Will the highest score achieved by a SpaceXAI Grok model on Humanity’s Last Exam in 2026 be 45% or higher?

- [2026-10-10] The official resolution source, agi.safe.ai's Humanity's Last Exam leaderboard, currently lists only one xAI entry in its quantitative-results table: "Judge Model: o3-mini | Dataset Updated: April 3rd, 2025 | Model | Accuracy (%) ↑ | CalibrationError (%) ↓ | Grok 4 logo Grok 4 | 24.5 | 56.4" — i.e. the official site still shows "Grok 4" at 24.5%, far below the 45% threshold, with no newer Grok variant listed. (https://agi.safe.ai/) - [2026-09-22] agi.safe.ai's own news feed shows its most recent update is the release of a new "HLE-Diamond" variant, not a new Grok score: the site's latest news items list "[09/22/2026]: HLE-Diamond is released" and no new Grok submission in the news log. (https://agi.safe.ai/) - [2025-07-09] xAI's own launch blog claims (self-reported, not on the official leaderboard) that "Grok 4 Heavy leads USAMO'25 with 61.9%, and is the first to score 50.7% on Humanity's Last Exam (text-only subset)" , while the tool-using configuration was reported at 44.4%; these figures have never appeared as an official agi.safe.ai leaderboard row. (https://x.ai/news/grok-4) - [2026-10-09] A third-party benchmark aggregator (Artificial Analysis data, via orcarouter.ai) reports the current flagship model, Grok 4.7 (released Sept 21, 2026), at "Humanity's Last Exam at 43.1%" — below 45% and from an independent tracker, not the official CAIS/Scale leaderboard. (https://orcarouter.ai/blog/step-5-preview-vs-grok-4-7) - [2026-10-05] Independent model-tracking coverage confirms xAI's current lineup as of early October 2026 is Grok 4.5/4.6/4.7 (plus a "4.20" preview build), with Grok 5 explicitly not yet released: "No — not as of 5 October 2026... Grok 4.6 shipped on 12 August 2026, and grok-4.7 is now xAI's most capable model" . (https://codersera.com/blog/grok-4-5-launch-guide-2026/) - [2026-09-23] A dedicated Grok 5 tracker confirms the next-generation model remains unreleased and un-benchmarked: "As of September 2026, Grok 5 has not been released. xAI's next flagship model is still in training, and the only thing the company has officially said about it is a single line in its January 2026 funding announcement confirming the model exists and is being trained. There is no model card, no spec sheet, and no public benchmark." (https://geotoolbox.ai/blog/grok-5) - [2026-10-08] A benchmark aggregator snapshot (BenchLM, mirroring Scale's public table as of Sept 30, 2026) places the newest tracked xAI model, "Grok 4.7," well below 45%: "Grok 4.7 xAI · Closed 23.4%" in its 12-row leaderboard snapshot, with other labs' models (e.g., GPT-6 Astra at 60.6%) leading. (https://benchlm.ai/benchmarks/scale-hle-diamond) - [2025-07-11] At the time of the Grok 4 Heavy claims, outside observers noted the gap between company marketing and the official leaderboard: "The results from xAI's internal testing have yet to appear on the leaderboard for HLE... it remains unclear whether this is because xAI has yet to submit the results or because those results are pending review." (https://www.scientificamerican.com/article/elon-musks-new-grok-4-takes-on-humanitys-last-exam-as-the-ai-race-heats-up/) - [2026-09-02] A related prediction market tracking whether "Grok 4 Heavy" would ever formally appear on the agi.safe.ai leaderboard resolved NO, with the market's own description noting resolution would be NO if "The next major iteration of Grok models is released by xAI without Grok 4 Heavy being generally accessible... in the official xAI API. Examples include Grok 4.2, Grok 4.5, Grok 5" — confirming Grok 4 Heavy's high self-reported score never became an official leaderboard entry. (https://manifold.markets/Bayesian/grok-4-heavy-gets-on-humanitys-last) - [2026-10-08] Independent intelligence-index tracking (Artificial Analysis, via themodelpress.com) shows the current HLE frontrunner across all labs is Anthropic's Claude Opus 5.5 at 61.4%, not any Grok model, underscoring that Grok has fallen behind rivals on this metric in 2026. (https://themodelpress.com/news/2026/10/claude-opus-5-5-intelligence-index) BOTTOM LINE: As of October 10, 2026, the official agi.safe.ai leaderboard — the stated resolution source — still shows only a single, outdated "Grok 4" entry at 24.5% accuracy, and no Grok model (through the current Grok 4.7) has an official score anywhere near 45%; independent trackers put the newest Grok variants in the low-to-mid 20s–43% range, and Grok 5 remains unreleased and unbenchmarked. NOT FOUND: No evidence that xAI has ever formally submitted Grok 4 Heavy's self-reported 44.4%/50.7% figures (or any later Grok model's score) to the official agi.safe.ai leaderboard; no confirmed official leaderboard update for any Grok 4.x model beyond the original "Grok 4 = 24.5%" row; no release date or benchmark data for Grok 5.

- [2025-07-09] xAI's own press release claimed Grok 4 scored 25.4% (no tools) and Grok 4 Heavy 44.4% (with tools) on HLE at launch, with a later-cited "text-only subset" figure of 50.7% — the first time any model's self-reported HLE number crossed 45%. https://x.ai/news/grok-4 - [2025-07-11] Despite xAI's self-reported 44.4%/50.7% figures, independent benchmark trackers (e.g., Artificial Analysis) measured Grok 4 far lower on HLE; Zvi Mowshowitz noted "Artificial Analysis only gave Grok 4 a 24% on HLE, versus the 44% claimed above, which is still an all-time high score but much less dramatically so." This shows a recurring gap between xAI marketing numbers and third-party/official measurement. https://thezvi.substack.com/p/grok-4-various-things - [2025-09 (resolved)] A Manifold Markets contract asking whether "Grok 4 Heavy has a score in the leaderboard section of https://agi.safe.ai/, regardless of settings (text-only, tools-allowed, etc.)" resolved NO, confirming that xAI's headline HLE claims for Grok 4 Heavy were never actually posted to the official CAIS/Scale leaderboard — the resolution source this Polymarket question depends on. https://manifold.markets/Bayesian/grok-4-heavy-gets-on-humanitys-last - [2026-01 (reported)] Elon Musk publicly stated "Even Grok 4, which is primitive at this point, got, I think, 52%, excluding visual questions." — another unofficial/self-reported figure above 45%, again not tied to a confirmed official leaderboard entry. https://www.latestly.com/socially/technology/grok-5-release-in-2026-elon-musk-says-ai-could-be-nearly-perfect-on-humanitys-last-exam-7265812.html - [2026-08-10] A sibling Polymarket market on whether the next Grok model (4.6+) would debut with an HLE score ≥35% showed the current leading outcome is "35%+" at 100%, followed by "40%+" at 0% , implying market consensus that Grok 4.6's debut score landed in the 35–40% band, below the 45% threshold in this question. https://polymarket.com/event/next-grok-model-4pt6-humanitys-last-exam-debut-20260810180408165 - [2026-08-16] An analysis of the official public leaderboard concluded "As of August 16, 2026, no model listed on the official public leaderboard has reached 50%; future performance cannot be inferred from this benchmark trajectory alone." and separately noted the public leaderboard's leading score was 46.44% (not specific to Grok). https://intuitionlabs.ai/articles/humanitys-last-exam-ai-benchmark - [34 days old, ~2026-09] A Wikipedia-cited AI Index-style comparison table lists xAI's newest entry, "Grok 4.6," at 42.9% — below several rivals (GPT-6 Astra 54.7%, Gemini 3.8 Flash 47.8%, Kimi K3 46.9%) and below the 45% bar for this question. https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam - [2026-09-30] A third-party mirror of the "Scale score" HLE snapshot lists "Grok 4.7 xAI" at only 23.4% in one table (likely a harder/HLE-Diamond-style variant), underscoring how scores vary enormously (20s to 50s%) depending on which leaderboard/variant/tool setting is used for the same Grok generation. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-07-23 to present] No source found confirms any SpaceXAI/xAI Grok model officially listed at ≥45% on the specific agi.safe.ai leaderboard (the named resolution source) during 2026; all ≥45% Grok figures found (50.7%, 52%) are self-reported/press-release or hearsay numbers, historically shown not to transfer directly onto the official leaderboard. - [Base-rate pattern] Across the 2025–2026 history of this benchmark, xAI has repeatedly announced self-reported HLE numbers exceeding 45% (Grok 4 Heavy 44.4–50.7% in mid-2025; Musk's 52% claim in Jan 2026) faster than any other lab's cadence, but the independently verified/official leaderboard entries for Grok models have consistently landed materially lower (mid-20s to low-40s%), a gap that has persisted for over a year. BOTTOM LINE: The reference class shows a recurring and durable gap between xAI's press-released HLE numbers (which have exceeded 45% at least twice since mid-2025) and what actually gets posted to the official agi.safe.ai leaderboard that governs this market, where Grok entries have so far topped out in the low-to-mid 40s%; this base rate favors continued "No" outcomes unless xAI specifically submits a qualifying high score to the official leaderboard before year-end. NOT FOUND: No confirmed, dated entry on the live agi.safe.ai leaderboard itself (the page required sign-in and could not be rendered) showing any current Grok model's official listed score; no official xAI/CAIS statement from the last 30 days confirming a ≥45% Grok entry on that specific leaderboard; no explicit historical count of how many times (if any) a Grok model has previously appeared on the official leaderboard at all.

YES means: before Dec 31, 2026 11:59 PM ET, some SpaceXAI/xAI Grok model attains an official "HLE Accuracy" (or clearly equivalent) score of ≥45% on the agi.safe.ai leaderboard (or an approved alternative if that site is down), as read by the market's resolution criteria. - [2026-07] SpaceX's absorption of xAI, completed by February 2026 and renamed in July 2026, means the question's "SpaceXAI" entity is simply the current corporate name for the Grok developer, so any qualifying Grok score (Grok 4.x, 4.20, or a future Grok 5) counts toward resolution; this is confirmed independently by Wikipedia's chatbot entry listing "Developers: SpaceXAI Cursor" for the Grok line. (https://en.wikipedia.org/wiki/Grok_(chatbot)) - [2025-07-09] xAI's own launch blog already demonstrated a Grok variant crossing the 45% bar once, with Grok 4 Heavy "the first to score 50.7% on Humanity's Last Exam (text-only subset)," showing the 45% threshold is not architecturally out of reach for an aggressive multi-agent/tool configuration. (https://x.ai/news/grok-4) - [2026-01] Elon Musk separately claimed a Grok 4 (not even the then-newest model) scored "52%, excluding visual questions," a second independent instance of an xAI-affiliated figure publicly asserting a sub-model crossed 45%, indicating company messaging keeps reaching for and claiming this bar even if not yet officially posted. (cited in prior reports, https://www.latestly.com/socially/technology/grok-5-release-in-2026-elon-musk-says-ai-could-be-nearly-perfect-on-humanitys-last-exam-7265812.html) - [2026-09-22] The CAIS/Scale team does actively and recently update official HLE data tied to new Grok releases — a Grok 4.7 entry on the HLE-Diamond variant is dated to the CAIS/Scale blog itself on 2026-09-22, confirming there is a live, working pipeline for new xAI submissions to reach an official-adjacent CAIS/Scale source within days of a model launch, not months. (https://klu.ai/benchmarks/humanitys-last-exam) - [2026-10-10] The live agi.safe.ai site's news feed shows continued active curation (e.g., the Sept 22, 2026 "HLE-Diamond is released" entry), indicating the resolution source is operationally maintained and capable of posting a new, higher Grok score before the Dec 31 deadline if one is submitted. (https://agi.safe.ai/) - [2026-10-08] A SpaceXAI-focused industry roundup confirms the renamed entity remains a top-tier active lab still shipping models monthly (Grok 4.7 on Sept 21, 2026) rather than stagnating, preserving multiple more release windows (4.8, 4.9, and ultimately Grok 5) before year-end in which a qualifying score could appear. (https://slash-digital.io/en/insights/ai-models-2026/) - [2026-09-30] Grok 5 remains unreleased but is confirmed "in training" on the Colossus 2 supercluster with a reported 6–10 trillion parameter scale, roughly double-to-triple Grok 4's size; if xAI follows its historical pattern of large generational jumps (Grok 3→4 added roughly 10x compute per contemporaneous reporting), a Grok 5 release before Dec 31 could plausibly jump well past the mid-40s% band where Grok 4.6/4.7 currently sit. (https://geotoolbox.ai/blog/grok-5) - [2026-10-08] Independent trackers show rapid frontier movement on HLE across labs in just the past month (Claude Opus 5.5 at 61.4%, GPT-6 Astra near 54.7–60.6%), establishing that scores in the 45-60% range are now achievable with current-generation techniques, so a comparably scaled Grok 5 reaching ≥45% would not require an unprecedented leap relative to contemporaries. (https://themodelpress.com cross-referenced via https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam) - [2026-09-14] A September 14, 2026 X reply from Musk ("That will be Grok 5") is the most recent on-the-record signal that the model is advancing toward release, keeping open the possibility of a Q4 2026 launch inside the market's window, even though no date or benchmark has been confirmed. (https://geotoolbox.ai/blog/grok-5) BOTTOM LINE: The clearest YES path is a Grok 5 (or a surprise high-tool "Heavy"-style Grok 4.8/4.9 variant) launching before Dec 31, 2026 with an HLE score that xAI submits to, or CAIS/Scale independently posts on, the official leaderboard; this requires both an on-time release (which has slipped repeatedly) and the submitted number landing above 45% rather than in the 20s–43% band seen in every official Grok entry to date — a precedent (Grok 4 Heavy's 50.7%/44.4%) shows xAI is capable of generating such scores, but has never gotten them onto the official leaderboard itself. NOT FOUND: No confirmed Grok 5 release date, benchmark card, or HLE score as of Oct 10, 2026; no evidence xAI has a pending official leaderboard submission above 45% for any current Grok model; no statement from CAIS/Scale about plans to post a non-Diamond, full-HLE Grok score before year-end.

