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

date	predicted	market price	confidence
2026-08-30	33%	58%	46%
2026-08-21	33%	60%	44%
2026-08-12	43%	50%	43%

1. Planner Fable 5.1 · Fermi decomposition + tool plan

Tools chosen: polymarket_direct resolution_source polymarket_related kalshi_related claude_news article_search wikipedia code_execution

Sub-questions (Fermi decomposition)

What is the highest HLE Accuracy currently shown for any Anthropic Claude model on the official agi.safe.ai leaderboard, and does the leaderboard distinguish tool-augmented vs. no-tool scores?
What is the highest HLE score achieved by any lab's model to date (e.g., GPT-5.x, Gemini 3, Grok 4 Heavy), and how fast have top scores climbed over the past 12 months?
How often does the agi.safe.ai leaderboard get updated with new Anthropic models, and does Anthropic self-report HLE in its model cards at levels the leaderboard adopts?
What Anthropic model releases (Claude Opus/Sonnet successors) are expected in 2026, and what HLE scores has Anthropic reported for its latest models?
What do related Polymarket/Kalshi markets imply about the distribution of the top Claude HLE score in 2026 (other thresholds like 40%, 50%, 70%) and the top overall HLE score?
Has any model with tool use/search already exceeded 60% on HLE, indicating the threshold is reachable within the 2026 timeframe?

Planner reasoning

This is a Polymarket question resolving on the official HLE leaderboard at agi.safe.ai, so the market price and the current leaderboard state are the two key anchors. The key drivers are the current best Claude HLE score on the official leaderboard, the rate of improvement across frontier labs (OpenAI/Google/xAI scores as a ceiling proxy), whether the leaderboard even lists Claude models with tools vs. no-tools, and Anthropic's expected model release cadence through 2026.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.3s	1	## This Market's Polymarket Data Will the highest score achieved by an Anthropic Claude model on Humanity’s Last Exam in 2026 be 60% or higher? - Current price (probability): 33.50% - 7-day price change: -8.50% - 30-day price change: -31.50% - Total volume: $47,564 (USD notional) - Price range:
resolution_source	OK	5.2s	1	already_resolved=unclear; matches_criteria=True; last_updated=01/28/2026 (most recent news item shown). The fetched page content shows only the HLE homepage description, news updates, and dataset contributor list. It does not display the actual leaderboard rankings or any model's HLE Accuracy score,
polymarket_related	OK	0.5s	24	Polymarket search over 3 keyword(s), kept 24 matches. keyword "Humanity's Last Exam": 8 markets \| keyword 'HLE': 8 markets \| keyword 'Claude': 8 markets
kalshi_related	OK	4.6s	2	2 related markets / summaries. keyword "Humanity's Last Exam": ok \| keyword 'Claude': no matches \| keyword 'Anthropic': ok
claude_news	OK	27.7s	8	Here are the key findings on Claude models' Humanity's Last Exam (HLE) performance as of September 2026: Anthropic's official/system-card reported scores: - Claude Opus 4.5 (Nov 2025): Opus 4.5 scores 30.8% without search and 43.2% with search on Humanity's Last Exam, slightly ahead of GPT
article_search	OK	1.4s	30	Fetched 30 unique articles across 4 queries (lookback=120d). "Humanity's Last Exam Claude score": 10 articles \| 'HLE leaderboard new state of the art': 10 articles \| 'Anthropic Claude Opus new model benchmark': 10 articles \| "Humanity's Last Exam 60% accuracy": 10 articles
wikipedia	OK	0.4s	2	Fetched 2 Wikipedia entries (0 missing pages).
code_execution	OK	52.4s	0	## Key Findings - Historical fits (frontier scores, tool-augmented, % accuracy on HLE): overall best-model logistic fit gives ceiling L≈56.6%, growth rate k≈0.258, inflection t0≈5.7 months (Jan 2025=t0); linear fit ≈ 3.26 pts/month + 9.8% intercept. - Claude-only trajectory (≈5%→8%→10%→14%→

