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

- [2026-10-18] The agi.safe.ai page, fetched today, has a static results table headed "Judge Model: o3-mini | Dataset Updated: April 3rd, 2025". The only Grok row is Grok 4 at 24.5% accuracy and 56.4% calibration error. The table's top row is Gemini 3 Pro at 38.3%. The page also loads a separate "Loading HLE results…" widget that I could not read. https://agi.safe.ai/ - [2026-09-22] The agi.safe.ai news list shows HLE-Diamond released on 09/22/2026. It is described as a migration path for when frontier models hit the noise ceiling on the original HLE. The list also shows HLE-Rolling (10/08/2025) and the Nature publication (01/28/2026). If the official board is restructured, a "clear equivalent metric" clause could matter. https://agi.safe.ai/ - [2026-09-22] Wikipedia cites Artificial Analysis (text-only subset, 22 Sep 2026). It lists Grok 4.6 at 42.9%, the highest xAI entry there. It lists Claude Opus 5.5 at 61.4%, Gemini 4 Argon at 57.1%, GPT-6 Astra at 54.7%, Muse Spark 1.3 at 48.7% and Kimi K3 at 46.9%. https://en.wikipedia.org/wiki/Humanity's_Last_Exam - [2026-10-12] A BenchLM mirror of Scale's HLE text-only board (64 models) has GPT-6 Astra at 54.2% on top. A 30 Sep snapshot of Scale's main board has Grok 4.7 at 23.4% and GPT-6 Astra at 60.6%. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-09-21] Grok 4.7 is the newest Grok model, released 21 Sep 2026. Artificial Analysis scores it 46 on its Intelligence Index, 2 points above Grok 4.6. I found no HLE figure for it from SpaceXAI. https://en.wikipedia.org/wiki/Grok_(chatbot) - [2026-08-12] Grok 4.6 was released on 12 Aug 2026. Its Artificial Analysis Intelligence Index score was 44. https://greenflagdigital.com/top-ai-models-ranked/ - [2026-09-30] Grok 5 has not been released. xAI has said only that it is "in training" (Jan 2026 funding post). Musk's dates have slipped repeatedly: "4 weeks" on Jul 24, "10 days" on Sep 1, and "within 2026" on Aug 5. His 14 Sep roadmap puts Grok 4.8 and 4.9 ahead of it, with no dates. https://www.izzedo.chat/blog/grok-5 - [2026-09-29] A tracker reports Grok 5 is still in training with no model card or benchmark. It describes xAI's 4.7-to-4.9 roadmap as slipping. https://geotoolbox.ai/blog/grok-5 - [2026-10-16] Leading HLE scores elsewhere are far above 45%. BenchLM shows Claude Fable 5.1 at 65%, Claude Opus 5 at 64.7% and Muse Spark 1.1 at 62.1%. Grok is not in the top 11 on that list. https://benchlm.ai/benchmarks/hle - [2026-10-06] Three boards show different HLE top scores. Scale has GPT-6 Astra at 54.8%, Artificial Analysis has Claude Opus 5.5 at 61.4% on text-only questions, and HLE-Diamond has GPT-6 Astra at 59.9% without tools. https://iqmetrics.org/news/humanitys-last-exam-scores-what-the-leaderboards-show/ - [2025-07-09] Historical context: xAI's Grok 4 Heavy was reported at 44.4% on the full HLE with tools, and 50.7% on the text-only subset. The agi.safe.ai board never shows these numbers. https://x.ai/news/grok-4 - [2026-09-24] Musk said GPT-6-level overall capability is still 2–3 months away. He also said SpaceX would reach "pole position" in about 6 months. https://www.tikr.com/blog/musk-claims-rank-1-for-spacex-as-grok-tops-gpt-6-in-cyber-defense BOTTOM LINE: The official agi.safe.ai board is stale and shows only Grok 4 at 24.5%, so a Yes would need an update that adds a Grok row of at least 45%. Third-party boards put Grok 4.6 at about 43% (Artificial Analysis) and Grok 4.7 at 23% (Scale). A Grok 5 or 4.8/4.9 release before 31 Dec is the main path to ≥45%, and none has a date. NOT FOUND: The contents of the dynamic "HLE results" widget on agi.safe.ai (it did not render). Any official SpaceXAI HLE number for Grok 4.7. Any scheduled release date for Grok 4.8, 4.9 or 5. Any evidence that the agi.safe.ai board has added a Grok model newer than Grok 4.