YES means: some SpaceXAI/xAI Grok model's score, as listed on the official agi.safe.ai Humanity's Last Exam leaderboard (or an official successor source), reaches at least 45% accuracy at any point before December 31, 2026, 11:59 PM ET. - [2026-10-06] The only place a Grok model clears 45% is a third-party aggregator (llm-stats.com), which lists "Grok-4 Heavy xAI" at 0.507 — but this figure is flagged as one of "0 verified results and 104 self-reported results" on that tracker, not an independently-confirmed or CAIS-run score, and it still is not the official agi.safe.ai page named in the resolution criteria. https://llm-stats.com/benchmarks/humanity's-last-exam - [2026-09-22] On the actual CAIS/Scale-run benchmark (the "HLE-Diamond, no tools" table, sourced to the official CAIS and Scale blog), the newest xAI model, Grok 4.7, scored just 22.8%, dead last among the nine flagships tested, nowhere near 45%. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-06] On Scale Labs' own Diamond leaderboard snapshot, Grok 4.7 again sits near the bottom at 23.4% overall, confirming the low score is reproduced across two independent CAIS/Scale-linked tables, not a one-off. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-06] Even on Scale's full-set (non-Diamond) leaderboard — the closest proxy to the exact format behind the 24.5% figure already on agi.safe.ai — no Grok model appears in the top ten shown (which bottoms out at Claude Opus 4.6's 34.44%), implying any current Grok full-set score sits below roughly 34%. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-06] The field has moved further away from Grok, not closer: Claude Opus 5.5 leads Anthropic's own harness at 67.7% and GPT-6 Astra leads Scale's full board at 54.80%, widening the gap Grok would need to close even if it is later submitted and scored. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-09] Grok 5 — the one plausible vehicle for a step-change score — remains unreleased with "no model card, no spec sheet, and no public benchmark," and even xAI's own incremental cadence has run through 4.7/4.8/4.9 without reaching 45% on any officially benchmarked table. https://geotoolbox.ai/blog/grok-5 - [2026-08-13] An independent Grok-5 tracker confirms that as late as mid-August 2026 "the xAI model list at docs.x.ai still tops out at grok-4.6," underscoring how much of 2026 has already elapsed with xAI's flagship stuck in the 4.x line rather than a fresh model capable of resetting the HLE conversation. https://www.orcarouter.ai/blog/grok-5-release-date - [2025-07-11] The precedent is direct: a Manifold market asking specifically whether Grok 4 would "debut on HLE's leaderboard with a 45 percent score or greater... within a month of its release" was priced by the market at only "a 1 percent chance," and it in fact never happened — xAI's claimed 44.4% never became an official leaderboard row. https://www.scientificamerican.com/article/elon-musks-new-grok-4-takes-on-humanitys-last-exam-as-the-ai-race-heats-up/ - [2026-10-10] As of the fetch date, the official agi.safe.ai site (per the resolution source itself) still shows a single outdated "Grok 4 | 24.5%" entry, meaning the literal resolution mechanism has zero newer Grok rows to even evaluate against 45% — a pure status-quo/inertia case for NO, since xAI has shown no pattern of submitting updated scores to this specific table in over a year. - [2026-10-10] Even if CAIS someday scores a newer Grok model, the trend across every CAIS/Scale-linked table found (HLE-Diamond no-tools 22.8%, Diamond overall 23.4%) is that Grok's relative standing has fallen as rivals' scores have risen, making a jump to 45%+ before December 31, 2026 require a dramatic reversal rather than a continuation of trend. BOTTOM LINE: Under every officially-sourced or CAIS/Scale-run measurement found (as opposed to xAI's own marketing claims), current and recent Grok models score in the low-to-mid 20s, the official agi.safe.ai page itself still shows only the stale 24.5% "Grok 4" row, Grok 5 remains unreleased and unbenchmarked, and a direct historical precedent (the Grok 4 Heavy 44.4%/50.7% claims never posting to the official leaderboard) shows that even a crossing of the 45% line in xAI's own PR has not translated into a qualifying official entry — all pointing toward NO absent a sudden, unprecedented leaderboard submission before year-end. NOT FOUND: No evidence of any pending or scheduled xAI submission of a newer Grok score to agi.safe.ai specifically; no official agi.safe.ai screenshot/table dated after the April 2025 "Grok 4 = 24.5%" row; no confirmed Grok 5 release date or any Grok 5 HLE benchmark figure, official or self-reported.