Gemini 3.1 Pro: Google's Mid-Cycle Update Puts Reasoning Front and Center

Google released Gemini 3.1 Pro today, February 19, 2026. It’s the first time the company has shipped a “.1” increment as a mid-cycle model update rather than the “.5” jumps we’ve seen in past generations. The naming signals exactly what this is: not a new model family, but a targeted upgrade to Gemini 3 Pro’s reasoning capabilities, with the Deep Think architecture that previously lived only in a separate mode now baked into the base model.

The result is a model that tops 13 of 16 benchmarks Google tested against, including several where it beats both Claude Opus 4.6 and GPT-5.2. But the full picture, as usual, is more complicated than a leaderboard.

What Changed From Gemini 3 Pro

The core upgrade is reasoning. Gemini 3 introduced Deep Think as a separate mode for problems that needed extended cognition. With 3.1 Pro, that reasoning architecture is integrated into the model itself, available to all users rather than gated behind a special configuration.

The most dramatic benchmark shift reflects this: Gemini 3.1 Pro scores 77.1% on ARC-AGI-2, a test of abstract reasoning. That’s more than double Gemini 3 Pro’s score and a significant gap over Claude Opus 4.6’s 68.8%. ARC-AGI-2 tests the kind of novel pattern recognition that has historically been difficult for language models, so a jump this large is worth paying attention to.

Google also improved agentic task performance. APEX-Agents, which measures how well models handle multi-step tool-using workflows, went from 18.4% on Gemini 3 Pro to 33.5% on 3.1 Pro. Nearly double. For anyone building agent systems on top of Gemini, that’s the number that matters most.

The context window stays at 1 million tokens input and 64,000 tokens output, same as Gemini 3 Pro. The multimodal capabilities (text, images, audio, video, PDFs, code) carry over unchanged.

The Benchmark Picture

Here’s where Gemini 3.1 Pro lands across the benchmarks that matter, compared to the current competition:

Where Gemini 3.1 Pro leads:

GPQA Diamond (expert scientific knowledge): 94.3%, vs. Claude Opus 4.6 at 91.3% and GPT-5.2 at 92.4%
ARC-AGI-2 (abstract reasoning): 77.1%, vs. Claude Opus 4.6 at 68.8%
SWE-Bench Verified (agentic coding): 80.6%, first place
Terminal-Bench 2.0: 68.5%, first place
APEX-Agents: 33.5%, first place

Where competitors still lead:

GDPval-AA Elo (expert tasks): Claude Sonnet 4.6 Thinking scores 1633 vs. Gemini 3.1 Pro’s 1317. A wide gap.
Humanity’s Last Exam (with tools): Claude Opus 4.6 edges out at 53.1% vs. 51.4%
SWE-Bench Pro: GPT-5.3-Codex leads at 56.8%
Terminal-Bench 2.0 “other”: GPT-5.3-Codex at 77.3%

The pattern is clear: Gemini 3.1 Pro dominates on scientific reasoning and abstract problem-solving. Claude holds its advantage on expert-level evaluation tasks and some coding benchmarks. OpenAI’s Codex variant remains strong on certain coding-specific tests. No single model sweeps every category, which is roughly what you’d expect at this point in the market.

Pricing: Identical to 3 Pro

Google kept pricing exactly the same as Gemini 3 Pro:

	Under 200k tokens	Over 200k tokens
Input	$2.00 / 1M tokens	$4.00 / 1M tokens
Output	$12.00 / 1M tokens	$18.00 / 1M tokens
Batch input	$1.00 / 1M tokens	$2.00 / 1M tokens
Batch output	$6.00 / 1M tokens	$9.00 / 1M tokens

Context caching is supported at $0.20 per million tokens (under 200k) or $0.40 (over 200k), plus $4.50 per million tokens per hour for storage. Google Search grounding comes with 5,000 free prompts per month, then $14 per 1,000 queries.

This pricing keeps Gemini 3.1 Pro cheaper than Claude Opus 4.6 ($5/$25 per million tokens) and competitive with OpenAI’s offerings. For developers already on the Gemini API, the upgrade is free. Same price, better model.

Where to Use It

Gemini 3.1 Pro is available today in preview through the Gemini app, NotebookLM (for AI Pro and Ultra subscribers), the Gemini API via Google AI Studio, Vertex AI, Gemini Enterprise, Gemini CLI, and Android Studio.

The “preview” label matters. Google is using this rollout to validate the model before general availability, particularly around what they describe as “ambitious agentic workflows.” If you’re building production systems, test thoroughly before committing. Preview models can and do change behavior between preview and GA.

What This Means for Developers

The practical question is whether 3.1 Pro changes your model selection calculus. A few scenarios where it might:

Scientific and research applications. If you’re building tools for researchers, analysts, or anyone working with complex technical problems, the GPQA Diamond and ARC-AGI-2 improvements are real and meaningful. A 94.3% score on expert-level scientific knowledge is the highest any model has posted.

Agent-heavy architectures. The jump from 18.4% to 33.5% on APEX-Agents suggests Google has specifically optimized for multi-step tool use. If you’re building agentic systems and you’ve been frustrated by Gemini 3 Pro dropping the thread mid-workflow, this is worth testing.

Cost-sensitive deployments. At $2/$12 per million tokens (standard tier), Gemini 3.1 Pro offers strong reasoning performance at a lower price point than Claude Opus 4.6 or GPT-5.2. For high-volume applications where cost per query matters, the math favors Google here.

Where you probably stick with what you have. If your workload is primarily expert-level evaluation tasks (GDPval-AA style), Claude’s lead there is substantial. If you need the absolute best coding performance on complex repositories, GPT-5.3-Codex and Claude Opus 4.6 still trade blows at the top. And if you’re already deep in one ecosystem’s tooling, switching costs are real.

The Bigger Picture

Google’s decision to ship a “.1” update rather than waiting for a full generation bump tells you something about the pace of competition. When your competitors release strong models on a near-monthly cadence, holding improvements for a big splash six months later is a luxury you can’t afford.

The Deep Think integration is the real story here. Moving advanced reasoning from a separate mode into the base model makes it available to every API call, every consumer interaction, every agentic workflow. That’s a meaningful architectural choice. It suggests Google believes reasoning capability should be a default, not an opt-in feature.

Whether Gemini 3.1 Pro holds its benchmark leads once Claude and OpenAI respond is an open question. The cycle of one-upmanship has been relentless. But for today, Google has a model that leads on more benchmarks than any competitor, at a price point that undercuts both major rivals, with the same million-token context window that’s become table stakes at the top end.

For developers, the right move is straightforward: try it on your actual workloads, compare the results, and let your use case decide. Benchmarks set expectations. Your data tells you the truth.

Sources: