Scan: Will the highest score achieved by an OpenAI model on Humanity’s Last Exam in 2026 be 60% or higher?

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 for any OpenAI model on the official agi.safe.ai leaderboard, and does the leaderboard show tool-use/search-augmented scores or only no-tools text scores?
What has the trajectory of top HLE scores been since the benchmark launched (Jan 2025) — how many percentage points per quarter has the frontier improved, and is it accelerating or plateauing?
What HLE scores has OpenAI claimed in its own model cards/announcements (e.g., GPT-5, o3, Deep Research, GPT-5.x, agent modes) and do those match what appears on the official leaderboard?
What OpenAI model releases are expected before end of 2026 (GPT-5.5/GPT-6, o-series successors), and what HLE numbers are rumored or projected?
How do competitor frontier models (Gemini 3, Grok 4/5, Claude) currently score on HLE, indicating whether 60% is within reach of the frontier by 2026?
What does the Polymarket implied distribution across HLE thresholds (40/50/60/70/80%) for OpenAI look like, and are there analogous markets for other labs?
Is there any risk of the agi.safe.ai leaderboard being discontinued, restructured, or not updated with new OpenAI models (which would push toward 'No')?

Planner reasoning

This is a Polymarket question about whether an OpenAI model will reach ≥60% HLE accuracy on the official agi.safe.ai leaderboard by end of 2026. Key drivers are the current top OpenAI score on the leaderboard, the historical rate of improvement on HLE, whether the leaderboard lists tool-augmented vs. no-tool scores (and which OpenAI has claimed), and upcoming OpenAI releases. The market price and sibling thresholds (50%, 70%, etc.) give an implied distribution to anchor against.

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 an OpenAI model on Humanity’s Last Exam in 2026 be 60% or higher? - Current price (probability): 42.00% - 7-day price change: +3.50% - 30-day price change: +9.00% - Total volume: $34,568 (USD notional) - Price range: 17.00% - 4
resolution_source	OK	4.9s	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 table, model names, or HLE Accuracy scores for OpenAI or any other models—only site navigation, news updates, and a long list of dataset c
polymarket_related	OK	0.6s	32	Polymarket search over 5 keyword(s), kept 32 matches. keyword "Humanity's Last Exam": 8 markets \| keyword 'HLE': 8 markets \| keyword 'OpenAI score': 0 markets \| keyword "Humanity's Last Exam Google": 8 markets \| keyword "Humanity's Last Exam xAI": 8 markets
kalshi_related	OK	3.9s	3	3 related markets / summaries. keyword "Humanity's Last Exam": ok \| keyword 'HLE benchmark': ok \| keyword 'OpenAI model': ok
article_search	OK	47.2s	34	Fetched 34 unique articles across 4 queries (lookback=120d). "Humanity's Last Exam score OpenAI leaderboard": 10 articles \| 'HLE benchmark new state of the art model': 10 articles \| "GPT-5 Humanity's Last Exam accuracy": 10 articles \| 'OpenAI next model release 2026 benchmark': 10 articles
claude_news	OK	32.1s	12	Here are the key research findings on HLE scores for OpenAI vs. competitor models: Historical trajectory (early 2025 → 2026): - When first released in early 2025, state-of-the-art models scored only a few percent – GPT-4o managed just 2.7% and Claude 3.5 Sonnet 4.1% – while subject-matter expe
wikipedia	OK	0.4s	3	Fetched 3 Wikipedia entries (0 missing pages).
code_execution	OK	64.5s	0	HLE Trend Analysis — OpenAI Frontier Score Extrapolation to Dec 2026 - No-tools track (pure model, OpenAI only): o1 (9.1%, Jan'25) → o3 (20.3%, Apr'25) → GPT‑5 (25.0%, Aug'25). Linear fit: slope ≈ 26.6 pts/yr → naive Dec‑2026 projection = 61.8%. But growth is clearly decelerating (Δ=11.

