Will the highest score achieved by a Meta model on Humanity’s Last Exam in 2026 be 55% or higher?

- [2026-09-22] The official HLE site (agi.safe.ai) confirms its current "Latest News" update is the release of HLE-Diamond, a new/revised variant of the benchmark, as its most recent listed change; the page links to a "Submission Dashboard" for official results but did not return a browsable score table in a direct fetch. (https://agi.safe.ai/) - [2026-10-01] Wikipedia's Humanity's Last Exam results table (updated within the last ~7 days) lists the Meta Superintelligence Labs entry as "Meta Superintelligence Labs | Muse Spark 1.3 | 48.7" , placing Meta below OpenAI's GPT-6 Astra at 54.7 and several other labs, and still short of 55%. (https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam) - [2026-08-16] A secondary tracker reports that on the "official public-question leaderboard" Gemini 3.1 Pro Preview (thinking high) at 46.44%, GPT-5.4 Pro at 44.32%, and Muse Spark at 40.56% , and explicitly notes no model listed on the official public leaderboard has reached 50% as of that date. (https://intuitionlabs.ai/articles/humanitys-last-exam-ai-benchmark) - [2026-09-23] A benchmark mirror of the Scale Labs public HLE leaderboard snapshot shows Meta's best public-leaderboard row as "Muse Spark" at 44.3% , well under the question's 55% bar. (https://benchlm.ai/benchmarks/scale-humanitys-last-exam) - [2026-04-08] Meta's own launch materials for Muse Spark (original version) self-reported a score of 42.8 (No Tools) and 50.4 (With Tools) on HLE, while independent audits by Artificial Analysis tracked the model at 39.9% — i.e., Meta's own no-tool number and the independent audit are both well below 55%. (https://venturebeat.com/technology/goodbye-llama-meta-launches-new-proprietary-ai-model-muse-spark-first-since) - [undated, ~90 days old] Meta's self-published table for a later release, Muse Spark 1.1, claims "Humanity's Last Exam, with tools | 62.1 ... Humanity's Last Exam, no tools | 52.2" , both figures from Meta's own evaluation report rather than confirmed entries on the official agi.safe.ai leaderboard. (https://kingy.ai/blog/muse-spark-1-1-benchmarks-specs-evals/) - [2026-09-29] Artificial Analysis' HLE ranking (a widely used but third-party, not the official CAIS/Scale source) shows the overall frontier topped by Claude Opus 5.5 at 61.4%, with Meta's Muse Spark line not among the very top scores shown, underscoring that even on generous third-party trackers Meta is not the HLE leader. (https://benchgecko.ai/benchmark/aa-humanitys-last-exam) - [2026-10-02] Another third-party aggregator (BenchLM) places "Muse Spark 1.1" at 62.1% and the earlier "Muse Spark" at 50.4% on its own compiled HLE table, but this page mixes self-reported "with tools" marketing numbers from multiple vendors and is not the resolution source. (https://benchlm.ai/benchmarks/hle) - [2026-10-07, today] No evidence was found of any Meta model officially listed on agi.safe.ai (or its CAIS/Scale AI leaderboard twin) at or above 55% accuracy; the highest credible, source-linked Meta figure on leaderboard-style trackers is 48.7% (Muse Spark 1.3), with most trackers in the 40–50% range depending on tool-use conditions. (https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam) - [2026-09-22] HLE-Diamond was just released as a new/updated variant of the exam; if the official leaderboard transitions scoring to HLE-Diamond before Dec 31, 2026, this could reset or rescale reported accuracies, adding resolution uncertainty about which "HLE Accuracy" figure counts. (https://agi.safe.ai/) BOTTOM LINE: On every source that plausibly reflects the official HLE leaderboard (Scale AI/CAIS), Meta's best model (Muse Spark/Muse Spark 1.3) currently sits roughly in the 40–49% range, and even Meta's own generous self-reported "with tools" marketing figures (~50–62%) are not confirmed as official leaderboard entries; no current data shows a Meta model at or above 55% on the resolution source. NOT FOUND: A directly browsable, dated table from agi.safe.ai/dashboard itself (the fetch to that URL failed/was inaccessible), and any official CAIS/Scale AI leaderboard entry for Meta explicitly exceeding 50%, let alone 55%, as of October 2026.