3. Evidence Brief Sonnet 5 · 7506 chars

# Current state The market resolves on the highest HLE score any Claude model achieves by Dec 31, 2026, per the official agi.safe.ai leaderboard. As of mid/late-Sept 2026, Anthropic's own system card for Claude Opus 4.8 (May 2026) reports 57.9% with tools (49.8% without), and independent tracker Artificial Analysis shows newer "Fable 5.1" at 59.1% with tools — both just under 60% and unconfirmed on the resolution source itself. # Timeline of key events - 2025-01: HLE launched by Center for AI Safety/Scale AI; early scores single-digit (GPT-4o 2.7%, Claude 3.5 Sonnet 4.1%). [wikipedia, claude_news] - 2025-11: Claude Opus 4.5 released — 30.8% no-tools / 43.2% with search (vs Gemini 3: 37.5%/45.8%). [claude_news, confirmed via LessWrong/system-card citation] - 2026-02: Claude Opus 4.6 released; Anthropic claims it "leads all frontier models on HLE" (exact % not given in excerpt). [claude_news, confirmed] - 2026-05: Claude Opus 4.8 released — 49.8% no-tools / 57.9% with tools. [claude_news/DataCamp, confirmed] - 2026-06: Claude Fable 5 (Mythos-class, public) launched; Mythos 5 limited release; export-control changes announced. [CNBC, confirmed] - 2026-09: Claude Fable 5.1 released; Artificial Analysis reports 59.1% (max effort, with tools), 58.7% (xhigh effort) — highest credible Claude figure found, not yet 60%. [claude_news/Artificial Analysis, reported — methodology differs from Scale's official leaderboard] - 2026-09 (concurrent): Scale AI's official leaderboard shows GPT-6 Astra (OpenAI) leading at 54.80%, with Fable 5.1 (xhigh) at only 46.50% — a materially lower figure, likely reflecting no-tool or stricter scoring. [Scale Labs leaderboard, confirmed but conflicts with Artificial Analysis] - 2026-01-28: agi.safe.ai homepage last news update; actual leaderboard table not retrievable from fetched excerpt. [resolution_source tool] # Event Will any Anthropic Claude model score ≥60% on Humanity's Last Exam (per agi.safe.ai) by Dec 31, 2026? # Outcomes to forecast Yes / No # Kalshi market anchor No kalshi_direct data was returned for this ticker. The closest anchor is Polymarket's identical-question market: current YES ≈33.5%, down sharply from a 66.5% high and from ~42% a week ago (7d: −8.5pts; 30d: −31.5pts). Volume $47.6k over 60 days — moderate liquidity, meaningful downtrend suggesting the market has grown more skeptical as scores stall near 57-59%. # Resolution source agi.safe.ai homepage fetched (status 200, last news item dated 01/28/2026) but the actual leaderboard table/scores were not present in the retrieved content — cannot confirm current Claude score directly from the primary source. Condition not confirmed met or unmet from this source alone. # Sub-question answers 1. **Leaderboard's current top Claude score / tool distinction?** Not visible on agi.safe.ai excerpt; Scale AI's own leaderboard (likely underlying data source) shows Fable 5.1 (xhigh) at 46.50±2.00% — notably below Anthropic's/Artificial Analysis's self-reported tool-augmented figures, implying the leaderboard uses no-tool or stricter conditions than lab press releases. [Scale Labs, claude_news] 2. **Highest score by any lab / pace of growth?** GPT-6 Astra leads the official Scale leaderboard at 54.80%; Artificial Analysis (independent, tool-augmented methodology) shows Fable 5.1 at 59.1%, highest across trackers. Scores rose from ~3-4% (early 2025) to ~50-59% by Sept 2026 — roughly 3+ points/month, an extraordinarily fast climb. [claude_news, code_execution] 3. **Leaderboard update cadence / self-reporting gap?** Scale's leaderboard is actively updated with new frontier models (GPT-6 Astra, Fable 5.1 both marked "NEW" as of Sept 2026), but its scores for Claude are lower than Anthropic's own system-card-reported tool-augmented figures — indicating a persistent reporting/methodology gap. [claude_news] 4. **Expected 2026 Claude releases / reported scores?