Gemini 3.1: Features, Benchmarks, Hands-On Tests, and More

Gemini 3.1 Pro correctly answered [Red Square]. The model identified both rules independently: colors toggle (Red ↔ Blue) and shapes cycle (Square → Circle → Triangle → Square). It then traced through the logic step by step, showing how Blue Triangle becomes Red (color toggle) and Square (next in shape cycle), which is exactly the kind of compositional reasoning this test targets.

Test 2: The disguised sequence

This test checks hypothesis elimination across multiple layers. We give the model two sequences and ask it to identify what the first sequence is (partition numbers from the OEIS) and work out the two transformations applied to produce the second.

Here are two sequences. The second was derived from the first in two separate steps. 
Identify the named mathematical sequence that Sequence A belongs to, and work out 
both transformations that were applied to produce Sequence B.

Sequence A: 1, 1, 2, 3, 5, 7, 11, 15, 22, 30, 42, 56, 77
Sequence B: 2, 3, 5, 8, 3, 9, 8, 1, 7, 9, 8, 7

Explain your reasoning step by step.

Gemini 3.1 Pro correctly identified Sequence A as partition numbers (A000041) and explained what partition numbers represent in number theory. It then systematically worked through both transformations: first summing consecutive pairs to generate an intermediate sequence, then calculating the digital root of each result. The model verified each step against Sequence B, showing the complete chain of reasoning from the original sequence to the final output.

Test 3: The broken clock network

This test targets constraint consistency checking. Six clocks are networked, each applying a fixed 20-minute offset. One clock is broken. The model has to trace both paths through the network and spot the contradiction.

Here is the prompt I used:

Six clocks (A, B, C, D, E, and F) are connected in a network. Each clock applies 
a fixed offset to the time it receives. A is the root and shows 12:00. You observe:

- B receives from A and shows 12:20
- C receives from A and shows 11:40
- D receives from B and shows 12:40
- E receives from C and shows 11:00
- F receives from both D and E and shows 13:00

There is exactly one broken clock in the network. Based on the pattern of offsets, 
identify which clock is broken, and give two possible answers for what it should 
actually show (one for each path through the network).

Explain your reasoning step by step.

Gemini 3.1 Pro correctly identified F as the broken clock and inferred two possible values for it: 13:00 via D’s path and 10:00 via E’s path. The model treated the right-hand path as a constant +20 minute offset and the left-hand path as an arithmetic sequence of −20, −40, then −60 minutes.

Hands-On With Gemini 3.1 Pro

Beyond abstract reasoning tests, I wanted to see how the model handles practical tasks that showcase its new features.

Animated SVG generation

Google made a big deal of code-based visual output in the launch, so I tested it directly with a simple brief and no template.

Here is the prompt I used:

Create an animated SVG loading spinner with three bouncing dots. Make it smooth, 
professional, and suitable for embedding on a website. Output only the SVG code.

Gemini 3.1 Pro returned clean SVG code with CSS animations. The output was a working three-dot loader with staggered bounce timing, exactly what was asked for. It rendered correctly in a browser on the first try, no adjustments needed. File size was tiny, and because it's vector-based code, it scales cleanly to any size.

This is one of those features that sounds gimmicky in a press release but turns out to be really practical. Lightweight, embeddable, infinitely scalable animated graphics from a text prompt is a solid tool for frontend prototyping or quick visual assets.

How Can I Access Gemini 3.1 Pro?

Gemini 3.1 Pro is currently in preview. Google has said it will reach general availability soon, after addressing feedback and making further improvements to agentic workflows.

Here are the main access options:

Gemini CLI

The Gemini CLI is an open-source terminal agent that gives the model direct access to your local environment. Install it using the following code:

npm install -g @google/gemini-cli
# Or run directly: npx @google/gemini-cli

The CLI uses a ReAct loop, meaning it can write code, run it, read errors, fix issues, and iterate on its own. With 3.1 Pro's improved terminal coding performance, this loop is noticeably more reliable. The free tier gives you 60 requests per minute and 1,000 requests per day.

Gemini API

The Gemini API gives developers direct programmatic access to Gemini 3.1 Pro. 

