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

The key fact is that Grok is already operating near the cutoff rather than far below it. Recent public benchmark reporting places Grok 4.6 around 42.9% and Grok 4.7 around 43.1% on Humanity’s Last Exam, which means the model family has essentially closed most of the gap to 45% and only needs a modest additional gain. That is meaningfully different from a market where the model is sitting in the 20s or low 30s, because a few percentage points of improvement is within the range of what a new release, better prompting, or a different evaluation configuration could plausibly deliver.

At the same time, the most relevant standard scores still appear to be below the threshold, and the distinction between standard scoring and with-tools scoring is critical. Some reported figures above 45% come from with-tools or otherwise non-equivalent setups, while the market explicitly resolves on the official HLE accuracy metric or a clear equivalent. If the year ends without a clean standard leaderboard update above 45%, then even strong auxiliary results will not matter. That means the path to Yes is not just general model progress, but a specific public result on the right metric before the deadline.

The market price implies a much higher probability than my estimate, so I am implicitly fading the idea that a threshold-crossing is nearly certain. The bullish case is real because the gap is thin and xAI has shown the ability to iterate quickly on flagship models. The bearish case is that benchmark improvements often slow as models approach higher-performance territory, and the last few points can be much harder than the first few. On balance, I see Yes as more likely than No, but not overwhelmingly so, because the market needs one more measurable step and there is still enough uncertainty around timing, evaluation conditions, and whether the next public score actually clears 45%.