An ARC-AGI-3 “above human” bet just spiked to 92% on Manifold after OpenAI’s GPT-6 Astra. A Manifold market asking whether an AI will post above-human scores on the ARC-AGI-3 benchmark in 2026 jumped to a 92% implied probability of YES as of September 04, 2026, with 46% of its all-time volume changing hands in a single day. The trigger was OpenAI’s September 3 launch of GPT-6 Astra, whose ARC-verified results surpassed the human action-efficiency baseline on most levels while still falling short of full human completion under the strictest test setup. The result is a market pricing near-certainty on a question whose resolution wording is anything but certain.
The market that suddenly caught fire
The market in question is “Above human scores on ARC-AGI-3 in 2026?” hosted on Manifold and created by Zvi Mowshowitz, a widely followed writer on artificial intelligence. As of September 04, 2026 it shows a 92% implied probability that traders expect an AI system to reach above-human scores on the benchmark before year end.
What makes the move notable is not just the level but the speed. Roughly 46% of the market’s entire lifetime volume traded within the last 24 hours, one of the highest single-day volume readings on the platform for a market of its size, with 106 unique accounts active. On a play-money exchange like Manifold, that pattern signals a sharp shift in belief rather than gradual drift. The catalyst is not hard to find: it sits one day earlier in the calendar.
What ARC-AGI-3 actually measures
ARC-AGI-3 is a benchmark released by the ARC Prize Foundation on March 25, 2026. Unlike its predecessors, which asked models to infer static patterns from example grids, ARC-AGI-3 drops an agent into interactive game environments with no instructions and requires it to explore, form hypotheses about the hidden rules, and complete long-horizon tasks. The design targets the kind of on-the-fly reasoning that untrained humans handle easily and current models do not.
The human baseline is the headline. According to the ARC Prize Foundation, untrained humans solve 100% of the environments on their first attempt. At launch, every frontier model tested scored below 1%: Gemini 3.1 Pro Preview at 0.37%, GPT-5.4 at 0.26%, Claude Opus 4.6 at 0.25%, and Grok-4.20 at 0.00%, as documented by The Decoder. To spur research, ARC attached a prize pool reported at $2 million, including a grand prize for any agent that matches or exceeds human-level action efficiency on unseen environments, with a competition deadline in November 2026. For a primer on how a probability like 92% is derived from market prices, see our explainer on implied probability.
The odds and what 92% implies
A 92% implied probability is the market’s collective statement that “above human scores in 2026” is very likely, but not settled. In plain terms, traders are pricing roughly a one-in-twelve chance that the year ends without a qualifying result, which reflects genuine ambiguity in how the market resolves rather than doubt that models are improving. If you are new to reading these numbers, our guide to how prediction markets work walks through the mechanics.
The ambiguity matters here more than usual. “Above human” can mean matching the 100% human completion rate, or it can mean beating human efficiency, the number of actions taken to solve a level. As of early September, one of those thresholds has arguably been crossed and the other has not, which is exactly the kind of split that keeps a market off 99%.
What moved the market: GPT-6 Astra
On September 3, 2026, OpenAI launched GPT-6 Astra, which it called its most powerful model to date, with a headline emphasis on computer use. The ARC Prize Foundation published Astra’s ARC-AGI-3 results the same day. Under ARC’s Standard harness, Astra scored 62.7% for a reported compute cost of about $26,098. Under a Provider Adapter harness, it reached 99.9% for about $19,000.
The number driving the “above human” narrative is efficiency. ARC reported that Astra used fewer actions than the human baseline on 96.0% of levels and averaged 51.7% fewer actions per level. On that specific measure, the model is already operating above the human reference. ARC was careful to frame the milestone narrowly, calling it “meaningful progress towards generalization” while stating plainly that it is “not claiming that it is AGI” and that saturating the benchmark “would not represent proof of achieving AGI.”
OpenAI executives were less restrained. As reported by Fortune, OpenAI president Greg Brockman said it is “not unreasonable to feel that we are now in the AGI era.” Trade coverage from The Register struck a more cautious tone, noting that the top-line figure depends heavily on the test configuration.
The harness problem: why the score is contested
The gap between 62.7% and 99.9% for the same model is the crux of the debate, and it is not new. A benchmark score on ARC-AGI-3 measures the whole agent system, meaning the model plus the scaffolding that feeds it context and manages its reasoning between turns. Change the harness and the number moves, sometimes dramatically.
OpenAI demonstrated this earlier with GPT-5.6 Sol. In the official harness the model scored 7.8%. Running it through OpenAI’s own Responses API with two settings, retained reasoning across turns and context compaction, lifted the same weights to 38.3% on the public task set, a claim OpenAI detailed in its own writeup and that The Decoder flagged as unverified on the private set. The lesson traders have absorbed is that “above human” may depend on which harness the resolver accepts, which is why a 92% market has not simply gone to 99%.