** Opus 4.6 (Feb), Opus 4.8 (May, 57.9% w/tools), Fable 5 (June, public), Fable 5.1 (Sept, 59.1% per Artificial Analysis), plus Mythos (limited). Next Opus expected by Sept–Dec 2026 per Polymarket (90-98% probability). [claude_news, polymarket_related] 5. **Related markets' implied distribution?** Polymarket ladder for Claude: 55%→86.5% YES, 60%→33.5%, 65%→18.5%, 70%→11.5% — implies mode of outcome distribution centered ~55-62%. Comparable Meta-model thresholds are far lower (55%→26%, 60%→18%), showing markets expect Claude to clearly outperform Meta but still see 60%+ as a coinflip-or-worse. 6. **Has any tool-augmented model already exceeded 60%?** No model has confirmed >60% on the resolution-relevant leaderboard; the closest is Fable 5.1 at 59.1% (Artificial Analysis, tools+max effort) — suggesting the threshold is very close but not yet crossed as of Sept 2026. # Key facts (high-confidence, factual) 1. [claude_news/DataCamp] Opus 4.8 (May 2026): 57.9% with tools, 49.8% without. 2. [claude_news/Artificial Analysis] Fable 5.1 (Sept 2026): 59.1% (max effort+tools) — highest credible Claude figure to date. 3. [Scale Labs] Official leaderboard shows Fable 5.1 at only 46.5%, GPT-6 Astra at 54.8% — conflicting/lower than lab self-reports. 4. [Polymarket] Same-question market YES=33.5%, down from 66.5% high, reflecting waning confidence as scores plateau near 57-59%. # Cross-market signals - Kalshi related: no direct Claude/HLE Kalshi market found; unrelated Anthropic markets (IPO race 94%, RSI announcement 17%) don't inform HLE directly. - Polymarket: identical question at 33.5% YES; higher Claude thresholds (65%→18.5%, 70%→11.5%) and lower thresholds (55%→86.5%) form a coherent ladder implying median outcome ~58-62%. - Sportsbook implied: N/A. # Analyst opinions and speculation - code_execution Monte Carlo (frontier saturation + Claude-lag model): central P(≥60%) ≈30-36%, sensitive to assumed lag-to-frontier (10-50% depending on assumptions); naive linear extrapolation of Claude's own historical trend implies ~0%, while optimistic logistic fits imply ~62%. - claude_news synthesis: "crossing 60% before end of 2026 appears plausible but not yet confirmed," given ~7-10pt gains per release cycle and current 57.9-59.1% ceiling. # Directional lean per outcome - **Yes**: Momentum extremely strong (near-vertical HLE gains 2025-2026); Fable 5.1 already at 59.1% w/tools; more releases (next Opus/Fable) expected before year-end; Anthropic increasingly emphasizes tool-augmented scoring. - **No**: Official Scale/agi.safe.ai leaderboard shows much lower Claude figures (46.5%) than lab press claims; gains may be plateauing (Opus 4.8→Fable 5.1 only +1.2pts); resolution requires the *official* leaderboard number, which lags self-reported figures; Polymarket has fallen sharply (-31.5pts/30d), suggesting real-time sentiment turned skeptical. # Gaps / unknowns - Cannot verify actual current agi.safe.ai leaderboard entry for Claude (fetch failed to return table). - Ambiguity whether "official" leaderboard will adopt tool-augmented scores (which are higher) or stick to no-tool/lower scores — pivotal for resolution. - No visibility into whether further major Claude releases (e.g., Fable 5.2, Opus 5) are planned before Dec 31, 2026. # Calibration anchors - Polymarket YES for identical question: 33.5% (down from 66.5% peak) — primary cross-market anchor. - Precedent: benchmark scores on hard evals (e.g., FrontierMath, ARC-AGI) have shown similar rapid-then-plateauing patterns; official leaderboards often lag self-reported/marketing figures by 5-15 points.