- [2026-09-23] Reference-class base rate: sibling Polymarket markets use the identical "highest score ≥ X%" format per lab. For Anthropic Claude, a 55%+ threshold already prices near resolution (~89¢ "Yes" on a related 2026 full-year Claude market), reflecting that Claude passed 55% months ago; a 65%+ Claude market still trades live, showing thresholds above the current frontier are genuinely uncertain while thresholds below it resolve YES almost immediately once the leaderboard updates. https://polymarket.com/@longdated-poli - [2026-10-02] A comparable "Gemini Argon" HLE debut market (next major Gemini release) prices 55%+ at only ~50% and 60%+ at ~13%, while 45%+ and 50%+ are near-certain (95%/91%) — showing the market systematically treats each additional ~5-point threshold as a large, non-trivial hurdle even for a top-tier lab's next release. https://polymarket.com/event/gemini-argon-humanitys-last-exam-debut-20261002 - [2026-09-29] A similar "Next Fable Model (5.2+)" debut market prices 50%+ at 93% but 55%+ at only 87-88% and 60%+ at 28%, reinforcing that each 5-10 point jump above a lab's current best is priced well below certainty even late in the year. https://polymarket.com/event/next-fable-model-5pt2-humanitys-last-exam-debut-20260928 - [2026-10-07 cache] The Wikipedia HLE results table (sourced to the AI Index report lineage) lists Meta Superintelligence Labs' best entry as Muse Spark 1.3 at an accuracy of 48.7%, below the 55% strike, versus OpenAI's GPT-6 Astra at 54.7% in the same table — i.e., Meta trails the leading lab by several points on the standard (no-tools) leaderboard metric. https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam - [2026-10-01] Third-party leaderboard mirrors (Artificial Analysis-sourced) place Meta's best model (Muse Spark family) well behind the top of the board, which by October is led by Anthropic/OpenAI models scoring 54.7%-65% (with-tools variants reaching into the 60s); Meta is not shown among the top three anywhere in these snapshots. https://artificialanalysis.ai/evaluations/humanitys-last-exam - [2025 vs 2026 trend] HLE organizers originally projected models "could exceed 50% accuracy on HLE by the end of 2025" but as of August 2026 no model on the official public leaderboard had reached 50% under standard (no-tools) protocol, indicating the benchmark has historically been harder to saturate than forecast, which is a relevant outside-view caution against assuming rapid further gains. https://intuitionlabs.ai/articles/humanitys-last-exam-ai-benchmark - [2026-04-08] Meta's own "Muse Spark" launch blog reported 58% only in a non-standard, multi-agent "Contemplating" mode (tools-assisted), while its standard/no-tools score was lower (~42-43%); this illustrates a recurring pattern where labs' marketing numbers (often tool-augmented) diverge substantially from the stricter "HLE Accuracy" metric likely used on the official leaderboard that this market cites. https://ai.meta.com/blog/introducing-muse-spark-msl/ - [2026-07-10] Meta's Muse Spark 1.1 update gained about 5 points on HLE (40%→45%) in roughly three months versus 1.0, the most concrete historical "rate of improvement" data point available for a Meta model generation-to-generation. https://x.com/ArtificialAnlys/status/2075677416295739660 - [2026-09-02] Muse Spark 1.3 released, reaching 48.7% per Wikipedia's table — consistent with a roughly similar incremental (3-4 point) gain pattern across successive Meta releases (1.0→1.1→1.2→1.3) rather than a sudden jump to 55%+. https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam BOTTOM LINE: The base-rate/outside-view picture is mixed-to-unfavorable for YES: other labs' 55%+ HLE thresholds have resolved YES this year only after visible leaderboard crossings, and sibling markets price 55%+ thresholds at roughly 50-90% depending on how close the lab already is to that level — but Meta's own trajectory (steady ~3-5 point gains per release, last confirmed official score 48.7%) implies it would need one more above-average generational jump within under three months to clear 55%, a plausible but not yet demonstrated outcome. NOT FOUND: No confirmed Meta model score ≥55% on the official agi.safe.ai "HLE Accuracy" standard leaderboard as of the search date; could not access the live agi.safe.ai dashboard directly (JS-rendered) to verify the absolute latest entries; no explicit historical count of how many similar per-lab HLE threshold markets have resolved YES vs. NO to compute a numeric base rate.

