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

date	predicted	market price	confidence
2026-08-28	70%	74%	48%
2026-08-21	32%	78%	48%

1. Planner Fable 5.1 · Fermi decomposition + tool plan

Tools chosen: polymarket_direct resolution_source polymarket_related kalshi_related article_search claude_news wikipedia code_execution

Sub-questions (Fermi decomposition)

What is the current highest 'HLE Accuracy' shown for any Grok model on the official leaderboard at agi.safe.ai, and does the leaderboard include tool-augmented / Heavy / agentic configurations or only no-tools text runs?
What are the current top HLE scores across all labs (OpenAI, Google, Anthropic) on the official leaderboard, and how quickly have top scores risen over the past 12-18 months (e.g., from ~9% in Jan 2025 to what now)?
What HLE numbers has xAI/SpaceXAI publicly claimed for Grok 4, Grok 4 Heavy, and any newer models (Grok 4.x, Grok 5), and do those self-reported numbers match what agi.safe.ai lists?
When is Grok 5 (or the next frontier Grok) expected to release in 2026, and has xAI stated HLE targets for it?
How frequently and how quickly does the CAIS/Scale HLE leaderboard add new frontier models after release, and does it evaluate via API access that xAI provides?
What probabilities do sibling Polymarket markets assign to other Grok HLE thresholds (e.g., 40%, 50%, 55%, 60%) and to other labs' models reaching similar thresholds in 2026?

Planner reasoning

This resolves strictly on the 'HLE Accuracy' value shown for Grok models on the agi.safe.ai leaderboard by end of 2026, so the key uncertainties are what that page currently shows for Grok (and whether tool-augmented/agentic runs are listed), how fast frontier HLE scores are rising, and whether xAI/SpaceXAI ships a Grok 5-class model in 2026 that gets evaluated and posted. The Polymarket price and sibling threshold markets (e.g., 40%, 50%, 55%) give a crowd-implied distribution to anchor against.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.5s	1	## This Market's Polymarket Data Will the highest score achieved by a SpaceXAI Grok model on Humanity’s Last Exam in 2026 be 45% or higher? - Current price (probability): 77.55% - 7-day price change: +9.05% - 30-day price change: +7.40% - Total volume: $44,089 (USD notional) - Price range: 46.5
resolution_source	OK	5.7s	1	already_resolved=unclear; matches_criteria=True; last_updated=01/28/2026 (most recent news item shown). The fetched page content does not display any leaderboard data, model names, or HLE Accuracy scores for any models, including SpaceXAI Grok models. It only shows general information about the Huma
polymarket_related	OK	0.8s	27	Polymarket search over 4 keyword(s), kept 27 matches. keyword "Humanity's Last Exam": 8 markets \| keyword 'Grok HLE': 8 markets \| keyword "highest score Humanity's Last Exam 2026": 8 markets \| keyword 'Grok 5': 3 markets
kalshi_related	OK	6.5s	2	2 related markets / summaries. keyword "Humanity's Last Exam": ok \| keyword 'Grok': no matches \| keyword 'HLE benchmark': ok
article_search	OK	46.5s	36	Fetched 36 unique articles across 4 queries (lookback=120d). "Grok Humanity's Last Exam score": 10 articles \| 'Grok 5 release HLE benchmark': 10 articles \| "Humanity's Last Exam leaderboard new record": 10 articles \| 'xAI SpaceX merger Grok model release 2026': 10 articles
claude_news	OK	31.9s	13	Key findings: - At original July 2025 launch, xAI claimed Grok 4 Heavy saturates most academic benchmarks and is the first model to score 50% on Humanity's Last Exam , specifically Grok 4 Heavy leads USAMO'25 with 61.9%, and is the first to score 50.7% on Humanity's Last Exam (text-only subse
wikipedia	OK	0.5s	3	Fetched 3 Wikipedia entries (0 missing pages).
code_execution	OK	56.0s	0	## HLE Trend Extrapolation Analysis — Key Findings Data anchors used: Jan 2025 ~9% (o1/o3-mini) → Apr 2025 ~21% (o3) → Jun 2025 ~25% (Grok 4, no tools) → Jul 2025 ~44% (Grok 4 Heavy, with tools). Target horizon: Dec 2026 (month 24). Naive linear extrapolation is unusable: Fitting a straigh

