What prediction markets are and why they give a better picture than press releases

Prediction markets operate on a simple yet extraordinarily effective principle: thousands of participants put real money on the outcome of a specific event. The price of each "share" — expressed in cents from 0 to 100 — directly corresponds to the probability the market assigns to that outcome. It's not a public opinion poll or an expert estimate detached from reality. It's the collective wisdom of people risking their own capital. Platforms like Polymarket (the world's largest prediction market) and Metaculus (a more expert-oriented platform with a reputation system) became an unexpectedly accurate mirror of the AI industry in 2026. While technology companies churn out press releases full of superlatives and marketing promises, prediction markets reveal what those who invest time and money into the field actually believe. Studies repeatedly show that prediction markets are on average more accurate than individual experts and often even than consensus analyst estimates. The reason is simple: a participant who is wrong loses money. That creates a strong motivation for objectivity that simply doesn't exist in the marketing departments of technology companies.

Anthropic at the top: 87% favorite with a painful lead

The current Polymarket odds as of June 8, 2026 are unambiguous. In the market "Which company has the best AI model at the end of June?" Anthropic leads with a crushing 87% probability. Google — a technology giant with the DeepMind research division and virtually unlimited resources — has just 10%. OpenAI, until recently the undisputed leader of the field, didn't even make it among the main candidates in this particular market. Anthropic's dominance is even more pronounced in programming. 94% probability for the best coding model versus 5% for OpenAI — that's no longer a competition, it's a demolition of the competition. The reason is obvious: Claude Opus 4.8, released in May 2026, set a new standard in SWE-bench Verified benchmarks and in practical programming. According to Anthropic, the model is also "more honest" — when it's not sure of an answer, it says so plainly instead of hallucinating. And the most interesting part is yet to come. Claude Mythos, the long-awaited next generation model, has a 73% chance of release by the end of June and 87% by the end of July. Project Glasswing, within which Mythos has already discovered over 10,000 critical software vulnerabilities, suggests this will be a fundamental leap — not just in answer quality but also in practical usability for security auditing and development.

OpenAI: GPT-5.6 as the key response

OpenAI is not falling behind. Polymarket gives an 89% chance that GPT-5.6 will be released by the end of June, with 37% probability falling on the week of June 15–21. The release could come literally any day — and if it does, the odds could change dramatically within hours. Sam Altman called GPT-5.6 a "significant intergenerational leap." The market partially believes him: the model has a 97% chance of release by the end of July. But the long-term outlook is more interesting. GPT-6 has an 84% probability of release by the end of 2026, but only 47% by the end of September. OpenAI appears to be preparing a two-stage strategy: GPT-5.6 as an immediate countermeasure against Claude Mythos, and GPT-6 as the final ace up the sleeve by year's end. For comparison: Claude 5 has a 93% chance of release by the end of September 2026. The timelines of both companies overlap strikingly — and both labs know that every week of lead means millions of dollars in subscriptions and enterprise contracts.

Google Gemini: Why the tech giant is losing

The biggest surprise of the prediction markets is Google's position. A 10% chance of having the best AI model is frankly dismal for a company that has the DeepMind research team, its own TPU chips, and an annual AI research budget exceeding tens of billions of dollars. The next generation of the Gemini Pro model has an 81% chance of release by the end of June (96% by the end of July), but the market doesn't believe it will be enough. The problem isn't the technology — Gemini 3.5 Flash and Gemini Omni for video editing show that Google can innovate. The problem is in the product strategy: Gemini models excel in multimodal tasks and working with enormous context, but in purely coding and mathematical benchmarks they lag behind both Claude and GPT. Moreover, Google struggles with brand perception. While Claude is synonymous with safety and reliability and ChatGPT with accessibility, Gemini is searching for its identity — and the prediction markets relentlessly reflect that.

IPO: Who goes public first and at what valuation

Prediction markets also offer a unique window into corporate strategy. 83% chance that Anthropic's IPO will beat OpenAI's — that's one of the clearest signals on all of Polymarket. OpenAI has an overall 74% probability of going public by the end of 2026. Even more interesting are the valuations. According to the markets, Anthropic is heading for a 94% chance of reaching a $1.1 billion valuation by year's end. For context: that's more than most technology companies outside the very top tier have today. The market is essentially betting that Claude will be the backbone of enterprise AI much like Windows is the backbone of corporate computers. OpenAI is moving in similar numbers — but with one crucial caveat: the 74% chance of an IPO by year's end is significantly lower than the 94% probability of Anthropic's valuation milestone. In other words: the market believes Anthropic has a better-prepared path to the stock exchange.

What this means for Czech users and businesses

For Czechs who use AI or are considering deployment, the prediction markets yield specific conclusions: First — competition drives prices down. Claude Pro costs $20 per month (roughly 460 CZK), ChatGPT Plus also $20, Gemini Advanced €21.99 (about 550 CZK). All three services are commonly available in the Czech Republic. The price war that prediction markets forecast means better models for the same or less money. OpenAI also recently launched ChatGPT Go for $5 per month (roughly 120 CZK), which currently targets Asian markets — but its arrival in Europe is only a matter of time. Second — Czech language support is at a historic best. Claude 4 and GPT-5.5 handle Czech with minimal grammatical errors and good knowledge of Czech realities. Business communication, marketing copy, or technical documentation in Czech are no longer a problem for AI. Third — European regulation is an invisible player. The EU AI Act is in force and applies to all companies operating in the EU, including Czech ones. It may slow the deployment of certain advanced models in the European market — and prediction markets currently barely account for this factor. This creates a potential disconnect between global odds and European reality. Fourth — Anthropic is gaining the trust of enterprise customers. Thanks to its emphasis on safety, transparency, and a "constitutional" approach to AI development (Constitutional AI), Claude is building a position as the preferred model for companies that cannot afford hallucinations or data leaks. In the Czech Republic, Claude is used by a growing number of development teams, startups, and larger companies — including the Czech National Bank, which operates its own AI infrastructure.

How to read prediction markets — and when they can be wrong

Prediction markets are a powerful tool, but they are not infallible. Their main advantage — aggregating information from thousands of motivated participants — is also their Achilles' heel. If no participant has a key piece of information, the market cannot price it in. DeepSeek is a case in point. In December 2024, virtually no one expected that a Chinese startup would shake up the entire industry within a month with a model that competes with GPT-4 in performance for a fraction of the price. Prediction markets completely missed this "black swan" moment. Markets say: the collective estimate of people risking their own money, aggregating publicly available information, rumors, and expectations as of June 8, 2026. Markets don't say: objective truth. They are bets — and bets sometimes lose.

Czech footprint: How to get involved

Polymarket is essentially accessible from the Czech Republic as well, although gambling regulations create certain barriers. For academic and research purposes, Metaculus is more suitable — it doesn't require financial deposits and operates on a reputation score basis: the more accurate your predictions, the higher weight they carry. For Czech researchers, journalists, and AI enthusiasts, prediction markets represent a unique opportunity: to monitor the development of the field through the lens of data and collective intelligence, instead of relying on the marketing departments of companies whose main product is, ultimately, hype.