Highest Meta score on Humanity’s Last Exam in 2026?

The core question is whether any Meta-published model in 2026 can reach at least 55% accuracy on Humanity’s Last Exam before the year ends. That is a demanding threshold on a benchmark built to separate strong general models from truly exceptional ones, so the challenge is not simply releasing a bigger model but producing a system that improves enough on difficult reasoning, knowledge retrieval, and robustness to clear a fairly high line. With only a few months left in 2026, the market is effectively asking whether Meta already has a near-threshold model in development or can compress a meaningful leap into a short remaining window.

Arguments for Yes are real because Meta has both the resources and the flexibility to try multiple approaches before the deadline. The market only needs the best Meta model published during 2026 to reach the threshold, so an earlier model can fail while a later one succeeds. If Meta ships a larger foundation model, a stronger reasoning-focused variant, or a release with more aggressive post-training and test-time compute, the score could improve nonlinearly rather than gradually. Hard benchmarks also sometimes show sharp gains once a lab aligns training, inference strategy, and data curation well enough, so a late-year release could still surprise to the upside.

Arguments against Yes are stronger to me because 55% feels like a meaningful jump rather than a marginal improvement, and Meta has not historically been the most aggressive benchmark-maximizer on the hardest frontier evaluations. Even if Meta continues improving its models, the remaining time is short relative to the engineering, evaluation, and deployment cycle needed to deliver a system that is clearly above the cutoff. In addition, these very hard exams often resist easy scaling gains, so routine iteration may leave Meta in the high 40s or low 50s without quite crossing the line. The current market price near fifty-fifty reflects the uncertainty, but my read is that the threshold is just hard enough that No deserves a modest edge.