A niche AI benchmark bet just became one of the busiest markets on Manifold. A contract asking whether the top Artificial Analysis Intelligence Index score will reach 65 or higher by September 30, 2026 is trading at roughly 98% implied probability of YES as of September 02, 2026, and most of its lifetime volume changed hands in a single day. The trigger is not a mystery: Anthropic shipped Claude Fable 5.1 on September 1, and it posted a score of 66 on the index, clearing the threshold weeks before the market is set to resolve.
What the market actually asks
The contract lives on Manifold, the play-money prediction platform, and was created by user Bumin Anlakoğlu. The question is narrow and testable: will the highest score on the Artificial Analysis Intelligence Index leaderboard read 65.0 or higher when the market closes at 23:59 UTC on September 30, 2026? Resolution is mechanical. A single screenshot of the public leaderboard is taken at the closing timestamp, and the score has to be published on the page at that moment. An announcement, a preprint, or a vendor blog post does not count on its own.
That design matters because it turns a fuzzy question about AI progress into a binary, date-stamped event. Traders are not betting on whether a lab claims a number. They are betting on what a specific third-party leaderboard displays on one specific night.
The odds and what they imply
As of September 02, 2026, the market sits at about 98% implied probability of YES on Manifold. In prediction-market terms, an implied probability is the price of the YES share expressed as a percentage, and 98% means the crowd views the outcome as close to settled. The market lists a small book, with roughly two dozen holders and a few dozen trades, which is typical for a play-money contract that only recently caught attention. The notable signal is concentration: the outlet tracking this market recorded that a large majority of its all-time volume traded in the most recent 24 hours, a spike that lines up precisely with the model release that made the threshold look trivial.
Because this is Manifold, the stakes are in the platform’s virtual currency rather than dollars, so the 98% figure is a measure of conviction, not a regulated financial price. If you want a primer on how these instruments turn opinions into prices, see our explainer on what prediction markets are and how they work.
What is driving the sudden move
The catalyst is Claude Fable 5.1. Anthropic released it alongside Claude Mythos 5.1 on September 1, 2026, describing the pair as its most advanced models for coding and knowledge work. Artificial Analysis, which ran a pre-release evaluation, reported that Fable 5.1 at maximum effort scores 66 on its Intelligence Index, calling it the highest score the firm has measured. That is a four-point gain over Fable 5 and enough to pass the market’s 65 line outright.
The release was not only about raw capability. Anthropic kept base pricing at $10 per million input tokens and $50 per million output tokens, but cut cached input reads by 75%, from $1.00 to $0.25 per million. Artificial Analysis noted that the top score carries a cost: at max effort Fable 5.1 runs about $3.76 per index task, roughly 20% more than Fable 5 at $3.14, because it uses around 1.7 times the output tokens. Step the model down to its “xhigh” effort setting and it still lands at 65 on the index for about $2.72 per task. For a tech or finance reader, that trade-off between headline score and per-task cost is the more durable story, and it echoes the model comparison work in our look at how GPT-5.6 Sol, Terra, and Luna stack up. For the full launch rundown, see our coverage of the Claude Fable 5.1 and Mythos 5.1 launch.
What the Artificial Analysis Intelligence Index measures
The Artificial Analysis Intelligence Index is not a single test. The current version, v4.1.1, is a weighted composite of nine hard evaluations, then scaled from 0 to 100. The component benchmarks are GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. Together they span reasoning, coding, scientific problem solving, general knowledge, and multi-step agentic tasks. It is a text-only, English-language suite, so it does not directly capture image, audio, or video performance.
Version history is part of why the 65 threshold was interesting in the first place. Artificial Analysis has repeatedly recalibrated the index to keep headroom, at one point resetting top scores toward the low 50s so that further progress had somewhere to go. Under the stricter v4.1.1 regime, watching a model climb back into the mid-60s is a meaningful signal about the pace of capability gains rather than a quirk of easy grading.
The leaderboard right now
Two public views of the index agree on the ordering. Artificial Analysis’s own model page lists Claude Fable 5.1 at 66, the top of 184 models evaluated. A display-only snapshot on BenchLM, updated September 2, 2026, shows the same leader at 65.7% with a full top five. The small gap between 66 and 65.7 reflects effort settings and snapshot timing, not a disagreement about who leads.
| Rank | Model | Developer | Index score |
|---|---|---|---|
| 1 | Claude Fable 5.1 | Anthropic | 65.7% (66 at max effort) |
| 2 | Claude Opus 5 | Anthropic | 63.0% |
| 3 | Claude Fable 5 | Anthropic | 62.1% |
| 4 | Grok 4.6 | xAI | 60.9% |
| 5 | Kimi K3 | Moonshot AI | 59.7% |
Source: BenchLM snapshot of the Artificial Analysis Intelligence Index, updated September 2, 2026, cross-checked against the Artificial Analysis model page. Scores reflect the v4.1.1 scale.
