Gemini 3.1 Pro Review: Google Finally Built the Assistant That Beats Its Search Bar

What it’s actually good at

The reasoning jump is the headline, and it holds up in daily use. On ARC-AGI-2, a benchmark that evaluates a model’s ability to solve entirely new logic patterns, 3.1 Pro achieved a verified score of 77.1%, more than double the reasoning performance of 3 Pro. That’s not a marginal update; that’s the kind of number you see when a lab has actually cracked something. And it shows up on other tests too: it recorded 94.3% on GPQA Diamond, a graduate-level science test, which, according to PCMag, is the highest score ever reported on that benchmark. On Humanity’s Last Exam, Gemini 3.1 Pro scored 44.4% without tools, versus 37.5% for Gemini 3 Pro, 40.0% for Claude Opus 4.6, and 34.5% for GPT-5.2.

Translated out of benchmark-speak: it holds the thread on multi-step problems that used to trip Gemini up. Ask it to plan a migration, work through a five-way scheduling constraint, or synthesize a stack of research papers into a single argument, and it stays coherent to the end. That’s the failure mode I used to complain about with 2.5 Pro. It’s mostly gone.

The other big shift is invisible, which is the point. Gemini 3.1 Pro always applies what Google calls “dynamic thinking,” automatic chain-of-thought reasoning that scales in depth based on the complexity of the task. A simple factual question gets a quick answer. A multi-step research synthesis gets extended internal deliberation before the model responds. That’s a real UX win. You stop toggling between “fast” and “reasoning” modes because the model just does the right thing. And for developers, this behavior is now configurable through a new thinking_level parameter with four settings: low, medium (new in 3.1) and higher, so you keep the override when you need it.

Speed is the other quiet win. Gemini 3.1 Pro Preview generates output at 133.2 tokens per second (based on Google’s API), well above average compared to other reasoning models in a similar price tier (median: 69.9 t/s). Once the model gets going, it’s noticeably faster than Claude Opus for the same job.

Multimodal is where Gemini has always outpunched the field, and 3.1 Pro widens the gap. It can comprehend vast datasets and challenging problems from massively multimodal information sources, including text, audio, images, video, and entire code repositories, with a token context window of up to 1M. That 1M window isn’t a party trick anymore. It works. I’ve fed it whole codebases and 400-page contract PDFs and it still picks out specific clauses on page 380 without losing the plot.

Agentic and coding work landed better than I expected too. On the SWE-Bench Verified benchmark, which tests the model’s ability to resolve real-world software issues pulled from open-source Python repositories, Gemini 3.1 Pro achieved a pass rate of 80.6%. That metric requires the AI agent to comprehend an issue description, navigate a codebase, devise an implementation plan, write the corrective code, and verify the fix without introducing secondary errors. It’s competitive with the top of the field, and it’s why Google can plausibly ship things like Antigravity and Jules on top of the same model.

Where it lets you down

Two real gripes. First, it’s not universally best-in-class. Google’s marketing suggests it is, but the independent data is more nuanced.

GDPval-AA measures enterprise task performance (finance, legal, etc.). Gemini 3.1 Pro scored 1317, trailing Sonnet 4.6 (1633) and Opus 4.6 (1606) by nearly 300 points. Artificial Analysis independently confirmed that Gemini “improved but did not take the lead.” If your job is regulated-industry document drudgery, contract markup, compliance memos, financial reporting, Claude is still the sharper tool. On the Artificial Analysis Intelligence Index, Gemini 3.1 Pro scored 57, taking the top spot 4 points ahead of Opus 4.6 (53). But in the overall Text category, Gemini 3.1 Pro sits just 4 points behind the leading Opus 4.6, and human blind testing shows the two models as essentially tied.

Take the “beats everyone” framing with a pinch of salt. It’s the best model, or close to it, on a lot of tests. It’s not the best on all of them.

Second, latency at the start of a response is genuinely bad. Gemini 3.1 Pro Preview has a time to first token of 31.34s (based on Google’s API), which is at the higher end compared to other reasoning models in a similar price tier (median: 2.81s). That’s ten times the median. Once the tokens start flowing it flies, but the ~30-second stare into a “thinking” spinner before anything appears is the part that made me swear at my laptop more than once.

Everything else is annoyance-tier, not dealbreaker-tier. The pricing page is still opaque about limits. At I/O 2026 Google moved paid plans away from fixed daily prompt caps to compute-based usage limits, so a simple text prompt now spends far less of your allowance than a long video or coding session, which is fairer in principle but means you don’t really know what “4x higher usage than Free” means in prompts. If you want the top of the top, Deep Think for hardest reasoning and Gemini Agent for background tasks, you’re paying Ultra: $99.99/month with 5x higher usage limits vs. AI Pro, or $199.99/month with 20x higher usage limits, unlocking the most powerful features like video generation with Veo 3.1, Deep Research, audio overviews, and Google’s most capable AI models, like 3.1 Pro.

Should you pay for it?

For most people, yes, and it’s the easiest recommendation I’ve made in this category all year. Here’s how the tiers actually shake out.

The free Gemini tier costs $0. You get Gemini 3.6 Flash as the default model, a daily allotment of Gemini 3.1 Pro for harder reasoning, image generation with Nano Banana 2, up to five Deep Research reports per month, and Gemini Live voice mode. Storage is the standard 15 GB on Google One. No credit card and no subscription required. For casual use, that’s genuinely enough. Most people asking me “do I need to pay for this” don’t.

Google AI Pro is $19.99/month in the U.S. It includes 4x higher usage access than Free, higher access to Gemini 3 Pro in Google Search, 1,000 Google Flow credits, NotebookLM benefits, Gemini in Gmail, Docs, Vids and more, Google Home Premium Standard, YouTube Premium Lite, Gemini in Chrome early access, and 5 TB of cloud storage. That’s the tier that earns its keep. If you’re already paying for Google One storage or YouTube Premium, you’re effectively getting the frontier model for free. The consumer tiers broadly match the market: ChatGPT Plus costs $20/mo, Claude Pro costs $20/mo, and Gemini AI Pro sits at $19.99/mo. At this price point, Gemini differentiates through the Google ecosystem bundle: storage, Cloud credits, and native Workspace integration. The bundle is doing the work here.

Ultra is for a narrower crowd. $99.99/month for 5x higher usage limits vs. AI Pro, or $199.99/month for 20x higher limits. That’s the tier for people running Deep Think all day, Antigravity agents in the background, and Veo 3.1 video work. If that isn’t you, don’t touch it.

For developers, the API math is straightforward. Gemini 3.1 Pro costs $2.00/1M input (prompts up to 200k tokens) and $12.00/1M output, with a 1M token context window for deep reasoning. Prompts above 200K tokens move to $4.00 per 1M input tokens and $18.00 per 1M output tokens. Watch that 200K cliff on RAG pipelines; it’ll silently double your bill if you’re not paying attention. But at those numbers, it’s still cheaper per token than Opus for most workloads.

The bottom line

Gemini 3.1 Pro is the assistant most people should be using in August 2026. The reasoning is real, dynamic thinking makes it feel smart without any dial-fiddling, and the bundle at $19.99/month is the best value in the category. Opus 4.6 still edges it on the hardest coding and enterprise document work. If that’s your day, pay for Claude Pro instead. If you’re closer to the median use case (research, drafting, PDF wrangling, some code, some data work) Gemini 3.1 Pro is the one to beat, and the fact that it’s already stitched into Gmail, Docs, Search and NotebookLM makes the switch cost roughly zero. It earns the Editors’ Choice.