3. Evidence Brief Sonnet 5 · 6980 chars

# Current state No SpaceXAI Grok model has yet been confirmed at ≥45% on the official CAIS/Scale HLE leaderboard as of the latest available data (Sept 2026): best confirmed entry is Grok 4.6 at 42.9% (Wikipedia HLE table). xAI's self-reported, tools-enabled 44.4–50.7% claims for Grok 4 Heavy (Jul 2025) were never fully reflected on the leaderboard and a sibling Manifold market resolved NO on that basis. Grok 5 — xAI's likeliest path to a leap — remains unreleased and repeatedly delayed. # Timeline of key events - 2025-01: HLE launched; top scores across labs ~9% (o1/o3-mini). [Wikipedia/code_execution] - 2025-04: o3 reaches ~21% no-tools. [code_execution] - 2025-07: Grok 4 launches; official leaderboard shows 25.4% (no tools); xAI self-reports Grok 4 Heavy at 44.4% (tools) / 50.7% (text-only subset claim) — confirmed self-report, NOT yet leaderboard-confirmed. [claude_news, dig.watch, Scientific American] - 2025-07 (shortly after): Manifold market resolves NO on Grok 4 hitting 45%+ on official leaderboard — confirmed. [claude_news] - 2026-02: xAI acquired by SpaceX, rebranded SpaceXAI (confirmed, Wikipedia). - 2026-04-30: Grok 4.3 released as current flagship (1M context, video/file input) — confirmed. [claude_news] - 2026-08: Grok 5 still unreleased; only confirmed to exist via Jan 2026 funding announcement; release date slipped from late-2025 → Q1 2026 → Q2 2026 — confirmed pattern of delay. [claude_news] - 2026-09 (mid-Sept): Best xAI leaderboard entry = Grok 4.6 at 42.9%; top overall scores: Claude Fable 5 55.5%, Claude Opus 5 54.9%, GPT-6 Astra 54.7% — confirmed via Wikipedia/pricepertoken/llmrun.dev. - 2026-01-28: Most recent news item on agi.safe.ai per resolution_source fetch (leaderboard values not captured in fetch). # Event Will any SpaceXAI Grok model reach ≥45% HLE Accuracy on the official agi.safe.ai leaderboard (or equivalent) by Dec 31, 2026? # Outcomes to forecast Yes / No # Kalshi market anchor No kalshi_direct data was returned for this ticker. The only direct market-price data available is Polymarket for this exact question: **YES = 77.55%**, up +9.05% (7d) and +7.40% (30d), volume $44,089, range 46.5–97.75% over 57 days. Treat this as the best available cross-market consensus proxy in place of Kalshi. # Resolution source agi.safe.ai (and mirror lastexam.ai) fetched successfully (200 status) but the crawl only returned intro/news/contributor content — no live leaderboard table was captured. Most recent news item dated 01/28/2026. Cannot confirm current listed Grok score directly from this fetch; relying on secondary sources (Wikipedia, llmrun.dev, pricepertoken.com) for leaderboard state. # Sub-question answers 1. **Current highest Grok HLE score / tool inclusion**: Not directly visible in the agi.safe.ai fetch. Secondary sources put the leaderboard-recognized Grok high at 42.9% (Grok 4.6, Sept 2026); the leaderboard has historically lagged/excluded xAI's tools-enabled "Heavy" self-reported numbers (44.4–50.7%). [claude_news, Wikipedia] 2. **Top scores across labs**: Sept 2026 leaders are Claude Fable 5 (55.5%), Claude Opus 5 (54.9%), GPT-6 Astra (54.7%) — up from ~9% in Jan 2025, a ~45-50 point industry rise in 20 months, but Grok has not kept pace (42.9%). [pricepertoken.com, Wikipedia] 3. **xAI's public HLE claims vs leaderboard**: Grok 4 Heavy claimed 44.4% (tools) / 50.7% (text-only) at July 2025 launch; official leaderboard showed only 25.4% then, and Grok 4.6 shows 42.9% as