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

date	predicted	market price	confidence
2026-09-02	20%	44%	65%
2026-08-24	40%	49%	44%

1. Planner Fable 5.1 · Fermi decomposition + tool plan

Tools chosen: polymarket_direct resolution_source polymarket_related kalshi_related claude_news article_search gdelt_news wikipedia code_execution

Sub-questions (Fermi decomposition)

What is the current highest HLE Accuracy on the official agi.safe.ai leaderboard, and does it include tool-augmented/agentic runs or text-only scores?
What is the highest HLE score any Meta model (Llama 4 Maverick/Behemoth or newer) has achieved, and are Meta models currently listed on the leaderboard at all?
What frontier models is Meta (Meta Superintelligence Labs) expected to release in 2026, and what is the expected timeline and capability level relative to OpenAI/Google/Anthropic/xAI?
How fast has the HLE state-of-the-art been rising (e.g., ~3% Jan 2025 → ~25-45% by late 2025), and does extrapolation put frontier labs above 55% during 2026?
What prices do sibling Polymarket markets (other Meta HLE thresholds, and HLE thresholds for OpenAI/Google/xAI/Anthropic) imply about the probability distribution?
How does the leaderboard treat closed vs. open models and self-reported vs. CAIS-run evaluations — could a Meta claim fail to appear on the resolution source?

Planner reasoning

This is a Polymarket question about whether any Meta model reaches ≥55% HLE Accuracy on the official agi.safe.ai leaderboard by end of 2026. Key uncertainties are: the current frontier ceiling on HLE (and whether tool-augmented runs are listed), Meta's current standing and its frontier model roadmap (Meta Superintelligence Labs, next-gen model timing), and whether the leaderboard even lists Meta models. The market price is the primary anchor, with sibling threshold markets giving an implied distribution.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.4s	1	## This Market's Polymarket Data Will the highest score achieved by a Meta model on Humanity’s Last Exam in 2026 be 55% or higher? - Current price (probability): 33.00% - 7-day price change: -10.50% - 30-day price change: -15.50% - Total volume: $37,806 (USD notional) - Price range: 28.00% - 67
resolution_source	OK	3.8s	1	already_resolved=unclear; matches_criteria=True; last_updated=01/28/2026 (most recent news item listed). The fetched page does not display any leaderboard scores or model accuracy figures for Meta or any other organization on Humanity's Last Exam; it only shows news updates, dataset information, and
polymarket_related	OK	0.5s	24	Polymarket search over 4 keyword(s), kept 24 matches. keyword "Humanity's Last Exam": 8 markets \| keyword 'Meta HLE': 8 markets \| keyword 'HLE 2026': 0 markets \| keyword 'Meta model': 8 markets
kalshi_related	OK	4.6s	2	2 related markets / summaries. keyword "Humanity's Last Exam": ok \| keyword 'Meta AI model': ok \| keyword 'HLE': no matches
claude_news	OK	45.3s	21	Here are the key findings from research on this question: HLE Leaderboard Status (as of Sep 2026) - Top overall HLE scores (no-tools, text-only) as of mid-September 2026: the top-scoring model on Humanity's Last Exam is Claude Fable 5 at 55.5%, followed by Claude Opus 5 at 54.9% and GPT-6 Astr
article_search	OK	0.8s	31	Fetched 31 unique articles across 4 queries (lookback=120d). 'Meta Superintelligence Labs new frontier model 2026': 10 articles \| "Humanity's Last Exam leaderboard score record": 10 articles \| "Llama Humanity's Last Exam accuracy": 10 articles \| 'Meta AI model benchmark Alexandr Wang': 10 articles
gdelt_news	OK	97.9s	10	GDELT: 10 articles across 2 queries (lookback=60d). 'Meta Superintelligence Labs model': error GDELT rate-limited after retries (429) \| "Humanity's Last Exam": 10 hits
wikipedia	OK	0.3s	3	Fetched 3 Wikipedia entries (0 missing pages).
code_execution	OK	70.9s	0	## Key Findings Frontier trend fit (linear‑in‑logit, 4 anchor points: Jan‑25 9%, Feb‑25 26%, Jul‑25 44%, ~Nov‑25 37.5% midpoint): - Regression: `logit(score) = ‑1.64 + 0.135·t` (t = months since Jan‑2025), residual σ ≈ 0.75 in logit space (only 2 d.o.f. → very wide uncertainty). - Naïve extrapo