3. Evidence Brief Sonnet 5 · 7223 chars

# Event Will any OpenAI model reach ≥60% accuracy on the Humanity's Last Exam (HLE) leaderboard (agi.safe.ai) by Dec 31, 2026? # Outcomes to forecast - Yes (≥60% achieved by an OpenAI model in 2026) - No (highest OpenAI score stays <60%) # Kalshi market anchor No Kalshi-direct price was returned for this ticker in the raw research; the closest Kalshi-adjacent data is unrelated markets (OpenAI/Anthropic IPO race, Apple-OpenAI lawsuit) that provide no signal on this question. Treat Polymarket price as the best available cross-platform anchor: **42% YES** (up from 33% a month ago; range 17–47% over 58 days; $34.6k volume, rising trend). # Resolution source agi.safe.ai (and mirror lastexam.ai) was fetched but the returned page content only contained site navigation, news blurbs, and contributor lists — no live leaderboard table was captured. Most recent visible news item dated 01/28/2026. Cannot confirm current top OpenAI score directly from the primary source; third-party trackers (see below) must supplement. Status: unresolved/inconclusive from primary source as fetched. # Sub-question answers 1. **Current OpenAI leaderboard high & tools distinction** — Not directly visible on agi.safe.ai fetch. Third-party mirrors (Wikipedia, PricePerToken, Artificial Analysis, BenchLM) place OpenAI's best model (GPT-6 "Astra," as of Sept 2026) at ~54.7%, below the 60% bar. The official site does distinguish tool-augmented vs. no-tool scores (confirmed historically: GPT-5 no-tools 25%, GPT-5 w/tools 42%). 2. **Trajectory since Jan 2025** — o1 9.1% (Jan'25) → o3 20.3% (Apr'25) → GPT-5 25.0% no-tools (Aug'25) → GPT-5.5 Pro 43.1% no-tools (Apr'26) → GPT-6 Astra 54.7% (Sep'26) [claude_news, code_execution]. Growth is decelerating: large early jumps, then a multi-month plateau across labs in the 42–45% "tools" band mid-2025 to early-2026, followed by a fresh jump to ~54-59% range by late 2026. 3. **OpenAI's own claims vs. leaderboard** — GPT-5 pro claimed 42.0% (tools); GPT-5.5 announcement did not headline a standalone HLE score, instead folding it into the Artificial Analysis Intelligence Index — suggesting OpenAI is de-emphasizing HLE as a marketed metric [claude_news, openai.com]. 4. **Expected 2026 releases** — GPT-5.5, GPT-5.6 ("Sol"), and GPT-6 "Astra" have already shipped within 2026 per article_search/claude_news; no confirmed GPT-6.x/o-series successor beyond Astra is specified with a rumored HLE number exceeding 55-59%. 5. **Competitor scores** — As of Sept 2026: Claude Fable 5.1 leads at 59.1% (Artificial Analysis) or 65% (BenchLM); Claude Opus 5 ~54.9-64.7%; Gemini 3.7 Flash 47.9%. Anthropic models are consistently at or above the 60% frontier per multiple trackers, while OpenAI trails at 54.7% (Astra) — the competitive frontier is at/near 60%, but OpenAI specifically is not yet there. 6. **Polymarket distribution** — 55% threshold: 79.5% YES; 60%: 42% YES; 65%: 15.5% YES; 70%: 9.5% YES — a smooth, steeply declining CDF consistent with the market pricing 60% as a real but sub-even-odds outcome. Analogous Grok markets: 45% threshold 76.7% YES, 50% 56.5%, 55% 30.5%, 60% 16.5% — implying the market sees Grok as less likely than OpenAI to hit 60%. Gemini-Pro-debut markets show high confidence (86-92%) at 40-45% thresholds only. 7. **Leaderboard discontinuation risk** — Fetched agi.safe.ai page lacked visible leaderboard data (site restructuring or fetch limitation), but the market's rules allow using published results elsewhere if the source becomes permanently unavailable; no explicit evidence of discontinuation, just an incomplete fetch snapshot. # Key facts (high-confidence, factual) 1. [code_execution] OpenAI no-tools HLE trajectory: 9.1%→20.3%→25.0% (Jan-Aug 2025), decelerating. 