of Sept 2026 — self-reported figures have consistently outpaced/diverged from leaderboard-confirmed figures. [dig.watch, Scientific American] 4. **Grok 5 timeline**: Not released as of Aug/Sept 2026; slipped from late 2025 → Q1 2026 → Q2 2026, still unshipped; current flagship remains Grok 4.3 (Apr 2026). No stated HLE target for Grok 5 found. [claude_news] 5. **Leaderboard update cadence**: Not explicitly documented; evidence shows a lag pattern (Grok 4's July 2025 claims weren't immediately posted, "pending review/submission" per Scientific American). 6. **Sibling Polymarket thresholds**: Grok 45%=77.55% YES, 50%=44%, 55%=30.5% (all same series). Gemini 50%=84.9%, 55%=30.5%, 60%=19.5%. Any-lab overall 60%=51.5%, 65%=28.5%, 70%=14.5%. [polymarket_related] # Key facts (high-confidence, factual) 1. [Wikipedia/claude_news] Best confirmed xAI leaderboard score as of Sept 2026: Grok 4.6 at 42.9% — below 45%. 2. [claude_news] Sibling Manifold market already resolved NO on Grok 4 reaching 45%+ officially (44.4% self-reported fell short after rounding). 3. [claude_news] Grok 5 remains unreleased through at least Aug 2026, with repeated timeline slippage. 4. [Wikipedia] Competing labs (Anthropic, OpenAI) exceed 54-55% as of Sept 2026 — Grok trails significantly. 5. [polymarket_direct] Market price for this exact question is 77.55% YES and rising. # Cross-market signals - Kalshi related: no direct Grok/HLE Kalshi markets found; unrelated matches only. - Polymarket: this market 77.55% YES (rising); Grok 50% threshold 44%, 55% threshold 30.5% — implies market believes Grok likely clears 45% but is a coin-flip on 50%. - Sportsbook implied: N/A. # Analyst opinions and speculation - code_execution model argues "gap-to-target" reasoning (44% self-reported was only 1pt from 45%) implies near-certainty (~85-95%) — but this pre-dates and is contradicted by the more recent Sept 2026 confirmed 42.9% leaderboard figure, suggesting the earlier 44.4% never became the official recognized number. - claude_news bottom line: no Grok model has officially cleared 45% through mid-Sept 2026, with competitors pulling further ahead and Grok 5 delayed — bearish for YES relative to Polymarket price. # Directional lean per outcome - **Yes**: Polymarket pricing (77.55%, rising) and history of large generational jumps (25%→44% in one release) support upside if Grok 5 ships in Q4 2026 with tools/agentic scoring counted. - **No**: Structural/leaderboard evidence (42.9% as of Sept 2026, ~3.5 months before deadline) is more current and more reliable than the aging code_execution extrapolation; Grok 5 still unreleased; leaderboard has historically discounted xAI's tools-enabled self-reported scores. This is the stronger, more recent factual anchor. # Gaps / unknowns - Exact current leaderboard-listed Grok score not directly verified from agi.safe.ai (secondary sources only). - Whether Grok 5 releases before Dec 31, 2026 and whether it would be evaluated/posted before the deadline. - Whether leaderboard will count tools/Heavy-mode scores as "HLE Accuracy" per resolution criteria. # Calibration anchors - Polymarket YES price (proxy anchor): 77.55%, uptrending. - Precedent: Grok 4 self-reported 44.4% (tools) failed to register as 45%+ on official leaderboard (Manifold NO resolution) — a directly analogous prior miss. - Latest confirmed leaderboard figure (42.9%, Sept 2026) sits below both the 45% bar and the earlier self-reported peak, suggesting continued shortfall risk into year-end.