- [2026-10-18] The official page agi.safe.ai has a static table headed "Dataset Updated: April 3rd, 2025". The best Grok on it is Grok 4 at 24.5%, and the top row is Gemini 3 Pro at 38.3%. The page also has a JavaScript "Loading HLE results" element that I could not read, so I could not check whether it shows newer Grok entries. The latest news item is the 09/22/2026 HLE-Diamond release. The resolution source therefore shows no Grok at or above 45% in what I could read. https://agi.safe.ai/ - [2026-10-01 to 2026-10-18] Wikipedia's HLE results table, which I did not fetch in full, lists xAI Grok 4.6 at 42.9%. Other entries are Kimi K3 46.9%, Muse Spark 1.3 48.7%, GPT-6 Astra 54.7%, Gemini 4 Argon 57.1% and Claude Opus 5.5 61.4%. Grok 4.6 is the best Grok score I found on any aggregator. It is about 2.1 points below the 45% threshold. https://en.wikipedia.org/wiki/Humanity's_Last_Exam - [2026-09-21] SpaceXAI released Grok 4.7. It uses a new, larger base model with a longer RL run. The launch post reports no HLE score. It gives coding, office-work and safety benchmarks only, such as CursorBench, GDPval and Terminal-Bench. I found no HLE figure for 4.7 in any source. https://x.ai/news/grok-4-7 - [2026-09-21] The release cadence is Grok 4.5 on 2026-07-08, Grok 4.6 on 2026-08-12 and Grok 4.7 on 2026-09-21, roughly one flagship every 5–6 weeks. The API release notes confirm this. At that pace another one or two releases (4.8, 4.9) could land before 2026-12-31. https://docs.x.ai/developers/release-notes - [2026-09-14] Musk posted that Grok 4.8 "will be a noticeable improvement" and that Grok 4.9 is "probably Astra/Fable class". Astra and Fable are the GPT-6 and Claude tier, which score about 55–59% on HLE. This is a Musk claim, not a result. https://www.izzedo.chat/blog/grok-5 - [2026-09-29] Grok 5 is still unreleased and in training. Musk's earlier targets (late 2025, Q1 2026 and Q2 2026) were all missed. An aggregator reports that Polymarket's "released by Dec 31, 2026" market priced it between the low 40s and 80s, and that coverage was mixed. This is a separate market with its own date. A Grok 5 release is the most likely route to a big HLE jump. https://geotoolbox.ai/blog/grok-5 - [2026-07-08] Grok 4.5 launched. Musk called it roughly comparable to Opus 4.7, so it was positioned just below best-in-class. https://techcrunch.com/2026/07/08/spacexai-releases-grok-4-5-which-elon-describes-as-an-opus-class-model/ - [2026-10-16] The frontier is well above 45%. Aggregators show Claude Fable 5.1 at 59.1% (pricepertoken) or 65% (BenchLM). Scale's text-only board on 2026-10-12 has GPT-6 Astra at 54.2%. The 45% bar is a mid-pack frontier score. Grok trails Kimi K3, MiMo-V2.6-Pro and Muse Spark. https://pricepertoken.com/leaderboards/benchmark/hle - [2025-07-09] History for xAI: Grok 4 launched with 25.4% (no tools), 38.6% (with tools) and 44.4% (Heavy, full set, with tools). xAI's own claims were much higher than independent or official-board numbers, which put Grok 4 at 24.5%. The gap between xAI self-reports and the agi.safe.ai board is large. https://x.ai/news/grok-4 - [2026 early] Grok 4.20 appears on aggregators at 50.7%. That figure matches Grok 4 Heavy's text-only claim and is probably a mislabelled or self-reported variant. This shows scores differ by variant and source: text-only vs full set, tools