first; it exists precisely because tool-call discipline is where generalist models most often wobble.
Google splitting its flagship into two endpoints is itself informative: agentic tool use has become important enough to warrant dedicated serving behaviour rather than a prompt-level workaround.
Gemini 3.1 Pro Benchmark Performance
Benchmarks are where Gemini 3.1 Pro earned its flagship badge. The DeepMind model card publishes a strong set of scores across reasoning, coding, knowledge and long-context retrieval, and Google’s launch materials showed the model leading on 13 of 16 headline benchmarks at release.
Benchmarks are imperfect proxies, of course, and every vendor curates its chart. But the model card figures below are specific, published and checkable — a firmer basis for a buying decision than launch-day adjectives.
The chart below visualises the five percentage-based headline scores from the model card, each discussed with its figure in the subsections that follow.
Gemini 3.1 Pro on ARC-AGI-2 and GPQA Diamond
The score Google shouted about loudest is ARC-AGI-2, the abstract-reasoning benchmark designed to resist memorisation. The flagship posts 77.1% — which Google says is more than double the score of Gemini 3 Pro. On GPQA Diamond, the graduate-level science question set, the model reaches 94.3%.
Together these two figures are the core of the flagship claim: novel-problem reasoning and expert-level knowledge, both at the top of what Google has published for any of its models.
Coding Scores: SWE-Bench Verified and LiveCodeBench Pro
On SWE-Bench Verified, the real-world software-engineering benchmark built from actual GitHub issues, Gemini 3.1 Pro scores 80.6%. On LiveCodeBench Pro, the competitive-programming leaderboard, it holds a rating of 2887 Elo.
For working developers the SWE-Bench figure is the meaningful one: it measures whether a model can read an existing codebase, locate a bug and produce a passing patch — exactly the loop that coding agents run all day.
Long-Context and Frontier Results for Gemini 3.1 Pro
Long-context retrieval is where the million-token window has to prove itself. The flagship scores 84.9% on MRCR v2 at 128k context, per the model card. On Humanity’s Last Exam — the deliberately brutal frontier-knowledge test — it reaches 51.4% when allowed search and code tools.
These are strong numbers, though as we will see later, the newest Flash workhorse has since overtaken the flagship on long-context retrieval specifically — one of several reasons the “best model” question now needs a nuanced answer.
Gemini 3.1 Pro API Pricing Explained
Flagship capability commands flagship pricing. Gemini 3.1 Pro is the most expensive text model in the Gemini API’s current line-up, with rates that step up beyond 200k tokens of context. The table below lays out the official Gemini API pricing for the flagship and its Flash siblings.
| Model | Input per 1M tokens | Output per 1M tokens | Notes |
|---|---|---|---|
| Gemini 3.1 Pro Preview | $2.00 (≤200k) / $4.00 (>200k) | $12.00 (≤200k) / $18.00 (>200k) | Flagship of the range |
| Gemini 3.5 Flash | $1.50 | $9.00 | Backs gemini-flash-latest |
| Gemini 3.7 Flash | $0.75 to 31 Dec 2026, then $1.50 | $3.75 to 31 Dec 2026, then $7.50 | Introductory pricing at launch |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | Cheapest 3.5-generation tier |
Gemini 3.1 Pro Token Rates
The flagship’s Preview rate card reads: $2.00 per million input tokens up to 200k of context, rising to $4.00 per million above that threshold. Output costs $12.00 per million tokens up to 200k, rising to $18.00 beyond it.
The 200k step matters for anyone planning to exploit the full million-token window — a maximum-context request is billed at double the input rate and 50% more on output, so long-context jobs deserve deliberate budgeting.
Batch Discounts and the Free Tier
Two levers soften the bill. The Gemini Batch API offers a 50% discount across most models for workloads that can tolerate asynchronous processing — an obvious fit for overnight document analysis or bulk classification on Gemini 3.1 Pro.
A free tier also remains available with limited model access, with the caveat that content on the free tier is used for product improvement. For anything commercially sensitive, paid tiers are the only sensible route.
Comparing Gemini 3.1 Pro Pricing With the Flash Line
The gulf between the flagship and its siblings is stark. At standard context, flagship input costs $2.00 per million tokens against $0.75 for Gemini 3.7 Flash on introductory pricing — and $0.30 for Gemini 3.5 Flash-Lite. The bars below are proportional to those stated input rates.