2. [claude_news/VentureBeat] GPT-5.5 Pro (no tools) scored 43.1% (Apr 2026), below Anthropic's Mythos Preview (56.8%). 3. [claude_news/PricePerToken/Wikipedia] By Sept 2026, GPT-6 "Astra" ~54.7%, still <60%, and behind Claude Fable 5 (55.5%) and Opus 5 (54.9%). 4. [claude_news/Artificial Analysis, BenchLM] Trackers disagree sharply on absolute values (54.7% vs 59.1% vs 65% for top models), reflecting methodology differences (tool use, reasoning effort) — no clean "official" agi.safe.ai number retrieved. 5. [polymarket_related] Sibling Polymarket markets show a coherent declining CDF: 55%→79.5%, 60%→42%, 65%→15.5%, 70%→9.5% YES. # Cross-market signals - Kalshi related: no direct comparable market found; unrelated OpenAI markets (IPO race, Apple lawsuit) provide no signal. - Polymarket: 42% YES on this exact question, up 9pts in 30 days — moderate/rising confidence in OpenAI reaching 60%. - Polymarket (competitor threshold markets): implies market believes OpenAI more likely than Grok to hit 60%, but less certain than for lower Gemini thresholds — consistent with OpenAI being the #2-3 lab on HLE. # Analyst opinions and speculation - code_execution's saturation/logistic model (most credible per its own analysis) estimates only 20% probability of OpenAI reaching 60% by Dec 2026, citing an industry-wide 4-5 month plateau at 42-45% (tools) in 2025. - Naive linear extrapolations (discounted as unreliable) would suggest far higher probabilities (57-99%), illustrating high model-form sensitivity. - Synthesized code_execution estimate: ~25-35% probability. - OpenAI itself appears to be de-emphasizing standalone HLE marketing (folding into composite indices), possibly signaling either strategic reframing or acknowledgment it's not leading on this specific metric. # Directional lean per outcome - **Yes (≥60%)**: Supported by continued rapid model releases (GPT-5→5.5→5.6→6 Astra within ~18 months), rising Polymarket price (33%→42%), and precedent of Anthropic models already crossing/nearing 59-65% — showing 60% is within the industry's demonstrated capability ceiling. Opposed by OpenAI's own models consistently trailing Anthropic (54.7% vs 55.5-65%) and multi-month plateaus at each new capability tier. - **No (<60%)**: Supported by OpenAI's HLE scores strictly trailing Anthropic's best across all recent snapshots (Sept 2026: 54.7% vs. 54.9-65%), OpenAI's de-emphasis of the metric, and saturation-aware modeling giving only ~20-35% Yes probability. Opposed by the sheer pace of releases (3+ major OpenAI models in 2026 alone) leaving ~3 more months for another leap. # Gaps / unknowns - No confirmed live read of the actual agi.safe.ai leaderboard table (fetch only returned navigation/news). - No Kalshi-direct YES price was provided for this specific ticker. - Wide disagreement (54.7% to 65%) across third-party trackers for "top score" undermines confidence in exact current OpenAI standing. - No confirmed rumors of a specific OpenAI release explicitly targeting/claiming ≥60% HLE before Dec 2026. # Calibration anchors - Polymarket YES price: 42% (current, rising trend) — primary quantitative anchor. - Sibling Polymarket buckets: 55%→79.5%, 65%→15.5%, 70%→9.5% YES, forming a coherent declining CDF around the 42% at 60%. - Precedent: HLE frontier has moved from ~2-9% (Jan 2025) to ~55-65% (top labs, Sept 2026) in ~20 months — rapid but decelerating; OpenAI specifically has lagged the absolute frontier by 5-10 points throughout 2026.