The model ID you want is: gemini-3.1-pro-preview

Here's some Python code to help you get started:

from google import genai

client = genai.Client()
response = client.models.generate_content(
    model="gemini-3.1-pro-preview",
    contents="Your prompt here"
)
print(response.text)

The pricing is the same as Gemini 3 Pro Preview.

The thinking_level parameter accepts low, medium, high, or max. Supported tools include Google Search, URL context, code execution, and file search. I'll cover the context window details in the comparison section below.

NotebookLM

NotebookLM is now powered by Gemini 3.1 Pro for Google AI Pro and Ultra subscribers. NotebookLM responds only based on documents you upload, making it a really useful research tool when you want the model to stay grounded in specific materials.

Consumer access

Google has started rolling Gemini 3.1 Pro out across its consumer and developer products, but it has not published a simple "plan X = model Y" mapping. Practically, you'll see 3.1 Pro in the Gemini app and API as it rolls out, with AI Ultra providing the broadest access.

Gemini 3.1 Pro vs. Claude Models

The February 2026 releases from Google and Anthropic have created a really interesting set of trade-offs. These aren't situations where one model clearly wins. The right choice depends heavily on what you're building.

The pricing gap is worth dwelling on. Gemini 3.1 Pro is much cheaper on both input and output than Claude Opus 4.6. If you're running high-volume API calls, that's not a small difference.

Choose Gemini 3.1 Pro when:

• Abstract reasoning and scientific analysis are the priority
• You need robust, natively multimodal support for video and audio in the same model
• You want the full 1M context window in its stable, non-beta form
• Cost efficiency matters, especially at scale

Choose Claude Opus 4.6 when:

• You need the full 128K output tokens (Gemini caps at 64K)
• Multi-agent orchestration is central to your workflow (Agent Teams is a real differentiator)
• Computer use via GUI is important
• You're doing knowledge-intensive work where deep research quality matters

Choose Claude Sonnet 4.6 when:

• Knowledge work, document analysis, or financial analysis are the main tasks
• You need near-flagship performance at a lower price point
• You're already using Anthropic tools and Sonnet is your default

Gemini 3.1 Pro Use Cases

Based on the benchmarks and hands-on testing, here are the areas where Gemini 3.1 Pro is particularly well-suited:

• Scientific research and analysis: Strong GPQA Diamond performance plus 1M context window makes it practical for literature review, hypothesis generation, and synthesizing across multiple papers at once.
• Autonomous research agents: Improved agentic benchmarks translate to real-world multi-step tasks like gathering information from multiple sources, verifying facts, and producing structured reports with minimal supervision.

• Codebase analysis and refactoring: A large context window plus improved reasoning handles tasks like identifying architectural inconsistencies across modules or tracing bugs through multiple files.

• Multimodal content analysis: Native video and audio support enables analyzing recorded meetings, extracting insights from lecture videos, or processing podcasts without preprocessing.

• Cost-sensitive production deployments: At roughly half the cost of Claude Opus 4.6, it makes sense for high-volume inference where reasoning quality matters but budget is a constraint.

• Prototyping and visual assets: Code-based animated output generates loading spinners, animated charts, or interactive dashboards from text prompts that you can embed directly.

Final Thoughts

Gemini 3.1 Pro is a good example of where these models are going. Less focus on new input types, more focus on better reasoning, more reliable agents, and handling longer contexts. Even though it's just a ".1" release, the benchmark improvements and the Deep Think connection make it feel like a bigger step forward in how these systems think.

For teams building real products, there's no single "best" model. Gemini 3.1 Pro works well for scientific reasoning, research agents, and analyzing large codebases, especially when you consider the price and video support. Claude is still better for knowledge work and using computers through the screen, and GPT-5.3-Codex still wins on some coding tests.

The interesting question is what happens when it leaves preview. Google has said they're working on agent improvements before the full release. If those ship alongside the current reasoning upgrades, the gap between research models like Deep Think and everyday models will get smaller. For now, it's a good time to try different models and build systems that can use what each one does best.

To get started with Google's AI tools, check out our Introduction to Google Gemini course. For working with the API in Python, our Working with the Gemini API tutorial covers the essentials.