3. Evidence Brief Sonnet 5 · 7508 chars

# Current state Meta Superintelligence Labs' newest model (Muse Spark 1.3, released Sept 2026) scores ~45-47% on HLE under independent no-tools evaluation (Artificial Analysis) — still below the 55% threshold. Meta's own self-reported "with tools" score for Muse Spark 1.1 is 62.1%, which would clear 55% if tool-augmented runs count, but the official agi.safe.ai leaderboard's exact treatment of tools vs. standard accuracy for this resolution is unconfirmed. Kalshi/Polymarket price has fallen from a high of 67.5% to 33%, reflecting growing skepticism that Meta's standard (no-tools) score reaches 55% by year-end 2026. # Timeline of key events - 2025-01: HLE benchmark publicly released by CAIS/Scale AI; SOTA models score single digits (Wikipedia — confirmed). - 2025-02 to 2025-07: Frontier SOTA HLE accuracy rises rapidly, ~9% (Jan) → ~26% (Feb) → ~44% (Jul) (code_execution trend data — confirmed). - ~2025-11: SOTA growth decelerates/plateaus to ~37.5% on some trackers, signaling benchmark saturation (code_execution — reported). - 2025-06: Meta Superintelligence Labs (MSL) founded; Alexandr Wang (ex-Scale AI) named Chief AI Officer (Wikipedia — confirmed). - 2025-11: Yann LeCun departs Meta to found separate AI lab (Wikipedia — confirmed). - 2026-04-08: MSL ships first proprietary model, Muse Spark (closed weights, multimodal reasoning) — Meta's first non-open flagship (claude_news/Wikipedia — confirmed). - 2026-07: Muse Spark 1.1 released (agentic/coding focus); Meta self-reports HLE "with tools" = 62.1%, "without tools" = 52.2% (claude_news — confirmed by Meta's own table, independently unverified). - 2026-08: Meta releases Muse Image, then open-weight Muse Glimmer (on-device model) and open-weight Muse Spark 1.2 variant; Zuckerberg publishes essay on superintelligence strategy, favoring hybrid open/closed approach (CNBC/CBS — confirmed). - 2026-09: Muse Spark 1.2 and 1.3 released; independent (Artificial Analysis, no-tools) scores hover 44-47% HLE, still below rivals; Wikipedia AI Index lists Muse Spark 1.1 at 46.2% (below GPT-6 Astra) (claude_news — confirmed/reported). - 2026-09: Overall HLE SOTA (no-tools) led by Claude Fable 5.1/Opus 5/GPT-6 Astra in 55-65% range depending on tracker methodology (pricepertoken, Artificial Analysis, llm-stats — reported, sources disagree). - 2026-09-09: CNBC references an unreleased "Watermelon" Meta model "on the near-term horizon" — no benchmark data yet (reported). # Event Will Meta's best HLE model score reach ≥55% accuracy (per agi.safe.ai leaderboard) at any point in 2026? # Outcomes to forecast Yes / No # Kalshi market anchor Polymarket-sourced price for this exact ticker: **33% YES** (No 67%). Down 10.5% over 7 days, down 15.5% over 30 days; range has been 28%-67.5% over 58 days of trading — a sharp decline from earlier optimism. Volume: ~$37.8K. # Resolution source agi.safe.ai fetched successfully (200, last news item dated 01/28/2026) but the captured page content did not include actual leaderboard numeric scores — only news/dataset info. Cannot confirm current Meta standing directly from source; treat as inconclusive/unresolved as of fetch. # Sub-question answers 1. **Current top HLE score & tools status** — Trackers disagree: llm-stats shows overall leader Claude Fable 5.1 at 65% (methodology unclear); Artificial Analysis (no-tools, 2,158-question subset) shows ceiling ~59.1%; pricepertoken shows top score 55.5%. Separately, "no-tools" vs "with-tools" leaderboards are tracked distinctly (Tencent Hy3: 47.0% no-tools vs 53.2% with tools) (claude_news). 2. **Meta's highest HLE score** — Llama 4 Maverick: 5.7% (legacy, irrelevant now). Muse Spark 1.3 (Sept 2026, current best, no-tools): ~45-47% independently verified. Muse Spark 1.1 self-reported "with tools": 62.1% (claude_news, Meta's own table — unverified independently). 