Next Claude Opus: Humanity’s Last Exam Debut?

The market’s central thesis is that the next Claude Opus will score around the low 50s on Humanity’s Last Exam under the configuration recognized by the official leaderboard. That inference explains why thresholds through 50% trade near certainty while 55% receives only a 0.8% Yes price. The dividing line closely tracks Anthropic’s published gap between tool-free and tool-assisted evaluation.

Anthropic’s own scores create the 55% fault line

Anthropic’s Claude Opus 4.8 system card reports 49.8% HLE accuracy without tools and 57.9% with tools. Those results bracket the market’s sharpest division: one sits just below 50%, while the other clears 55% by 2.9 percentage points.

This matters because the next Opus needs only a small improvement over the reported 49.8% no-tools result to reach 50%. Reaching 55% under a comparable configuration requires a larger gain of 5.2 points. The prices therefore imply confidence in incremental progress and deep skepticism that the settlement score will reproduce Anthropic’s tool-assisted result.

That interpretation is a market inference. The system-card figures are sourced facts, while settlement depends exclusively on the HLE accuracy displayed at agi.safe.ai for the next Claude Opus entry. Anthropic’s evaluation does not itself determine the outcome.

The market is implicitly choosing a testing regime

The largest hidden assumption concerns comparability. A benchmark score can depend on tool access, prompting, sampling, answer extraction and the exact model build submitted. The supplied record establishes two Anthropic scores with different tool conditions, yet it does not specify which condition the official HLE entry will use.

A leaderboard entry aligned with the 49.8% no-tools evaluation would validate the market’s low-50s anchor. A tools-enabled entry comparable to 57.9% would directly challenge the 55% pricing. A separately run evaluation could produce another result altogether, especially if the next Opus differs materially from the version covered by the system card.

The hierarchy also assumes that model progress will transfer to HLE. A new Opus release may improve coding, agentic tasks or product reliability without adding five points on this benchmark. Conversely, benchmark-specific reasoning gains could move HLE faster than broader capability measures suggest. Product branding alone provides limited evidence about either path.

The lower thresholds reveal confidence, plus pricing noise

The 35%, 40%, 45% and 50% contracts carry Yes prices of 99.3%, 95.9%, 96.8% and 98.1%, respectively. Their order is logically inconsistent: any score clearing 50% must also clear every lower threshold. The reversals between 40%, 45% and 50% mean the individual quotes cannot be treated as a single clean probability distribution.

This is the main counter-signal to reading too much precision into the hierarchy. With $23.79K in volume, $18.42K in liquidity and $8.15K in open interest across the event, separate contract conditions or stale quotes can influence the visible ordering. The broad message remains a cliff above 50%; the exact differences among lower bands carry weaker evidentiary value.

Official publication will overwhelm indirect capability evidence

The decisive catalyst is the first appearance of the next Claude Opus on the official HLE results site. Under the resolution criteria, its displayed accuracy at 12:00 PM ET on the following calendar day controls settlement. That next-day checkpoint makes corrections, relabeling or score updates during the initial publication window materially relevant.

Before publication, an Anthropic launch announcement or system card could shift expectations if it reports HLE performance and states the evaluation setup. A result above 55% under conditions matching the official leaderboard would weaken the market’s configuration thesis. A tool-free score near 50%, or explicit evidence that the leaderboard excludes tools, would strengthen it. Scores near a threshold would also make rounding and the site’s displayed precision important.

Timing remains an operational failure mode

The market closes on December 31, 2026 at 11:59 PM UTC, while the supplied criteria define settlement through the next model’s leaderboard appearance. The record provided here does not state how a failure to publish a new Opus entry before closing would be handled. That leaves launch timing and rules interpretation as secondary catalysts.

The strongest substantive failure mode is simpler: the market may be anchoring too heavily on the 49.8% no-tools figure. If the official entry follows Anthropic’s 57.9% tools-enabled setup, or the next generation adds more than five points under a comparable no-tools test, the apparent 55% barrier would lose its empirical basis.

Sources