On output the ratio is even wider: $12.00 per million for the flagship against $3.75 for Gemini 3.7 Flash until the end of 2026. That input-price spread — from $2.00 down to $0.30 per million tokens across one vendor’s line-up — is exactly why model selection has become a genuine architectural decision.
How to Access Google's Flagship Model
Google distributes its flagship unusually widely for a preview-labelled model. Per the launch post, Gemini 3.1 Pro is available to consumers in the Gemini app and NotebookLM, and to developers through the Gemini API, Google AI Studio, Gemini CLI, Android Studio, Google Antigravity, Vertex AI and Gemini Enterprise.
Gemini App and NotebookLM
In the consumer Gemini app and in NotebookLM, Gemini 3.1 Pro is reserved for Google AI Pro and Ultra plan subscribers. Free-tier app users do not get the flagship; the paid gate is how Google manages demand for its most compute-hungry model.
NotebookLM access is a quiet highlight: pointing the flagship’s million-token context at a notebook full of sources turns it into a formidable research assistant. Upload a quarter’s worth of board papers or an entire literature review, and the model can cross-reference the lot in a single grounded conversation.
Google AI Studio and the Gemini API
Developers can try Gemini 3.1 Pro in Google AI Studio and call it through the Gemini API under the gemini-3.1-pro-preview model ID. Tool-centric builders should also evaluate the customtools variant discussed earlier.
One recent API change to note: on 21 July 2026 Google deprecated the temperature, top_p and top_k sampling parameters across the Gemini API, so freshly written integrations should not rely on them.
Vertex AI, Gemini CLI and Developer Tooling
For production and enterprise work, Gemini 3.1 Pro runs on Vertex AI and in Gemini Enterprise — the route most regulated organisations will take. Terminal-first developers get it in the Gemini CLI, Android developers inside Android Studio, and agentic-coding teams within Google Antigravity.
That breadth means the flagship can sit in every stage of a delivery pipeline, from a developer’s terminal to a governed cloud deployment, without switching model families.
The Flash Line Racing Past the Flagship
Here is the strange shape of Google’s 2026: while the flagship has stood still since February, the Flash “workhorse” line has been sprinting. Three Flash generations have shipped since May — several of them beating Gemini 3.1 Pro on specific agentic benchmarks at a fraction of its price. The timeline below, drawn from Google’s changelog and announcements, shows the cadence.
| Date | Release |
|---|---|
| 18 Nov 2025 | Gemini 3 Pro launches across the Gemini app, Google Search, AI Studio and Vertex AI |
| 19 Feb 2026 | Gemini 3.1 Pro released in preview — the current flagship |
| 9 Mar 2026 | gemini-3-pro-preview endpoint shut down; redirects to gemini-3.1-pro-preview |
| 19 May 2026 | Gemini 3.5 Flash goes GA at Google I/O; Gemini 3.5 Pro promised “next month” |
| 1 Jun 2026 | Gemini 2.0 models shut down |
| 30 Jun 2026 | gemini-omni-flash-preview launches for video generation |
| 21 Jul 2026 | Gemini 3.6 Flash GA, Gemini 3.5 Flash-Lite GA, Gemini 3.5 Flash Cyber pilot |
| 13 Aug 2026 | Gemini 3.7 Flash GA — three weeks after 3.6 Flash |
Gemini 3.5 Flash: The I/O 2026 Agentic Star
At I/O 2026 in May, Google called Gemini 3.5 Flash its strongest agentic and coding model to date — pointed language, given the flagship’s positioning. The stated scores: 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, 83.6% on MCP Atlas and 84.2% on CharXiv, outperforming Gemini 3.1 Pro on key agentic benchmarks per Google Cloud’s own blog.
Gemini 3.5 Flash now backs the gemini-flash-latest alias, at $1.50 per million input tokens and $9.00 per million output.
The 21 July Triple Release
On 21 July 2026 Google shipped three models in one day: Gemini 3.6 Flash (GA), Gemini 3.5 Flash-Lite (GA) and Gemini 3.5 Flash Cyber — a cybersecurity variant restricted to governments and authorised partners in a pilot programme.
Gemini 3.6 Flash’s headline was efficiency, cutting token consumption by up to 17% versus its predecessor for coding and agentic planning. TechCrunch’s coverage that day carried the sharper story in its headline: three new models, but no Gemini 3.5 Pro.
Gemini 3.7 Flash: Three Weeks Later
Just three weeks on, Gemini 3.7 Flash went GA on 13 August 2026, billed as Google’s “most intelligent workhorse model yet for coding and agents”. Its model card matches the flagship’s shape — 1M-token input, 64K output — with a March 2026 knowledge cutoff, 85.8% on Terminal-bench 2.1, 47.9% on OSWorld-2.0 computer use and 90.7% on Harvey LAB-AA legal workflows.
Google’s launch figures against 3.6 Flash, charted below, show why “workhorse” undersells it. It also posted 1588 Elo on WebDev Arena, up from 3.6 Flash’s 1538.
Introductory pricing is aggressive: $0.75 per million input tokens and $3.75 per million output until 31 December 2026 — half 3.6 Flash’s launch price — rising to $1.50 and $7.50 from 1 January 2027.
Availability was equally aggressive on day one: Google AI Studio, the Gemini API, Android Studio, Google Antigravity, the Gemini Enterprise Agent Platform and the Gemini Enterprise app, alongside the consumer Spark rollout.
Gemini 3.1 Pro vs Gemini 3.7 Flash
So which should you actually use — the six-month-old flagship or the day-old workhorse? The honest answer is that Gemini 3.1 Pro and Gemini 3.7 Flash are optimised for different jobs, and the published numbers make the split fairly clean.
| Attribute | Gemini 3.1 Pro | Gemini 3.7 Flash |
|---|---|---|
| Launch | 19 Feb 2026 (preview) | 13 Aug 2026 (GA) |
| Positioning | “Smarter model for your most complex tasks” | “Most intelligent workhorse model yet for coding and agents” |
| Input context | 1M tokens | 1M tokens |
| Output limit | 64K tokens | 64K tokens |
| Long context at 128k | 84.9% (MRCR v2) | 97.0% (GDM-MRCR v2) |
| Input price per 1M | $2.00–$4.00 | $0.75 intro, $1.50 from 2027 |
| Output price per 1M | $12.00–$18.00 | $3.75 intro, $7.50 from 2027 |
| Consumer access | Gemini app and NotebookLM (AI Pro/Ultra) | Gemini app via Spark (AI Pro/Ultra) |
Where Gemini 3.1 Pro Still Wins
The flagship’s case rests on frontier reasoning. Its 77.1% ARC-AGI-2, 94.3% GPQA Diamond, 51.4% Humanity’s Last Exam and 2887 LiveCodeBench Pro Elo are the strongest such figures Google has published for any model, and the Flash model cards simply do not report comparable numbers on those tests.
For genuinely hard problems — novel research questions, gnarly architectural decisions, adversarial legal analysis — the flagship remains the model Google itself tells you to reach for.
Where Gemini 3.7 Flash Pulls Ahead
On long-context retrieval at 128k, the workhorse now beats the flagship on the stated card figures: 97.0% on GDM-MRCR v2 versus 84.9% on MRCR v2 for Gemini 3.1 Pro. Gemini 3.7 Flash also brings a fresher March 2026 knowledge cutoff, an 85.8% Terminal-bench 2.1 score for terminal-driven agents, and 47.9% on OSWorld-2.0 computer use.
Add the earlier I/O finding that Gemini 3.5 Flash already outperformed the flagship on key agentic benchmarks, and the pattern is clear: for agent loops, the Flash line is no longer the compromise option.
The Price-Performance Verdict
At introductory rates, Gemini 3.7 Flash undercuts Gemini 3.1 Pro by 62.5% on input ($0.75 versus $2.00) and 69% on output ($3.75 versus $12.00) at standard context. Unless your workload demonstrably needs the flagship’s reasoning ceiling, the workhorse wins the invoice argument comfortably.
Our rule of thumb: prototype on Gemini 3.7 Flash, escalate the failures to Gemini 3.1 Pro, and let the failure rate tell you which model each task truly needs.
How to Run a Fair Pilot
Benchmark tables never settle the argument on their own; your data does. Build a small evaluation set from real tasks — twenty representative tickets, documents or prompts — and run it through both models with identical instructions. Score the outputs blind, then weigh the quality gap against the price gap from the table above.
