Google just dropped another bombshell — on February 19, they released Gemini 3.1 Pro, and this is not just a cosmetic update. Honestly, I was even surprised when I saw they skipped version 3.0 and jumped straight to 3.1. Usually, Google follows a pattern of 1.0 - 1.5 - 2.0 - 2.5 - 3.0, but this time, something different.

The main feature isn’t increased capabilities but a real upgrade to the model’s intelligence. They integrated deep thinking technology directly into the core, so now the model can consider a problem from multiple angles simultaneously and choose the best solution. Previously, this was an optional feature for subscribers; now it’s standard.

What really impressed me during testing? The model identified a non-obvious logical error in code that many programmers would have simply overlooked. It wrote an optimal SQL query with the correct indexes and explained the Monty Hall paradox in a way even a layperson can understand. This isn’t just fact retrieval — it’s genuine reasoning.

Benchmark numbers are impressive:
- ARC-AGI-2 (logic test): 77.1% versus 31.1% in the previous version. That’s two and a half times higher.
- Humanity’s Last Exam: 44.4% — higher than GPT-5.2 (34.5%)

But it’s not just about the numbers. The demos they released are fantastic. The model understood the atmosphere of the novel "Wuthering Heights" and generated a portfolio website that truly conveys the dark mood of the book. Created an interactive 3D simulation of a flock of birds, where you can influence the flight path with your hand. Connected to the ISS API and built a real aerospace dashboard. This isn’t just code generation — it’s understanding context and creativity.

Multimodality remains at the flagship level: text, images, video, audio, PDF. The context window is still about a million tokens (roughly like loading "War and Peace" entirely), but the maximum output has increased to 64-65 thousand tokens. That’s a whole small book in one request.

Where the community found weak spots? Office tasks. In the GDPval-AA (presentations, tables, documents) test, Gemini scored 1317, while Claude Sonnet 4.6 scored 1633. In rankings where people vote for the most appealing answers, Claude still leads. So, for complex coding — Gemini is top; for beautiful presentations — Claude is better.

And now, the most delicious part — price. Google didn’t raise it. Entry for a million tokens is $2 (if context up to 200K) or $4 (if more). Output costs $12 or $18. For comparison: Claude Sonnet 4.6 charges $3 for input and $15 for output, GPT-5.2 is $10 and $30, and Claude Opus 4.6 is $15 and $75.

In practice: if you need to process 100K input tokens and 10K output tokens:
- Gemini: $0.32
- Claude Sonnet: $0.45
- GPT-5.2: $1.30
- Claude Opus: $2.25

Gemini is 7 times cheaper than Opus. For production workloads, this difference is significant.

Artificial Analysis ran tests: they processed all top models through their Intelligence Index. Gemini used 56 million tokens and cost $892. GPT-5.2 consumed 130 million ($2,304), Claude Opus — 58 million ($2,486). That means the same amount of intelligence on Gemini costs 2.6 times less.

For regular users, there are subscriptions. Google AI Plus for $8/month — this is the Pro mode Gemini 3.1 Pro, Deep Research, 1000 images per day. Enough for most. Google AI Pro for $20 — 100 requests per day and 20 Deep Research. Google AI Ultra for $250 — everything above plus Deep Think and maximum priority.

Why didn’t Google raise the price? Several reasons. First, they focus on accessibility — free AI Studio, generous free tokens, low API costs. Second, competition. Anthropic released Claude Sonnet 4.6 — a great price-to-quality ratio. OpenAI isn’t sleeping. To keep developers engaged, a good price is essential. Third, the model is still in preview, so Google can afford to underprice and gather feedback. When it officially launches, prices may change.

Overall? This is a very good update. Logic has genuinely improved, prices haven’t increased, code quality is rapidly growing. Not perfect, but very attractive — especially if you’re a developer who counts every dollar and can read technical specs.

Of course, there are nuances. If you need complex system code, GPT-5.3-Codex might be better. If you’re preparing presentations for the board, Claude Sonnet 4.6 is more polished. If working with confidential data, none of these models are suitable because all run on servers in the US.

But what’s really important: models released today may be outdated tomorrow. While I was writing this, engineers at Google, OpenAI, and Anthropic are probably already training something new. So my advice: get Gemini 3.1 Pro, try it on your real tasks. If it works — enjoy and save money. If not — there are plenty of other options. Many good models are available now, and that’s the best news.