YES means: at some point before December 31, 2026, 11:59 PM ET, a Meta-published model's score, as reported on the official Humanity's Last Exam leaderboard (agi.safe.ai / its Scale AI dashboard), reaches at least 55% HLE Accuracy (or a clearly equivalent labeled metric even if the presentation changes, e.g. a rescaled or "Diamond" variant of the benchmark). - [2026-10-07] Meta's own developer documentation (dev.meta.ai), a primary Meta source, lists Humanity's Last Exam ... Multidisciplinary reasoning (w/ tools) 62.1 for Muse Spark 1.1, comfortably above the 55% bar — showing Meta has already produced and published, in an official capacity, a qualifying-looking number, even if it isn't yet mirrored on agi.safe.ai itself. (https://dev.meta.ai/models/muse-spark-1-1) - [2026-10-07] On the same official Meta table, the prior-generation Muse Spark scored Humanity's Last Exam...50.4 "with tools," meaning the generational jump from Muse Spark to 1.1 alone added nearly 12 points on this tool-assisted reading of HLE in a few months. (https://dev.meta.ai/models/muse-spark-1-1) - [2026-10-07] Scale Labs' live leaderboard page (fetched today) now hosts a rebuilt "Humanity's Last Exam (Diamond)" board with materially higher absolute scores than the classic HLE scale — top entries gpt-6-astra60.60±3.002claude-opus-5-555.00±3.103claude-fable-5-151.30±3.10 — indicating the benchmark's newest official variant is already yielding scores well above 55% for frontier labs, which raises the ceiling Meta would need to match if/when it submits to this specific board. (https://labs.scale.com/leaderboard) - [2026-10-07] On that same live Scale Labs leaderboard, Meta's newest model is not a one-off: it is "NEW" and already topping multiple other Scale categories today, e.g. Professional Reasoning Benchmark - FinanceEvaluating Professional Reasoning in Finance1Muse Spark 1.3NEW59.54±1.66 and Professional Reasoning Benchmark - LegalEvaluating Professional Reasoning in Legal Practice1Muse Spark 1.3NEW61.56±1.72 , both above 55%, showing Meta's current flagship can clear this bar on Scale's own evaluation infrastructure when the task favors it. (https://labs.scale.com/leaderboard) - [2026-10-07] Meta's own eval methodology document confirms Meta directly audits and republishes HLE itself under controlled, replicable protocol rather than relying only on third parties: Humanity's Last Exam (HLE): We use the full 2,500-question dataset and GPT. o3-mini as the LLM-as-judge model, following the official HLE evaluation protocol. This shows Meta has the infrastructure and willingness to generate official-grade HLE numbers for each new release, which could be submitted to agi.safe.ai/Scale at any time. (https://ai.meta.com/static-resource/muse-spark-eval-methodology) - [2026-10-07] A third-party deep-dive on the original Muse Spark documents the exact scaling mechanism Meta already demonstrated for crossing 55% on a "with tools" HLE reading: Their data on Humanity's Last Exam (With Tools) shows that going from 1 agent (~50%) to 2 agents (~56%) to 4 agents (~57%) to 16 agents (~58.5%) delivers consistent accuracy gains at comparable latency. This is a concrete, already-validated engineering lever (more parallel "Contemplating" agents) Meta can dial up again on its newest base model. (https://www.eigent.ai/blog/meta-muse-spark-personal-superintelligence) - [2026-09-02] Meta's release cadence has been rapid and continuous — Muse Spark 1.3 (the current stable release) shipped just five months after the original April 2026 debut, per Wikipedia's infobox: Stable release: Muse Spark 1.3 / September 2, 2026 (2026-09-02) — implying at least one more point release (1.4 or a Muse 2 "Behemoth"-class model) is plausible before the Dec 31 deadline, each prior step having added several HLE points. (https://en.wikipedia.org/wiki/Muse_Spark) - [2026-08-10] Meta has publicly signaled it is not done scaling: The technical blog describes it as 'the first step on our scaling ladder' — a deliberately sized initial model that validates the new stack before Meta goes larger. A larger model is already in development. A larger, more capable successor landing before year-end is the most direct mechanism for a qualifying 55%+ no-tools (or tool-assisted) HLE score. (https://neuralcoretech.com/meta-muse-spark-review-benchmarks-2026/) - [2026-08-10] Zuckerberg has also committed to opening up the Muse Spark line and continuing rapid public releases: Meta CEO Mark Zuckerberg said the company would open the weights for Muse Spark 1.2, its latest AI model. Zuckerberg also said the company would launch a new family of open-source models called Muse Glimmer that are designed to run on laptops. Continued frequent public releases increase the chances that one of them clears 55% before the leaderboard snapshot used for resolution. (https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html) BOTTOM LINE: The most credible YES path is a combination of (a) Meta's demonstrated, already-published "with tools"/multi-agent HLE figures (50.4%, 62.1%) that exceed 55% under a reading the market's "clear equivalent metric" clause could accept, and (b) Meta's fast, ongoing release cadence (1.0→1.1→1.2→1.3 in five months, larger model "already in development") plus the newly observed higher-scoring HLE-Diamond variant on Scale's live leaderboard, any of which could push an official Meta leaderboard entry past 55% before December 31, 2026. NOT FOUND: No confirmed Meta entry on the agi.safe.ai/Scale "Humanity's Last Exam (Diamond)" leaderboard itself (the live fetch showed only non-Meta labs in the visible top-3 snippet); no dated announcement of a Muse Spark 1.4 or "Muse 2/Behemoth" HLE score; no explicit statement that Meta's tool-assisted 62.1% figure has been or will be submitted as the official agi.safe.ai "HLE Accuracy" entry.