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

claude-fable-5-1

47% conf

Yes	70%
No	30%

The best available market anchor (Polymarket, no Kalshi data) prices YES at ~77.5% and has been rising over 7d/30d, implying traders expect a late-2026 Grok release to clear the bar. The brief's most recent confirmed leaderboard figure is Grok 4.6 at 42.9% (Sept 2026), only ~2 points short, and xAI has shown large generational jumps (25%→~44% in one cycle), so any Grok 4.x refresh or Grok 5 shipping in Q4 would very plausibly exceed 45%. Countervailing factors are material: Grok 5 has slipped repeatedly (late-2025→Q1→Q2 2026, still unshipped as of Aug/Sept), the official leaderboard has historically lagged or excluded xAI's tools-enabled self-reports (the analogous Grok 4 45% market resolved NO), and only ~3.5 months remain including evaluation/posting lag. I shade below the market to ~0.70 to reflect delay and resolution-source-lag risk, while respecting that the gap to 45% is small and the industry trajectory is steep.

claude-fable-5-1

42% conf

Yes	68%
No	32%

The best available market anchor is Polymarket at 77.55% YES and uptrending, implying the crowd expects Grok to clear 45% by year-end (with 50% priced as roughly a coin flip). Against that, the brief's most recent factual anchor puts the leaderboard-recognized Grok high at 42.9% (Grok 4.6), Grok 5 remains unreleased after repeated slips, and there is a direct precedent of xAI's tools-enabled self-reported 44.4% failing to register as ≥45% officially (Manifold NO). However, the remaining gap is only ~2 points, xAI has shown large generational jumps, and any Q4 release (Grok 5 or a 4.x refresh) would plausibly exceed the bar given competitors sit at 54-55%. I shade below the market to reflect leaderboard-lag and tools-counting resolution risk plus the possibility Grok 5 slips into 2027.

claude-fable-5-1

40% conf

Yes	72%
No	28%

The best available market anchor (Polymarket, no Kalshi data) prices YES at ~77.5% and is trending up, implying the market expects a Grok release to clear 45% before year-end. The brief's most recent confirmed leaderboard figure puts Grok 4.6 at 42.9% as of Sept 2026—only ~2 points short with ~3.5 months remaining, and xAI ships frequent point releases (4.3, 4.6) even while Grok 5 slips. Countervailing evidence: the leaderboard has historically discounted xAI's tools-enabled self-reports (Grok 4 Heavy's 44.4% claim never registered officially, and a sibling Manifold market resolved NO), Grok 5 remains unreleased with a pattern of delays, and any Q4 release must be evaluated and posted before Dec 31. I shade modestly below the market to reflect leaderboard-lag and rounding risk, landing at 0.72 YES.

gpt-5.6-sol

64% conf

Yes	64%
No	36%

The 77.55% Polymarket YES price is the best consensus anchor, and Grok 4.6 needs only a 2.1-point improvement from its reported 42.9% to clear the threshold. Rapid industry-wide HLE gains and the possibility of a late-2026 Grok update support YES, even without Grok 5. I discount the market because Grok 5 remains repeatedly delayed, xAI's earlier tools-enabled claims were not accepted by the official leaderboard, and only about three and a half months remain for both release and leaderboard posting. The inability to verify the live resolution-source table directly adds uncertainty to the reported current score and eventual resolution treatment.

gpt-5.6-sol

64% conf

Yes	64%
No	36%

The 77.55% Polymarket YES price is the best available consensus anchor, but the latest secondary-source leaderboard evidence warrants a bearish adjustment. Grok 4.6 is reportedly at 42.9%, leaving a modest 2.1-point gap, while competing models above 54% show that further gains are technically plausible. However, Grok 5 remains repeatedly delayed, and xAI’s earlier 44.4–50.7% self-reported results were not accepted as an official ≥45% leaderboard score, demonstrating material submission and evaluation risk. With roughly three and a half months remaining and the resolution-source leaderboard not directly captured, YES remains favored but substantially less than the market price suggests.

6. Final plain mean of 5 draws

68% (51% confidence)

Yes: 68% No: 32%

Mean of 5 draws (0.70, 0.68, 0.72, 0.64, 0.64). The best available market anchor (Polymarket, no Kalshi data) prices YES at ~77.5% and has been rising over 7d/30d, implying traders expect a late-2026 Grok release to clear the bar. The brief's most recent confirmed leaderboard figure is Grok 4.6 at 42.9% (Sept 2026), only ~2 points short, and xAI has shown large generational jumps (25%→~44% in one cycle), so any Grok 4.x refresh or Grok 5 shipping in Q4 would very plausibly exceed 45%. Countervailing factors are material: Grok 5 has slipped repeatedly (late-2025→Q1→Q2 2026, still unshipped as of Aug/Sept), the official leaderboard has historically lagged or excluded xAI's tools-enabled self-reports (the analogous Grok 4 45% market resolved NO), and only ~3.5 months remain including evaluation/posting lag. I shade below the market to ~0.70 to reflect delay and resolution-source-lag risk, while respecting that the gap to 45% is small and the industry trajectory is steep.

Pipeline Timing

Total pipeline time: 164.1s

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