vs no tools. https://aitooltier.com/benchmarks/humanitys-last-exam - [Reference class] Successive xAI flagships have gained about 3–10 points of HLE per release. Examples are Grok 4.6 at 42.9% and the 4.5 to 4.6 step, though I did not find the 4.5 figure. Other labs moved from the mid-30s to the 55–60s in about 6 months. A gain of 2.1 points over Grok 4.6 is smaller than typical per-release jumps. The main uncertainty is whether the official agi.safe.ai board gets updated with newer Grok models. - [2026-09-22] HLE-Diamond was released as a migration path for when frontier models hit the noise ceiling on the original HLE. The official site may shift emphasis or metrics, which creates some resolution-source ambiguity. https://agi.safe.ai/ BOTTOM LINE: The best Grok score I found is 42.9% (Grok 4.6, Wikipedia and aggregators). That is about 2 points short of 45%, and one or two more Grok releases are likely by year-end, so a threshold-clearing score from a new model is plausible. The main risk to a Yes is how the official agi.safe.ai board is read. Its readable table is stale (best Grok 24.5%), and the Grok 4.7 launch post gave no HLE number. NOT FOUND: An HLE score for Grok 4.7 or 4.5. The contents of the dynamic agi.safe.ai leaderboard. Any official CAIS or Scale entry for Grok 4.6. Any expert forecast specific to Grok's HLE score.

YES means: some SpaceXAI Grok model is shown with an HLE accuracy of 45% or higher on the official agi.safe.ai leaderboard, or a clear equivalent official HLE result, by 2026-12-31. - [2026-10-06] Klu's HLE page says there is no single HLE leaderboard in 2026. It lists at least four official or semi-official boards, each with a different leader and scoring protocol: Anthropic's own runs, Artificial Analysis (text-only, 2,158 questions), Scale (full set), and HLE-Diamond. The resolution text allows "a clear equivalent metric" if the presentation is restructured. A Grok score of 45% or more on any official board, such as a CAIS/Scale refresh, gives a Yes path that does not depend on the stale table. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-06] Frontier labs' own harnesses give much higher numbers than the Scale/CAIS judge. Anthropic reports Opus 5.5 at 64.4% without tools and 67.7% with tools on the full set. HLE-Diamond with tools has GPT-6 Astra at 82.9%. If CAIS adds a tools column for Grok, 45% is easy to clear. Grok 4 Heavy already claimed 44.4% with tools in 2025. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-13] The nearest-in-time rivals show the size of a generational jump. DeepSeek V4.1 Flash went from 36.8 to 42.6 on no-tools HLE in one update (Opus 5 at 56.3 on the same chart). Step 5 Preview, an open-weights model, scores 46.5% on Artificial Analysis. Grok 4.6's 42.9% is a mid-pack score, so a modest gain on the next Grok clears 45%. https://xenospectrum.com/en/deepseek-v41-flash-1017-comparison - [2026-10-16] Step 5 Preview (open weights, 27B active parameters) hits 46.5% on HLE with an Intelligence Index of only 43.7. Grok 4.7 has an index of 46, higher than that model. By index it should score at or above Step 5 on Artificial Analysis HLE, which would put it over 45%. No Grok 4.7 HLE figure is published, so this is an inference. https://marktechpost.com/2026/10/16/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context - [2026-10-13] Artificial Analysis lists Grok 4.7 at 46 against Grok 4.6's 44 on its Intelligence Index. The post also lists Claude Opus 5.5 at 58. https://orcarouter.ai/blog/muse-wonder-vs-grok-4-6 - [2026-07-09] A Grok 4.5 HLE score of 0.40 is recorded on Artificial Analysis. Grok 4.6 is at 42.9%. Grok 4.7 has an Artificial Analysis HLE score I could not find. Extrapolating about 2–3 points per release from 40 to 42.9 gives roughly 45% for Grok 4.7, and more for 4.8/4.9. https://agiranker.com/ - [2026-09-14] Musk's roadmap rates Grok 4.8 as a "noticeable improvement". It is a 2.5T-parameter model on a new C++ stack, finishing pretraining the week of 14 Sep before RL. Cursor's server model list already contains a grok-4.8 family (commit seen 28 Sep). That is early evidence of an approaching release, but no xAI page, price or date exists as of 13 Oct. Grok 4.9 is described as "probably Astra/Fable class", which scores above 55%. https://cellcog.ai/blog/grok-4-8-release-date/ - [2026-10-13] No Grok 4.8 launch is confirmed in the official model catalog or release notes. Grok 4.7 is the documented flagship. xAI's observed cadence of 5–6 weeks (8 Jul, 12 Aug, 21 Sep) puts the next model in roughly the week of 2 Nov, well before 31 Dec. https://apimaster.ai/blog/grok-4-8-api - [2026-09-22] Grok 4.7 scores only 22.8% on HLE-Diamond with no tools, per the CAIS/Scale blog as cited by Klu. Scale's text-only board has Grok at 23.4%. That is a headwind. It does show that CAIS/Scale now run each new Grok within days of release (Grok 4.7 was scored on its release date). A new Grok would be added to the official CAIS/Scale tables quickly. The yardstick is harsher, though, so the Yes path needs a large jump, or tools. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-18] The agi.safe.ai page is a static snapshot, showing only Grok 4 at 24.5% with the o3-mini judge. A crawl of the page described it as updated April 2025, though the news list has HLE-Diamond (09/22/2026). A "Loading HLE results" widget loads separately, and CAIS also points to the CAIS AI Dashboard and SEAL leaderboards. If the question writer reads a dynamic widget that includes newer models, the answer could come from there. I could not render it. https://agi.safe.ai/ - [2025-07-12] Manifold put the chance that Grok 4 would appear on the official board at 45% or more within a month at 1%, and xAI claimed 44.4%. Scientific American noted xAI's results were not on the official leaderboard. The official board has historically lagged xAI's own claims, so a Yes needs a model that clears 45% under the official judge. https://www.scientificamerican.com/article/elon-musks-new-grok-4-takes-on-humanitys-last-exam-as-the-ai-race-heats-up/ BOTTOM LINE: The most likely path to Yes is a Grok 4.8/4.9 or Grok 5 release before 31 Dec, with a ≥45% score published or mirrored on an official CAIS/Scale HLE board. Grok 4.6 is only about 2 points short at 42.9% on Artificial Analysis (a third-party board), and rivals like Step 5 Preview, at a similar index level, already post 46.5%. The main obstacle is the official scoreboard, where Grok sits at 22–25%. NOT FOUND: A Grok 4.7 HLE score on Artificial Analysis or from xAI. The contents of the agi.safe.ai dynamic widget. Any dated release for Grok 4.8, 4.9 or 5. Any CAIS statement that a with-tools board will be added.