Why a near-settled market still has a tail
If the threshold is already cleared, why is the market at 98% rather than 100%? The remaining doubt is procedural, not about capability. The resolution rule specifies the score must be published on the leaderboard page at 23:59 UTC on September 30, so a display change, a page outage, or a model being pulled at the wrong moment could in theory matter. There is also version risk. The rules say that if the index version changes before resolution, scores will be chain-linked back to the current v4.1.1 scale using anchor-model adjustments, with a minimum of three anchors required, or the market resolves N/A. Artificial Analysis has recalibrated before, so a surprise methodology change is the low-probability scenario keeping the last two points of doubt alive.
That is the honest read of a 98% market. The event looks decided, and the residual price is the market paying respect to edge cases rather than to any real chance a leading model falls back below 65.
Who and what is involved
Three parties sit at the center. Anthropic is the model maker whose September 1 release moved the number. Artificial Analysis is the independent evaluation firm that publishes the index and ran the pre-release testing, making its leaderboard the market’s official resolution source. Manifold is the venue where the bet trades, with Bumin Anlakoğlu as the market creator. None of these is a regulated exchange, and none of the trading involves real money, which shapes how much weight to put on the 98% figure.
The capability race behind the number is broader than Anthropic. Independent snapshots put xAI’s Grok 4.6 and Moonshot AI’s Kimi K3 in the high 50s to low 60s, and coverage of the release noted GPT-5.6 Sol near 61 at max effort. The threshold cleared this month because one lab shipped first, but several labs are clustered close enough that the top score was likely headed past 65 regardless of which release landed first.
Broader context for tech and finance readers
Benchmark leaderboards have quietly become a market-moving genre. They shape enterprise procurement, influence how labs price their models, and increasingly seed prediction markets that let observers put a number on the pace of progress. The Fable 5.1 story shows why the cost line matters as much as the score: a model can top a leaderboard while charging 20% more per task, which changes the calculus for anyone routing production traffic. That tension between capability and unit economics is the same one enterprises weigh when they compare frontier vendors, a theme running through much of the current AI-model coverage.
Prediction markets add a second layer. A contract like this one compresses a diffuse debate about AI progress into a single, checkable outcome, and the volume spike shows how quickly attention concentrates once a real-world event resolves the underlying uncertainty. For a sense of how these benchmark and AI contracts behave on the same platform, our write-ups on the Waymo Marin County expansion market and the Polymarket IPO market track similar dynamics.
What to watch next
Three things over the next four weeks. First, whether Artificial Analysis touches the index version before September 30; that is the only real path to an N/A or surprise resolution. Second, how far past 65 the top score climbs, since a competing release from OpenAI, xAI, or a Chinese lab could push the ceiling higher and shift attention to a new threshold. Third, the cost picture, because the more interesting question after “did it clear 65” is whether the highest scores can be delivered at a price teams will actually pay. Watch the leaderboard on the closing night; the screenshot is what settles the bet.
FAQ
Is the top AI index score already above 65? Yes. As of early September 2026, both the Artificial Analysis model page (66) and the BenchLM snapshot (65.7%) show Claude Fable 5.1 above the threshold on the v4.1.1 scale.
Where does the market trade and for how much? It trades on Manifold in the platform’s play-money currency. There is no real-dollar stake, so the 98% figure reflects crowd conviction rather than a regulated price.
What could still make it resolve NO or N/A? Mainly procedural risk: the score must be published on the leaderboard page at 23:59 UTC on September 30, and if the index version changes, scores must be chain-linked to v4.1.1 using at least three anchor models, or the market resolves N/A.
What is the Artificial Analysis Intelligence Index? A weighted composite of nine hard text-based evaluations, scaled 0 to 100, covering reasoning, coding, science, knowledge, and agentic tasks.
Why did Claude Fable 5.1 move the number? Anthropic released it on September 1, 2026, and it scored 66 at max effort, a four-point gain over Fable 5 and the highest score Artificial Analysis has measured.
The Bottom Line
Sources
- Manifold — Will the top Artificial Analysis Intelligence Index score reach 65 or higher on September 30, 2026?
- Artificial Analysis — Claude Fable 5.1 tops the Artificial Analysis Intelligence Index
- Artificial Analysis — Intelligence Index v4.1.1 methodology and leaderboard
- VentureBeat — Anthropic’s Claude Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for Fable cache reads
- MacRumors — Anthropic launches Claude Fable 5.1 with lower costs and fewer false positives
- BenchLM — Artificial Analysis Intelligence Index leaderboard snapshot (September 2026)
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 with play-money currency. Age restrictions of 18+ or 21+ apply as applicable. If gambling is a problem for you or someone you know, call 1-800-GAMBLER for confidential help.
Related Coverage
- Bitcoin Higher in 7 Days? Traders Flood a Manifold Market at 91% After a 22% Weekly Surge
- Traders Pile Into a Yudkowsky AI-Job Market, but Manifold Odds Stay at 4%
- A Manifold Market Now Puts a Non-MF Group Proof at 70% After Astra's Math Shock
- Manifold Traders Bet 95% on a Non-MF Group Proof This Month After Astra's Math Surge
- Traders Pile Into a Trump Assassination Market as Iran Threats Escalate