Teams are often surprised in both directions: some discover the workhorse clears their bar comfortably, others find the flagship’s reasoning margin is worth every cent on their hardest ten per cent of cases.
The Missing Gemini 3.5 Pro and the Road to Gemini 4
No guide to Gemini 3.1 Pro is complete without addressing the elephant in the roadmap: its successor is late, publicly and repeatedly. Understanding the delay explains why the February flagship still tops the range in August.
The I/O 2026 Promise
At I/O 2026 in May, Google announced Gemini 3.5 Pro as “coming next month”, per the Google Cloud blog’s own round-up of the event. That month came and went. As of 14 August 2026 the model remains unreleased, and Gemini 3.1 Pro continues as the flagship by default.
What Google Says About the Delay
Google DeepMind product lead Logan Kilpatrick said in July 2026 that Gemini 3.5 Pro is being evaluated with partners and is expected to “land soon”. Independent reporting has been less soothing: Bloomberg reported on 13 August 2026 that Google’s top model is still delayed, having earlier reported internal setbacks in meeting performance benchmarks.
The subtext is straightforward: a successor to Gemini 3.1 Pro has to beat a very high bar, and Google appears unwilling to ship one that does not.
The Gemini 4 Pre-Training Signal
Buried in Kilpatrick’s July comments was a bigger tell: the team has begun “its most ambitious pre-training run yet for Gemini 4”. Whatever happens with 3.5 Pro, the next full generation is already in the oven.
For planners, the practical takeaway is that Gemini 3.1 Pro is not a stopgap to design around — it is the stable flagship for the foreseeable near term, with an upgrade path that will arrive on Google’s timetable, not yours.
There is even a silver lining for buyers: the delay means anything you build on the current flagship today will not be obsolete next week. Bloomberg’s reporting of internal benchmark setbacks reads as Google holding the release to a standard — and a successor that clears that bar decisively will be worth migrating to when it finally lands.
Consumer Plans: Where the Flagship Fits
Google restructured its consumer AI subscriptions at I/O 2026, and the changes decide who gets flagship access in the Gemini app. The line-up now runs from a $4.99 entry plan to a $199.99 top tier.
| Plan | Monthly price | Flagship-relevant highlights |
|---|---|---|
| Google AI Plus | $4.99 (cut from $7.99 on 8 Jun 2026) | Storage doubled to 400 GB; no stated flagship app access |
| Google AI Pro | $19.99 | Gemini 3.1 Pro in the Gemini app and NotebookLM; Spark agent |
| Google AI Ultra | $99.99 | 5x Pro usage limits; Gemini 3 Deep Think reasoning mode |
| Google AI Ultra (top tier) | $199.99 | 20x Pro usage limits; replaces the old $249.99 plan |
AI Plus, AI Pro and AI Ultra Compared
The restructure moved in one direction: down in price. Google AI Ultra now starts at $99.99 per month with 5x Pro usage limits, with a $199.99 tier offering 20x — down from the previous $249.99 price. At the bottom, AI Plus was cut from $7.99 to $4.99 on 8 June 2026, with storage doubled to 400 GB.
The mid-range AI Pro plan holds at $19.99 per month and remains the pivotal tier for most readers of this guide.
Which Plan Unlocks Gemini 3.1 Pro?
Per Google’s launch post, flagship access in the Gemini app and NotebookLM is reserved for Google AI Pro and Ultra plans. In other words, $19.99 per month is the entry ticket to the flagship as a consumer; the Ultra tiers buy you multiplied usage limits rather than a different base model.
Ultra does add one exclusive: Gemini 3 Deep Think, the more capable reasoning mode upgraded in February 2026 and targeted at science, research and engineering work.
Spark and the Subscriber Extras
Pro and Ultra subscribers also get Spark, the 24/7 personal AI agent announced at I/O 2026 that lives inside the Gemini app. Google began rolling Gemini 3.7 Flash into Spark on its 13 August launch day across 160+ countries.
The pairing is telling: your Spark agent runs on the Flash workhorse, while Gemini 3.1 Pro handles the deep-thinking conversations you open yourself. It previews how Google expects subscribers to work — an always-on Flash agent covering the routine, with the flagship on call for the questions that need depth.
Choosing Between Pro and Flash for Real Workloads
Model selection across the Gemini range is now a genuine engineering decision with real budget consequences. Here is the practical decision framework we use, grounded entirely in the published figures above.