YES means: some model published by Meta (e.g., a Muse Spark variant or any other Meta-released model) is confirmed, on the official Humanity's Last Exam leaderboard (agi.safe.ai) or an accepted official-equivalent source, to score 55% or higher "HLE Accuracy" at any point before Dec 31, 2026, 11:59 PM ET. - [2026-09-22] CAIS/Scale AI released **HLE-Diamond**, a new frozen 1,000-question variant (500 reasoning + 500 knowledge) explicitly designed to replace the saturating original 2,500-question HLE set, with the agi.safe.ai "Latest News" banner now pointing to it as the newest benchmark version. (https://agi.safe.ai/) - [2026-10-05] On Scale Labs' own official HLE-Diamond results table — the primary source the question's resolution criteria would plausibly track if "HLE Accuracy" is understood as "a clear equivalent metric" — **Meta's Muse Spark 1.3 scores only 25.4% overall** (31.6% reasoning, 19.2% knowledge), nearly 30 points below the 55% strike and dead last among non-open-weight entries except Grok and Kimi. (https://labs.scale.com/leaderboard/hle-diamond) - [2026-10-05] The same official table's leader, GPT-6 Astra, sits at only 60.6% overall and runner-up Claude Opus 5.5 at 55.0%, meaning on the newest official no-tools metric even the #2 lab in the world barely scrapes 55% — Meta is 30 points behind that frontier, not a few points short. (https://labs.scale.com/leaderboard/hle-diamond) - [2026-10-05] A third-party mirror of the same release (BenchLM) corroborates this exact figure, listing "Muse Spark 1.3 … No tools · high reasoning 24.6%" in its 12-model HLE-Diamond table, confirming it isn't a single-source data error. (https://benchlm.ai/benchmarks/hle-diamond) - [2026-10-07] The CAIS HLE-Diamond release notes explicitly state the headline "Headroom" finding that the strongest model tested lands at 60.6%, "well below the noise ceiling," meaning CAIS itself frames the whole benchmark — not just Meta — as far from saturated, i.e., no expectation of imminent 55%+ scores across the board, let alone for a trailing lab. (https://labs.scale.com/leaderboard/hle-diamond) - [2026-09-23] A separate news aggregator confirms the HLE-Diamond announcement carried **no benchmark scores or model results at launch**, meaning the resolution source itself only began publishing comparative Meta data in the following two weeks — a short, late-arriving data window that leaves little room for a 30-point Meta jump before year-end. (https://aibriefs.news/card/4ea77bae-c00b-42c3-8e30-8b9d0691ea20) - [2026-10-07] Even on the "with tools / highest effort" HLE-Diamond variant table (the most favorable condition tracked), the published leader GPT-6 Astra reaches only 66.2% and runner-up 61.2%, with Meta still absent from any top-tier row — there is no tool-assisted HLE-Diamond condition under which Meta's current model is shown clearing 55%. (https://benchlm.ai/benchmarks/hle-diamond-max) - [2026-10-05] Meta's own more favorable self-reported figures (e.g., "with tools 62.1%" for Muse Spark 1.1 cited by colleagues) apply only to the **original, pre-Diamond** 2,500-question HLE set and are not current entries on either agi.safe.ai's or Scale's HLE-Diamond table, creating a direct resolution-criteria mismatch risk if the official leaderboard now reports Diamond-based "HLE Accuracy" as its primary figure. - [2026-10-05] The gap between Meta's original-HLE figure (48.7%, per Wikipedia as of Oct 1) and its HLE-Diamond figure (25.4%) shows that question refreshes/rescales can move a lab's reported accuracy in the wrong direction by roughly half, illustrating how a benchmark-version technicality could mechanically keep Meta well under 55% for the rest of 2026 regardless of underlying model progress. - [2026-10-07] No source found shows Meta has released, or has announced an imminent release of, a next-generation flagship (e.g., "Muse Spark 2" or similar) between now and the Dec 31 deadline; absent a new model launch, status-quo inertia favors Meta's best official score staying anchored to Muse Spark 1.3's confirmed sub-50% (original HLE) / 25.4% (HLE-Diamond) results through year-end. BOTTOM LINE: On the newest and most likely "official" resolution surface (CAIS/Scale's HLE-Diamond table), Meta's best model scores just 25.4% — far further from 55% than the already-insufficient ~48.7% figure on the legacy HLE set — and even the global #1 model (60.6%) barely clears the bar, making a Meta jump to 55% within the remaining ~12 weeks of 2026 look structurally unlikely absent an unreleased, unannounced flagship model. NOT FOUND: Any confirmed Meta score ≥55% on either the original HLE or HLE-Diamond tables from a primary official source; any announcement of an imminent new Meta frontier model release before Dec 31, 2026; explicit agi.safe.ai guidance on whether "HLE Accuracy" for resolution purposes will be read as the legacy HLE score or the new HLE-Diamond score going forward.