YES means: a SpaceXAI Grok model shows at least 45% "HLE Accuracy" on the official agi.safe.ai leaderboard (or a clear official equivalent) by 2026-12-31 11:59 PM ET. - [2026-10-18] I rendered agi.safe.ai in a browser. The readable content is the news list, the dataset text and the contributor lists. The earlier static table (Grok 4 at 24.5%, dated April 2025) is the only Grok row I could confirm. The page has no visible new-model submission process. It has been essentially frozen for about 18 months while many frontier models shipped, so status-quo inertia favours No. https://agi.safe.ai/ - [2025-07 / Scientific American] Precedent: xAI's 44.4% Grok 4 Heavy claim never reached the official board, which instead shows 24.5%. The same article noted that Manifold priced a 45%+ debut on the HLE leaderboard within a month of release at 1%. Self-reported or tool-assisted numbers do not count. Even if a Grok 4.8 launch post carries a 45%+ number, the resolution source may not carry it. https://www.scientificamerican.com/article/elon-musks-new-grok-4-takes-on-humanity-s-last-exam-as-the-ai-race-heats-up/ - [2026-10-06 to 2026-09-22] The official CAIS and Scale HLE-Diamond board, the one actually being maintained, scores Grok 4.7 at 22.8% (no tools) on 2026-09-22 and 23.4% on Scale's Diamond board. That is about half the 45% bar and about 22 points below the leader. xAI's newest model is therefore far from the threshold on the CAIS-affiliated board. If the question migrates to Diamond or the "clear equivalent" is read as Diamond, a Grok jump of more than 21 points in about 10 weeks would be needed. https://klu.ai/benchmarks/humanitys-last-exam - [2026-09-21] The Grok 4.7 launch post (fetched) lists coding, office-work, legal, clinical and safety benchmarks and has no HLE figure. SpaceXAI is not leading with HLE in launches. That lowers the chance of a clean official self-report that CAIS would then ingest. https://x.ai/news/grok-4-7 - [2026-07-09] Grok 4.5 scored about 40% on the Artificial Analysis HLE measure (aggregator agiranker). Grok 4.6 reached 42.9% on that board, a gain of roughly 3 points over about 5 weeks. A third-party 45%+ requires another 2.1 points and a new model. Even the best-case aggregator path is not guaranteed, because 4.7 has no HLE figure and only a 2-point Intelligence Index gain. https://agiranker.com/ - [2026-10-11 to 2026-09-28] Grok 4.8 (2.5T parameters) has no release date. As of 2026-09-28 it has no xAI page, model ID or price. A 2026-10-11 X post guessed a release "this week", but I found no release by 2026-10-18. Musk's ten-day-class promises for 4.7 slipped at least five times (Jul 24 "four weeks" to Sep 21). A repeat slip would push 4.8 toward November or December and leave little time for any board to list it. https://cellcog.ai/blog/grok-4-8-release-date/ - [2026-09-14] Musk's own roadmap rates 4.7 as roughly Opus 5.0 parity, 4.8 as a "noticeable improvement" and only 4.9 as "probably Astra/Fable class". So 4.8 is the only plausible release before year-end, and "noticeable improvement" is not a claim of 45%+ on HLE. Musk said Grok 5 would arrive "within 2026" on 2026-08-05, but no date exists, and it has slipped from late 2025, then Q1 and Q2 2026. https://www.winzheng.com/en/article/xai-grok-4-8-2-5t-cpp-stack-agi-roadmap-grok-47-delay - [2026-10-08] Grok 5 has no benchmark card and no model ID. xAI's only statement is the 28 Jan 2026 line that it is "in training". Another aggregator update on 2026-10-08 still lists Q2 2026 as the expected window, which has already been missed. https://overchat.ai/ai-hub/grok-5-release-date - [2026-10-16] Competitors are on the board at higher scores (Step 5 Preview 46.5%, Kimi K3 46.9%). Grok's 42.9% sits below several open-weight models. This suggests xAI optimises for coding and agents, not HLE, and that the "mid-pack" 45% level is not trivially cleared. https://marktechpost.com/2026/10/16/stepfun-releases-step-5-preview-open-weights-600b-moe-27b-active-1m-context - [2026-10-18] Resolution wording: "HLE Accuracy" on the agi.safe.ai board. That is the o3-mini-judged table with its calibration column. Artificial Analysis, Scale, BenchLM and Wikipedia are not that source. The fallback to "official results published elsewhere" applies only if the source is unavailable or permanently gone. The source is available, so a strict reader resolves on the static table (24.5%). If no Grok ≥45% appears there, the market resolves No. https://agi.safe.ai/ BOTTOM LINE: The strongest No case is that the named resolution source is frozen at Grok 4 (24.5%). The only maintained CAIS and Scale board puts Grok 4.7 near 23%. Third-party Grok figures (best 42.9%) are below 45% and are not the named source. Any rescue depends on an unscheduled Grok 4.8 or 4.9 release that slips and on the board being updated, neither of which has precedent. NOT FOUND: The contents of the dynamic HLE results widget on agi.safe.ai. Any xAI-reported HLE score for Grok 4.5, 4.6 or 4.7. A confirmed Grok 4.8 release or date. Any CAIS statement on how it handles new submissions.