3. **Meta's 2026 roadmap** — MSL shipped Muse Spark (Apr), 1.1 (Jul), 1.2/1.3 (Sep) — roughly monthly-to-quarterly cadence; a further model ("Watermelon") is rumored/near-term (CNBC, Sept 2026). Compute investment reportedly comparable to OpenAI/Anthropic through 2027 (claude_news). 4. **HLE SOTA trend** — Rapid rise Jan-Jul 2025 (9%→44%), deceleration by Nov 2025 (~37.5%), current 2026 frontier 55-65%. Naive extrapolation implies ~76% by Dec-2026 (likely overstated due to plateau); damped estimate: frontier 45-60% by year-end (code_execution). 5. **Sibling Polymarket prices** — Meta 55%=33%, 60%=25.5%, 65%=17%, 70%=6.1% (declining monotonically, consistent structure). OpenAI 55%=80%, 60%=44%, 65%=15.5%, 70%=9.5%. Google Gemini 50%=84%, 55%=36%, 60%=19.5%, 65%=11.5%. Market clearly ranks OpenAI > Google ≈ Meta at the 55%+ threshold, with Meta priced lowest among the three at comparable thresholds. 6. **Leaderboard treatment of tools/self-reported vs CAIS-run** — Third-party trackers (llm-stats, Artificial Analysis) separate "no tools" and "with tools" leaderboards; Meta's self-reported 62.1% (with tools) may not be independently replicated/listed on agi.safe.ai's official leaderboard, creating resolution ambiguity (claude_news). # Key facts 1. [polymarket_direct] Current YES = 33%, down from 67.5% high. 2. [claude_news] Muse Spark 1.3 (Sept 2026) independently verified at ~45-47% no-tools HLE. 3. [claude_news] Meta self-reports 62.1% with-tools HLE for Muse Spark 1.1 — unverified externally. 4. [Wikipedia] Muse Spark released Apr 2026, closed weights, MSL's first proprietary flagship. 5. [resolution_source] agi.safe.ai reachable but leaderboard numbers not captured in fetch; unresolved status. # Cross-market signals - Polymarket sibling Meta thresholds (60/65/70%) all priced lower than 55%, consistent monotonic structure. - OpenAI 55%+ priced far higher (80%) than Meta (33%), reflecting perceived capability gap. - No Kalshi-native related markets on HLE found (only unrelated Meta stock-price/antitrust markets). # Analyst opinions and speculation - Independent evaluators (Artificial Analysis) view Meta's self-reported tool-augmented gains skeptically; Claude Fable 5.1 seen as ahead of Muse Spark 1.3 on both Intelligence Index and HLE (claude_news). - Analysts (SemiAnalysis) note Meta's compute buildout rivals top labs, suggesting future catch-up potential. - Monte Carlo model (code_execution) estimates P(Meta ≥55% by Dec 2026) ≈ 10-20% baseline, range 4-30%, highly sensitive to assumed "catch-up lag" (3-24 months) and whether Meta's score even appears on the official leaderboard. # Directional lean per outcome - **Yes**: Muse Spark's rapid iteration (Apr→Sep 2026, four releases) shows fast cadence; heavy compute investment; self-reported tool-use score (62.1%) already exceeds 55% if such scores count. - **No**: No-tools/independently-verified scores remain 8-10 points below threshold (~45-47%); market price fell sharply (67.5%→33%) suggesting resolved skepticism; ambiguity favors stricter no-tools interpretation, which Meta hasn't cleared; overall SOTA plateauing. # Gaps / unknowns - Whether agi.safe.ai's official leaderboard will list Meta models at all, and whether it counts tool-augmented scores. - No direct current agi.safe.ai leaderboard snapshot with figures obtained. - Uncertain cadence/capability of rumored next model ("Watermelon"). # Calibration anchors - Kalshi/Polymarket YES price (anchor): 33%, trending down. - Sibling Meta thresholds imply market's own probability curve: 55%→33%, 60%→25.5%, 65%→17%, 70%→6.1%. - Monte Carlo model estimate: ~10-20% (baseline), broader range 4-30%.

