Polymarket believes in Google: 96% chance that Gemini 3.2 arrives by the end of June

Prediction markets are an increasingly accurate barometer of expectations in the tech world. And right now, Polymarket, the world's largest prediction platform, is showing interesting numbers around Google's upcoming Gemini 3.2 model.

The market „Gemini 3.2 released by...?" currently has a volume of over 107 thousand dollars and shows a 96% probability that the model will arrive by June 30, 2026. Even more interesting is the detailed market „Gemini 3.2 released on...?", which breaks down individual days. The highest probabilities fall on May 11 and 12 (both 38%), followed by May 14 (34%) and May 13 (28%).

The analytical account BridgeMind on the X network noted that the cumulative chance of release in the next week (May 12–16) reaches 47%. That is an unusually high number on the prediction market for such a specific time window.

Why right now? Google I/O as the ideal springboard

Google I/O, Google's annual developer conference, traditionally takes place in May. Although Google has not yet officially announced any date for 2026, the May window is historically the moment when the company presents its biggest AI news. It was at Google I/O 2025 that Google unveiled Gemini 3.0 — and the community expects a similar scenario to play out this year as well.

After all, Polymarket also has a market on Gemini 3.5, where a 36% chance is assigned to release by July 31. If Google maintains its aggressive pace of introducing new models, we could see two significant updates during the summer.

The problem was not intelligence, but tools

Gemini 3.1 Pro, which Google introduced at the beginning of 2026, brought excellent results in reasoning and, according to BridgeBench, achieved the lowest hallucination rate of all tested models. BridgeBench is a specialized benchmark focused precisely on factual accuracy — that is, whether the model "invents" information that it does not have substantiated.

But there was one fundamental catch: tool calling, or the model's ability to reliably call external tools and APIs. In practice, this means that when you asked Gemini to perform a specific action — search data in a database, send an email, run code, or work with files — the model often failed, performed an incorrect operation, or did not respond at all.

For comparison: GPT-5.5 from OpenAI and Claude Opus 4.7 from Anthropic excel in this discipline. Both companies have massively invested in the agentic capabilities of their models — that is, in making AI not just "think", but also actually do something.

What is tool calling and why it matters

Tool calling is the ability of a language model to interact with external systems. Think of it as the difference between an advisor who tells you what to do, and an assistant who actually does it for you.

In agentic AI systems of 2026, reliable tool calling is absolutely crucial:

• Corporate agents need to read corporate data, write to CRM, send messages
• Development tools like Claude Code or GitHub Copilot rely on running commands, reading files, and interacting with the terminal
• Personal assistants must be able to work with calendar, emails, and other applications

Without reliable tool calling, a model is like a sports car with an empty tank — it has performance, but you will not get anywhere with it.

What Gemini 3.2 success would mean for the market

If Google truly fixes tool calling in Gemini 3.2 and at the same time maintains or improves existing qualities in reasoning and factual accuracy, it could become a full-fledged competitor to GPT-5.5 and Claude Opus 4.7.

The impact would be significant on several levels:

1. Price competition — Google traditionally offers more aggressive prices for API access than OpenAI. While GPT-5.5 costs approximately 15 USD per million input tokens via API, Gemini 3.1 Pro was around 2.50 USD. If Google maintains its pricing policy, Gemini 3.2 could significantly lower the cost of access to top-tier AI.
2. Google ecosystem — Gemini is integrated into Gmail, Google Workspace, Android (including Pixel devices), and the Chrome browser. An improved model would thus immediately enhance the AI experience for billions of users.
3. European availability — Gemini is fully available in the Czech Republic via gemini.google.com and supports Czech at a solid level. Google One subscription (from 59 CZK per month) provides access to Gemini Advanced. For Czech companies, Gemini API is available via Google Cloud with billing in crowns.

BridgeBench: Why accuracy matters

BridgeBench is a relatively new benchmark focused on measuring hallucinations — that is, the tendency of AI models to generate untrue or misleading information. At a time when AI is increasingly deployed in areas such as healthcare, law, or finance, factual accuracy is absolutely crucial.

The fact that Gemini 3.1 Pro led in this benchmark indicates that Google invested in reliability, not just raw performance. This is a trend that companies deploying AI in production environments will particularly appreciate — no one wants their corporate chatbot to invent non-existent facts.

What is next: Google I/O, Gemini 3.5, and the race for agentic AI

Polymarket is betting not only on Gemini 3.2, but also on Gemini 3.5 (36% by the end of July) and a new Gemini reasoning flagship (89% by the end of June). This suggests that the community expects several significant model releases from Google in the coming months.

For Czech developers and companies, this means one thing: watch Google's May announcement. If Gemini 3.2 truly arrives with fixed tool calling, it could be the model that rewrites the current order on AI leaderboards — and offers a solid alternative to the dominant models from OpenAI and Anthropic at a fraction of the price.