When Gemini 3.1 Pro Is the Right Choice
Reach for Gemini 3.1 Pro when the task is hard enough that failure costs more than tokens: multi-step reasoning over novel problems, research synthesis across a million tokens of sources, complex debugging where the 80.6% SWE-Bench Verified score earns its premium, or expert-domain questions where 94.3% GPQA-grade knowledge matters.
The customtools endpoint extends the same logic to high-stakes agents: when a mis-fired tool call is expensive, the flagship’s reliability is the cheaper option overall.
When a Flash Model Serves You Better
High-volume, latency-sensitive and cost-sensitive workloads belong on the Flash line: Gemini 3.7 Flash for coding agents and computer use, Gemini 3.5 Flash-Lite at $0.30 per million input tokens for classification and extraction at scale. The lightweight tier has been usable in surprising places for a while — see our hands-on look at Gemini 3.1 Flash Lite running in the browser from May.
If your agent benchmarks well on a Flash model, the flagship premium buys you nothing but a bigger invoice.
A Worked Cost Example on Gemini 3.1 Pro
Concrete numbers make the trade-off vivid. Take a document-analysis job of 100k input tokens and 10k output tokens per run — comfortably under the 200k threshold. On Gemini 3.1 Pro that costs $0.20 for input (0.1M at $2.00 per million) plus $0.12 for output (0.01M at $12.00): $0.32 per run, or $3,200 across 10,000 monthly runs.
The identical job on Gemini 3.7 Flash at introductory rates works out at roughly $0.11 per run ($0.075 input plus $0.0375 output), or about $1,125 per month. The Batch API’s 50% discount halves either bill for asynchronous work. Whether the flagship’s extra accuracy justifies the difference is a question only your own evaluation set can answer.
Migration Notes for Older Gemini Models
Housekeeping matters in a range that moves this fast. Gemini 2.0 models (gemini-2.0-flash and gemini-2.0-flash-lite) were shut down on 1 June 2026, while Gemini 2.5 Pro, 2.5 Flash and 2.5 Flash-Lite remain available as stable models alongside the 3.x line.
Anything still calling gemini-3-pro-preview has been silently redirected to Gemini 3.1 Pro since 9 March 2026 — worth verifying, since behaviour and pricing differ. And remember the 21 July 2026 deprecation of temperature, top_p and top_k when porting older code.
Business Adoption and Agentic Workflows
The commercial context around Gemini 3.1 Pro is enormous and growing, which is why model choice within the range — rather than whether to use Gemini at all — has become the real question for most organisations.
Gemini by the Numbers in 2026
Alphabet’s Q2 2026 earnings on 22 July reported the Gemini app at 950 million monthly active users, with Gemini models processing 22 billion API tokens per minute — up from 16 billion the previous quarter. Nearly 90% of the Fortune 100 use Gemini Enterprise, and more than 9 million developers build with Google’s models monthly.
Those numbers describe a platform, not a product — and the flagship sits at the top of it as the quality ceiling everything else is measured against. The token-throughput figure is the one to watch: growth from 16 billion to 22 billion tokens per minute in a single quarter is a 37.5% jump, and that is demand, not marketing.
Building Agentic Systems on Gemini 3.1 Pro
The architecture pattern we see working in 2026 is layered: Gemini 3.1 Pro as the planner and reviewer, Flash models as the executors. The flagship’s customtools endpoint anchors the tool-calling backbone, while cheap Flash calls handle the high-frequency steps. It is the same intelligent automation logic we apply in client systems: spend the expensive tokens only where judgement lives.
Gemini Enterprise, Vertex AI and the Gemini CLI give that pattern a deployment surface from prototype through to governed production.
Getting Strategic Help With Deployment
Choosing between a preview-labelled flagship, a three-week release cadence and a delayed successor is exactly the kind of moving target that stalls AI programmes. If your team needs a structured path — model selection, cost modelling, rollout governance — our AI strategy practice builds precisely these roadmaps around the Gemini range and its competitors.
The earlier you separate “flagship tasks” from “workhorse tasks” in your architecture, the less each Google release forces you to re-plan.
The Wider Gemini Ecosystem in 2026
Gemini 3.1 Pro headlines a much broader 2026 portfolio, and knowing the neighbouring lines prevents expensive mismatches — using a reasoning flagship where a specialist model was the right call.
Gemini Omni and Video Generation
Gemini Omni, announced at I/O 2026, is a separate multimodal video-generation and editing line accepting text, audio, image and video inputs — not a flagship reasoning model, despite occasional confusion over the name. Its first API release, gemini-omni-flash-preview, landed on 30 June 2026, producing 3–10 second 720p videos from text or still images. We covered that launch in detail in our Gemini Omni Flash preview report.
Image Models and Deep Think
Image generation went GA on 28 May 2026 with gemini-3.1-flash-image — better known as Nano Banana 2 — and gemini-3-pro-image, Nano Banana Pro. For pure reasoning beyond even the flagship, Gemini 3 Deep Think remains the Ultra-subscriber option, upgraded in February 2026 for science, research and engineering.
The specialist variants keep multiplying at the edges too: witness Gemini 3.5 Flash Cyber, piloted with governments and authorised partners for cybersecurity work.
Where Gemini 3.1 Pro Sits in the Stack
Seen whole, the 2026 stack has a clear shape: the flagship at the top for complex reasoning; the 3.5/3.6/3.7 Flash line as fast, cheap workhorses; Flash-Lite for bulk economy; and Omni, image and Cyber models as specialists. One flagship, many tools — and a Gemini 4 pre-training run already underway behind the curtain.
Frequently Asked Questions About Gemini 3.1 Pro
Is Gemini 3.1 Pro free to use?
Not meaningfully. In the Gemini app and NotebookLM, Gemini 3.1 Pro requires a Google AI Pro ($19.99/month) or Ultra plan. Developers pay per token via the Gemini API — from $2.00 per million input tokens — though the API’s free tier offers limited model access, with content used for product improvement.
What is the difference between Gemini 3 Pro and Gemini 3.1 Pro?
Gemini 3 Pro was the flagship launched on 18 November 2025; Gemini 3.1 Pro replaced it on 19 February 2026 with substantially stronger reasoning — Google says its 77.1% ARC-AGI-2 score is more than double its predecessor’s. The old gemini-3-pro-preview endpoint was shut down on 9 March 2026 and now redirects to the newer model.
Does Gemini 3.1 Pro have a 1 million token context window?
Yes. The DeepMind model card lists a 1 million-token input context window and a 64K-token output limit, with support for text, images, audio, video and code repositories as inputs.
When will Gemini 3.5 Pro replace Gemini 3.1 Pro?
Nobody outside Google knows. Gemini 3.5 Pro was announced at I/O 2026 as “coming next month” but remains unreleased as of 14 August 2026. Google’s Logan Kilpatrick says it is being evaluated with partners and will “land soon”; Bloomberg reports it has faced internal setbacks meeting performance benchmarks.
Is Gemini 3.7 Flash better than Gemini 3.1 Pro?
On some published measures, yes: Gemini 3.7 Flash scores 97.0% on GDM-MRCR v2 long context versus the flagship’s 84.9% on MRCR v2, and the Flash line leads on several agentic benchmarks. But the flagship’s ARC-AGI-2, GPQA Diamond and Humanity’s Last Exam scores remain unmatched in Google’s published figures — for the hardest reasoning, Gemini 3.1 Pro still leads.
Which older Gemini models were retired in 2026?
Google shut down the Gemini 2.0 models (gemini-2.0-flash and gemini-2.0-flash-lite) on 1 June 2026, and closed the gemini-3-pro-preview endpoint on 9 March 2026, which now redirects to the current flagship. The Gemini 2.5 Pro, 2.5 Flash and 2.5 Flash-Lite models remain available as stable options alongside the 3.x line.
References
Gemini 3.1 Pro: A smarter model for your most complex tasks (Google, 19 Feb 2026)
Gemini 3.1 Pro Model Card — Google DeepMind
Introducing Gemini 3.7 Flash: our most intelligent workhorse model (Google, 13 Aug 2026)
Gemini 3.7 Flash Model Card — Google DeepMind
Gemini API release notes and changelog — Google AI for Developers
Gemini API pricing — Google AI for Developers
Innovations from Google I/O 26 on Google Cloud — Google Cloud Blog
Google releases three new Gemini models — but no 3.5 Pro — TechCrunch
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