4. Forecasters 5 draws · Fable 5.1 + GPT-5.6 Sol

claude-fable-5-1

46% conf

Yes	29%
No	71%

The Polymarket anchor is 33% YES and has been falling steadily (67.5% → 33%), reflecting that Meta's best independently verified no-tools HLE score (Muse Spark 1.3, ~45-47%) sits 8-10 points below the 55% threshold with roughly one quarter left in 2026. Meta's self-reported 62.1% 'with tools' figure would clear the bar, but the resolution source (agi.safe.ai) treatment of tool-augmented or self-reported scores is unconfirmed and third-party trackers separate those leaderboards, so this path is a genuine but uncertain upside tail. A YES still requires either a rumored 'Watermelon'/next Muse release delivering a ~10-point jump by December or a favorable tools interpretation; Meta's fast release cadence and compute buildout keep this plausible, but the Monte Carlo estimate (~10-20%) and frontier plateauing argue for staying slightly below the market. I settle modestly under the anchor at 29% YES.

claude-fable-5-1

45% conf

Yes	28%
No	72%

Meta's best independently verified no-tools HLE score (Muse Spark 1.3, Sept 2026) sits at ~45-47%, roughly 8-10 points below the 55% threshold with only ~3 months left in 2026, and overall SOTA growth is plateauing. The main Yes path is either a strong 'Watermelon' release or the resolver counting Meta's self-reported 62.1% with-tools figure, but third-party trackers separate tools/no-tools and the agi.safe.ai leaderboard's treatment is unconfirmed. The Polymarket anchor is 33% YES and has fallen sharply from 67.5%, while the sibling curve (60%→25.5%, 65%→17%) and a Monte Carlo baseline of 10-20% both support a slight shade below the anchor. I set Yes at 0.28, close to the market with modest downward adjustment for the stricter no-tools interpretation being the likeliest resolution basis.

claude-fable-5-1

44% conf

Yes	28%
No	72%

Meta's best independently verified no-tools HLE score (Muse Spark 1.3, Sept 2026) is ~45-47%, roughly 8-10 points short of the 55% threshold, with only about three months left in 2026 and the overall frontier plateauing in the 55-65% range. Meta's self-reported 62.1% 'with tools' score would clear the bar, but third-party trackers and likely agi.safe.ai treat tool-augmented runs separately and Meta's self-reports have not been replicated, so resolution ambiguity cuts against YES. Fast release cadence (four models Apr–Sep) and a rumored 'Watermelon' model keep a real path to ~55% open, and Meta's own no-tools claim of 52.2% suggests the gap could be closed by one strong release. Anchoring on the Polymarket price of 33% (down from 67.5%) and the Monte Carlo estimate of 10-20%, I shade modestly below market to 28%.

gpt-5.6-sol

58% conf

Yes	30%
No	70%

The 33% YES market price is the primary anchor, but Meta’s independently evaluated no-tools score of roughly 45–47% remains 8–10 points below the threshold. Meta’s rapid release cadence and rumored Watermelon model leave a plausible path to 55%, while its self-reported 62.1% with-tools result could already qualify if agi.safe.ai accepts that methodology. However, evaluators distinguish tool-assisted and no-tools results, the official leaderboard treatment is unconfirmed, and broader HLE progress has decelerated. These factors support moving slightly below the market consensus to 30% YES.

gpt-5.6-sol

64% conf

Yes	29%
No	71%

The 33% YES market price is the primary anchor, but Meta’s independently measured no-tools score of roughly 45–47% remains 8–10 points short with only a few months left in 2026. Rapid Muse releases and the rumored Watermelon model provide a credible path to another improvement, while the broader frontier has already reached roughly 55–65%. Meta’s self-reported 62.1% with-tools result could make YES effectively achieved if accepted, but third-party leaderboards distinguish tool use and the agi.safe.ai treatment is unconfirmed. Given the recent plateau, independent-evaluation gap, and stricter-resolution risk, I move modestly below the market to 29% YES.

6. Final plain mean of 5 draws

29% (51% confidence)

Yes: 29% No: 71%

Mean of 5 draws (0.29, 0.28, 0.28, 0.30, 0.29). The Polymarket anchor is 33% YES and has been falling steadily (67.5% → 33%), reflecting that Meta's best independently verified no-tools HLE score (Muse Spark 1.3, ~45-47%) sits 8-10 points below the 55% threshold with roughly one quarter left in 2026. Meta's self-reported 62.1% 'with tools' figure would clear the bar, but the resolution source (agi.safe.ai) treatment of tool-augmented or self-reported scores is unconfirmed and third-party trackers separate those leaderboards, so this path is a genuine but uncertain upside tail. A YES still requires either a rumored 'Watermelon'/next Muse release delivering a ~10-point jump by December or a favorable tools interpretation; Meta's fast release cadence and compute buildout keep this plausible, but the Monte Carlo estimate (~10-20%) and frontier plateauing argue for staying slightly below the market. I settle modestly under the anchor at 29% YES.

Pipeline Timing

Total pipeline time: 181.9s

Per-tool research timings shown in the Research section above.