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

claude-fable-5-1

35% conf

Yes	40%
No	60%

No Kalshi-direct price exists, so the Polymarket 42% YES (rising from 33%) is the primary anchor, and its sibling buckets (55%→79.5%, 65%→15.5%) form a coherent CDF around it. The brief places OpenAI's best model (GPT-6 'Astra') at ~54.7% as of Sept 2026, consistently trailing Anthropic's 55–65% frontier, while a saturation-aware model gives only ~20–35% for a further +5.3-point jump in the remaining months. Countervailing factors are OpenAI's rapid release cadence and the tools-vs-no-tools ambiguity on agi.safe.ai (tool-augmented scores run ~15+ points higher), which makes a ≥60% official entry more reachable than the no-tools trajectory suggests. Netting these, I shade slightly below the market at 40% YES; confidence is limited because the primary leaderboard was not captured and third-party trackers disagree by 10+ points on absolute scores.

claude-fable-5-1

33% conf

Yes	43%
No	57%

No Kalshi-direct price exists; the Polymarket anchor on this exact question is 42% YES (rising from 33%), embedded in a coherent sibling CDF (55%→79.5%, 65%→15.5%). The brief's trackers put OpenAI's best at roughly 54.7% with rivals at or near 59-65%, so 60% is within the demonstrated industry ceiling but OpenAI has consistently trailed the frontier by 5-10 points and is de-emphasizing HLE in marketing. Saturation-aware modeling gives ~20-35%, while the sheer release cadence and remaining time argue for something closer to the market. Resolution-source ambiguity (tools vs no-tools, no live leaderboard captured) adds downside risk, so I stay close to the market anchor rather than moving materially.

claude-fable-5-1

42% conf

Yes	39%
No	61%

The best available anchor is Polymarket at 42% YES (rising from 33%), with a coherent sibling CDF (55%→79.5%, 65%→15.5%). Current evidence puts OpenAI's best model (GPT-6 'Astra', Sept 2026) at ~54.7%, roughly 5 points short with ~3 months left, and OpenAI has consistently trailed Anthropic's frontier by 5–10 points throughout 2026 while de-emphasizing HLE in its marketing. Saturation-aware modeling gives ~20–35%, and each new capability tier has shown multi-month plateaus, which argues for a slight discount to the market; offsetting this is OpenAI's rapid release cadence (3+ major models in 2026) and tracker disagreement (some show frontier models at 59–65%, so a tool-augmented OpenAI score could plausibly clear 60% on the official leaderboard). I land modestly below the market at 39% YES.

gpt-5.6-sol

58% conf

Yes	38%
No	62%

The best available market anchor is Polymarket at 42% YES, but third-party trackers place GPT-6 Astra around 54.7%, leaving a meaningful 5.3-point gain with only about three months remaining. OpenAI’s rapid progression from 25.0% in August 2025 to 43.1% in April 2026 and 54.7% in September 2026 keeps 60% attainable, while the lack of a confirmed successor to Astra and evidence of capability plateaus favor No. Anthropic’s roughly 59–65% results show the threshold is technically reachable, but OpenAI has consistently trailed that frontier and appears to be de-emphasizing HLE. I move modestly below the 42% anchor because saturation-aware estimates cluster around 20–35%, while retaining substantial YES probability due to tracker-methodology disagreement and the incomplete official leaderboard fetch.

gpt-5.6-sol

58% conf

Yes	38%
No	62%

The best available market anchor is Polymarket at 42% YES, with no direct Kalshi price. Third-party trackers place OpenAI’s September 2026 high near 54.7%, so another roughly 5.3-point gain is required despite no confirmed successor to GPT-6 Astra before year-end. Rapid progress from GPT-5.5’s 43.1% and competitors reaching the 60% frontier keeps YES plausible, but decelerating gains, OpenAI’s persistent lag behind Anthropic, and saturation-aware estimates of 20–35% justify moving modestly below the market anchor. Confidence is limited by the unavailable official leaderboard and substantial methodological disagreement among third-party trackers.

6. Final plain mean of 5 draws

40% (45% confidence)

Yes: 40% No: 60%

Mean of 5 draws (0.40, 0.43, 0.39, 0.38, 0.38). No Kalshi-direct price exists, so the Polymarket 42% YES (rising from 33%) is the primary anchor, and its sibling buckets (55%→79.5%, 65%→15.5%) form a coherent CDF around it. The brief places OpenAI's best model (GPT-6 'Astra') at ~54.7% as of Sept 2026, consistently trailing Anthropic's 55–65% frontier, while a saturation-aware model gives only ~20–35% for a further +5.3-point jump in the remaining months. Countervailing factors are OpenAI's rapid release cadence and the tools-vs-no-tools ambiguity on agi.safe.ai (tool-augmented scores run ~15+ points higher), which makes a ≥60% official entry more reachable than the no-tools trajectory suggests. Netting these, I shade slightly below the market at 40% YES; confidence is limited because the primary leaderboard was not captured and third-party trackers disagree by 10+ points on absolute scores.

Pipeline Timing

Total pipeline time: 149.5s

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