83% Odds: Manifold Bettors Say AI Will Play Any Game at Human Level by 2028. A once-quiet Manifold market asking whether an AI can play randomly selected computer games at human level, with no chance to practice, has jumped to an 83% implied probability of YES as of September 13, 2026. The move follows a wave of game-playing demonstrations from OpenAI’s newly released GPT-6 Astra, and the market now ranks among the most actively traded on the platform, with 596 unique bettors.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What This Market Actually Asks
The market, created by writer Scott Alexander on Manifold, poses a deceptively simple question: “In 2028, will an AI be able to play randomly selected computer games at human level without getting to practice?” The two conditions that make it hard are the words “randomly selected” and “without getting to practice.” A model that trains for weeks on a single title, or that is handed a walkthrough, does not qualify. The bar is zero-shot competence on games the system has never seen.
That framing maps almost exactly onto how researchers now define general-purpose reasoning in interactive environments. It is why the market has become a shorthand bet on whether frontier AI is closing the gap between narrow benchmark wins and open-ended adaptation. For readers new to these venues, our explainer on how prediction markets work covers the mechanics of how a question like this gets priced.
Current Odds and What They Imply
As of September 13, 2026, the market sits at roughly 83% implied probability of YES on Manifold. In prediction-market terms, that means the crowd is pricing the 2028 milestone as far more likely than not, though still well short of a settled outcome. Manifold runs on its own play-money currency, so the figure reflects belief and reputation rather than dollars at risk, a distinction worth keeping in mind and one we unpack in our guide to implied probability.
An 83% reading leaves meaningful doubt: roughly a one-in-six chance the crowd assigns to NO. The interesting part is not the level but the velocity. The question had traded quietly for long stretches before the September surge in trading volume and participation pushed it up the platform’s activity rankings.
What Is Driving the Move
The catalyst is GPT-6 Astra, which OpenAI released on September 3, 2026. The model’s headline feature is computer use: driving a desktop, reading a screen, and operating software directly. Within days, several demonstrations reframed the model as a game player rather than just an office assistant, and traders reacted.
The most viral example came on September 7, 2026, when OpenAI’s Sharif Shameem posted that Astra had cleared all 48 levels of Neal Agarwal’s browser game “I’m Not a Robot,” a title built specifically to separate humans from bots. The result was covered by Benzinga via TradingView and widely amplified elsewhere. Separately, an early tester reported that Astra finished Pokemon FireRed in 18 hours and 12 minutes using a screenshots-only setup with no walkthrough, compared with 96 hours for the prior GPT-5.6 model and an earlier version that had not finished after more than 200 hours.
None of these individually resolves the 2028 question, but together they shifted the narrative from “AI can pass benchmarks” to “AI can figure out unfamiliar interactive worlds,” which is precisely what the market is about.
The Benchmark That Anchors the Debate
Behind the hype sits a harder measure. ARC-AGI-3, the interactive benchmark launched by the ARC Prize Foundation, is effectively a set of turn-based video games with no instructions, no stated rules, and no goals. An agent has to explore, infer the mechanics, and carry what it learns forward. According to ARC Prize, this is the closest public test to the Manifold market’s exact conditions.
The gap it exposes is stark. As The Decoder reported, untrained humans score 100% while frontier models given full game state scored below 1% in early 2026 testing. That is the tension traders are wrestling with: flashy single-game feats on one side, a sub-1% score on the true zero-shot test on the other, and 27 months left on the clock.
Who and What Is Involved
The named players in this story are OpenAI, whose GPT-6 Astra release supplied the spark; ARC Prize and its founder Francois Chollet, whose benchmark defines the strict version of the task; and Scott Alexander, the writer and forecaster who authored the market. On the reaction side, Nvidia CEO Jensen Huang publicly declared that general intelligence had “arrived” with Astra, and OpenAI President Greg Brockman said the company was moving into “the AGI era,” per the Benzinga coverage. Those statements are opinion, not measurement, but they fed the trading momentum.
On the benchmark record, OpenAI reported that Astra scored 72.6% on OSWorld 2.0, a test of computer and browser use, up from 65.7% for the previous model, according to third-party summaries from Vellum. That is a genuine step up in software operation, which is a prerequisite for playing games through a screen, without being proof of zero-shot game mastery.