The road here: Opus 5, GPT-5.6 Sol, and Nvidia’s AVO
The September spike did not come from nowhere. Scores climbed through the summer as labs and outside teams built better scaffolding around existing models. The clearest example arrived on August 21, when Nvidia published results for an agent architecture called AVO, short for Agentic Variation Operators. Wrapped in AVO, Claude Opus 5 went from about 30% standalone to a reported 100.00 on the ARC-AGI-3 public set, solving all 183 levels across 25 environments in 6,624 environment actions, roughly 12% fewer than the prior leading system, as covered by The New Stack. Crucially, AVO changed the harness, not the model.
| Date | System / event | ARC-AGI-3 result | Setup |
|---|---|---|---|
| Mar 25, 2026 | Benchmark launch, frontier models | Below 1% (humans 100%) | Official harness |
| Jul 2026 | Claude Opus 5 | ~30.2% | Standard harness |
| Jul 2026 | GPT-5.6 Sol | 7.8% to 38.3% | Standard vs custom API settings |
| Aug 21, 2026 | Nvidia AVO + Opus 5 | 100.00 (public set) | Nvidia harness, public levels |
| Sep 3, 2026 | GPT-6 Astra | 62.7% / 99.9% | Standard vs Provider Adapter |
Read together, the trend is a step change in what scaffolding can extract from a fixed model. That trajectory, capped by an ARC-verified efficiency result from Astra, is what pushed the market to price the year-end question as near-settled.
Who and what is involved
The named actors are concentrated. The ARC Prize Foundation, co-founded by researcher Francois Chollet, owns the benchmark, the human baseline, and the official leaderboard, and it remains the arbiter of what counts. OpenAI produced GPT-6 Astra and the earlier GPT-5.6 Sol; Anthropic supplied Claude Opus 5; Nvidia built the AVO harness. On the market side, Manifold hosts the wager and Zvi Mowshowitz created it. Each of these parties has a different incentive, and the tension between lab marketing, benchmark stewardship, and trader positioning is precisely what the odds are trying to reconcile.
Why this matters for a tech and finance reader
Benchmarks like ARC-AGI-3 increasingly function as pricing inputs. Model capability claims move enterprise buying decisions, compute demand, and by extension the valuations of companies across the AI supply chain. When a lab reports a near-perfect score and a benchmark author simultaneously cautions that the number is scaffolding-dependent and not evidence of AGI, the disagreement itself is information. Prediction markets compress that disagreement into a single, tradeable probability, which is why they have become a quick read on contested technical claims. For comparison of the venues where these bets live, see our looks at Manifold, Kalshi versus Polymarket, and our Polymarket review.
What to watch next
Three things will decide whether this market resolves clean or messy. First, resolution criteria: whether the market accepts efficiency-above-human, a Provider Adapter completion score, or only a Standard-harness match of the 100% human rate. Second, ARC’s official leaderboard and the November 2026 prize deadline, which set the strictest bar and remain unclaimed at the full-benchmark level. Third, independent reproduction on private held-out environments, since every headline number so far leans on either a public subset or a lab-controlled harness. Until an ARC-supervised result crosses the human completion bar on unseen tasks, the remaining 8% of doubt in the market is doing real work.
Frequently asked questions
Has an AI officially beaten humans on ARC-AGI-3? Not unambiguously. GPT-6 Astra surpassed the human action-efficiency baseline on 96.0% of levels per ARC, but under the Standard harness it completed 62.7% of the benchmark versus the 100% human rate. Whether that counts as “above human” depends on the definition used.
Why did the odds jump on September 4, 2026? OpenAI launched GPT-6 Astra on September 3 and ARC published verified results the same day, including the above-human efficiency figure. Traders repriced the year-end question within 24 hours.
What is the difference between the 62.7% and 99.9% scores? They are the same model under different harnesses. The Standard harness is ARC’s controlled setup; the Provider Adapter allows richer context handling. The gap shows how much benchmark scores depend on scaffolding.
Is the $2 million ARC prize claimed? No source reports the grand prize as awarded. It requires matching human-level performance on the full, unseen benchmark under ARC’s protocol, with a deadline in November 2026.
Where can I see the market? The market is public on Manifold at the link in the Sources section below.
The Bottom Line
Sources
- ARC Prize — OpenAI’s GPT-6 Astra on ARC-AGI-3
- Fortune — OpenAI debuts GPT-6 Astra, Brockman says start of AGI
- The Register — OpenAI throws Astra into the top-tier model ring
- The Decoder — ARC-AGI-3 offers $2M as every frontier model scores below 1%
- The New Stack — Nvidia’s AVO lifts Claude Opus 5 to 100% on ARC-AGI-3
- OpenAI — How two settings tripled our ARC-AGI-3 scores
- The Decoder — OpenAI claims GPT-5.6 Sol beats Opus 5 with its own harness
- Manifold — Above human scores on ARC-AGI-3 in 2026?
Prediction markets carry risk and are not investment or betting advice. Market availability is restricted by jurisdiction: Polymarket is not available to US persons, while Kalshi is a CFTC-regulated US exchange; Manifold operates with play-money in many regions. Age restrictions of 18+ or 21+ apply as applicable. If gambling is a problem for you or someone you know, help is available at 1-800-GAMBLER.
Related Coverage
- Manifold Traders Bet 95% on a Non-MF Group Proof This Month After Astra's Math Surge
- A Manifold Market Now Puts a Non-MF Group Proof at 70% After Astra's Math Shock
- GTA VI 2026 Release Odds Hit 96% on Manifold After Netflix Extended Look
- AI Index Clears 65: Traders Flood a Manifold Market After Claude Fable 5.1 Scores 66
- Will Ghent Fire Nathan Cofnas Before 2028? Traders Pile Into Manifold Bet at 22%