Broader Context for a Tech and Finance Reader
The Astra release lit up more than one forecasting venue. Kalshi, the CFTC-regulated US exchange, saw its market on whether any company will formally announce AGI before April 1, 2027 jump from 32% to 62% in a single day, briefly touching 72%, according to the Benzinga report. The Manifold game market is a narrower, more concrete proxy for the same underlying question, which is why it moved in sympathy.
For investors, the signal is less about any single wager and more about how quickly sentiment on AI capability can reprice. If you want to compare the venues themselves, see our breakdowns of Kalshi versus Polymarket and our Manifold Markets review. This is also not the first offbeat AI question to draw a crowd on Manifold; we recently covered a similar surge in the market on whether an AI will perform magic.
Timeline of the Move
| Date (2026) | Event | Relevance to the market |
|---|---|---|
| Early 2026 | Frontier models score below 1% on ARC-AGI-3; humans score 100% | Sets the strict zero-shot baseline the question targets |
| Sept 3 | OpenAI releases GPT-6 Astra with computer-use focus | Supplies the capability behind later game demos |
| Sept 7 | Astra clears all 48 levels of “I’m Not a Robot” | Viral proof of screen-reading and multi-step control |
| Early Sept | Astra reportedly finishes Pokemon FireRed in 18h 12m, no walkthrough | Single-game speed run below typical human times |
| Sept 13 | Manifold market reaches ~83% YES with 596 bettors | Crowd prices the 2028 milestone as likely |
The Case for NO
The skeptical view is straightforward. Every headline win so far involves a single, named game or a benchmark with a known structure. The market demands human-level play across randomly selected titles with no practice, and the one public test built to those specifications still humbles frontier systems. Speed-running one Pokemon game, however impressive, is not evidence of generalization to an arbitrary game pulled from a shelf. Bettors on NO are effectively wagering that the last mile, from strong software operation to genuine zero-shot adaptation, is harder than the current momentum suggests.
What to Watch Next
Three signals will move this market between now and resolution. First, ARC-AGI-3 leaderboard scores: any jump above the low single digits would be a direct update. Second, standardized game-playing evals that use randomly selected titles rather than curated ones. Third, whether OpenAI, Google DeepMind, or others publish reproducible zero-shot results rather than one-off social media posts. Until an agent posts credible numbers on unseen games under controlled conditions, the 83% figure rests more on trajectory than on a met threshold.
FAQ
What does the Manifold market resolve on? It asks whether, in 2028, an AI can play randomly selected computer games at human level without practicing first. Resolution depends on the creator’s judgment of whether that bar is met.
Is 83% a real-money probability? No. Manifold uses play-money, so the figure reflects crowd belief and reputation rather than capital at risk. It still functions as a forecast, but it is not a regulated financial instrument.
Did GPT-6 Astra already meet the market’s condition? No. Astra’s wins were on specific games and benchmarks, not randomly selected titles under zero-shot conditions. The strict test, ARC-AGI-3, still shows frontier models below 1%.
Why did the odds move now? The September 3 release of GPT-6 Astra and its game-playing demonstrations shifted sentiment about how fast AI is generalizing, lifting this and related AGI-timing markets.
The Bottom Line
Sources
- Manifold — In 2028, will an AI be able to play randomly selected computer games at human level without getting to practice?
- Benzinga / TradingView — GPT-6 Astra beats “I’m Not a Robot,” and what prediction markets say about AGI
- OpenAI — GPT-6 Astra: A new generation of intelligence
- Vellum — GPT-6 Astra benchmarks explained (OSWorld 2.0)
- ARC Prize — Announcing ARC-AGI-3
- The Decoder — Frontier models score below 1% on ARC-AGI-3
Disclaimer: Prediction markets carry risk and are not investment or betting advice. Availability is restricted by jurisdiction: Polymarket is not available to US persons, and Kalshi is a CFTC-regulated US exchange; Manifold operates on play-money. Age restrictions of 18+ or 21+ apply where relevant. If gambling is a problem for you or someone you know, call 1-800-GAMBLER for confidential help.
Related Coverage
- Trump on US Currency? Manifold Traders Price 97% Odds His Face Appears Before 2028
- Will Polymarket IPO Before 2028? Manifold Bettors Now Put It at 43%
- Manifold Bettors Put 93% Odds on Bitcoin Hitting $85K Before the Strait of Hormuz Reopens
- Superconductor Bet Heats Up: Manifold Traders Give a Room-Temperature Breakthrough 15% Odds Before 2031
- Millennium Prize Problems Bet: Manifold Odds Jump to 24% on OpenAI Claim