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

claude-fable-5-1

47% conf

Yes	30%
No	70%

The identical Polymarket question sits at ~33.5% YES and has fallen 31.5 points over 30 days, reflecting a plateau in Claude scores (Opus 4.8 at 57.9% w/tools → Fable 5.1 at 59.1% w/tools, only +1.2pts per cycle). Critically, the resolution source (agi.safe.ai / Scale's official leaderboard) shows Fable 5.1 at just 46.5% and the overall leader GPT-6 Astra at 54.8% — no model is within striking distance of 60% on the official methodology, and official figures persistently lag lab self-reports by 5-15 points. A Yes requires either a large jump from a late-2026 Claude release (plausible given Anthropic's cadence) or the leaderboard adopting tool-augmented scoring, both uncertain within ~3 months. I lean slightly below the market anchor given the resolution-source gap and the Monte Carlo central estimate of ~30-36%.

claude-fable-5-1

44% conf

Yes	29%
No	71%

The Polymarket identical-question anchor sits at ~33.5% YES and has fallen sharply (-31.5pts over 30 days) as Claude scores plateau just under the threshold (Opus 4.8 57.9% and Fable 5.1 59.1% with tools, per lab/Artificial Analysis reporting). Crucially, the resolution source (agi.safe.ai / Scale leaderboard) shows Fable 5.1 at only 46.5% and the overall leader GPT-6 Astra at 54.8%, meaning the official number lags self-reported tool-augmented figures by 10+ points and no model is yet near 60% on the resolving source. A Yes requires either a large late-2026 Claude release (next Opus/Fable, highly likely to ship) that jumps ~13 points on the official leaderboard, or the leaderboard adopting tool-augmented scoring — plausible given ~3pts/month historical gains, but the recent +1.2pt step suggests deceleration. I shade slightly below the market given the source-vs-lab methodology gap and the 3-month remaining window.

claude-fable-5-1

47% conf

Yes	30%
No	70%

The Polymarket identical-question anchor sits at ~33.5% YES and has fallen sharply (−31.5pts over 30 days), reflecting that Claude's best figures have stalled just under the bar (Opus 4.8 at 57.9% with tools, Fable 5.1 at 59.1% per Artificial Analysis, only +1.2pts in four months). Critically, the resolution source is the official agi.safe.ai/Scale leaderboard, which lists Fable 5.1 at only 46.5% and the overall leader (GPT-6 Astra) at 54.8% — no model is within striking distance of 60% under the official methodology, and official leaderboards historically lag lab-reported tool-augmented numbers by 5-15 points. With roughly one more major Claude release plausible before Dec 31 and strong long-run momentum, a jump past 60% on the official board is possible but requires either a large step-change or the leaderboard adopting tool-augmented scoring. I shade slightly below the market anchor because of the source-vs-world gap.

gpt-5.6-sol

64% conf

Yes	34%
No	66%

The closest market anchor is Polymarket's 33.5% YES price, and the evidence does not justify moving far from it. Fable 5.1's 59.1% tool-augmented result and the likelihood of another 2026 Claude release make crossing 60% plausible. However, the resolution-relevant Scale leaderboard reportedly scores Fable 5.1 at only 46.5%, indicating a substantial methodology gap that a modest model upgrade may not overcome. The recent slowdown from 57.9% to 59.1% on the more favorable tracker and the market's sharp downward trend support keeping No favored.

gpt-5.6-sol

58% conf

Yes	34%
No	66%

The 33.5% Polymarket YES price is the best available consensus anchor, and I move only slightly upward to 34%. Fable 5.1's 59.1% tool-augmented result and the likelihood of another Claude release before year-end make crossing 60% technically plausible. However, the resolution-relevant Scale leaderboard reportedly places Fable 5.1 at only 46.5%, indicating a large methodology gap that a modest model update may not overcome. The recent slowdown from 57.9% to 59.1% on favorable external testing, combined with the market's sharp decline, keeps No favored despite rapid longer-term benchmark progress.

6. Final plain mean of 5 draws

31% (52% confidence)

Yes: 31% No: 69%

Mean of 5 draws (0.30, 0.29, 0.30, 0.34, 0.34). The identical Polymarket question sits at ~33.5% YES and has fallen 31.5 points over 30 days, reflecting a plateau in Claude scores (Opus 4.8 at 57.9% w/tools → Fable 5.1 at 59.1% w/tools, only +1.2pts per cycle). Critically, the resolution source (agi.safe.ai / Scale's official leaderboard) shows Fable 5.1 at just 46.5% and the overall leader GPT-6 Astra at 54.8% — no model is within striking distance of 60% on the official methodology, and official figures persistently lag lab self-reports by 5-15 points. A Yes requires either a large jump from a late-2026 Claude release (plausible given Anthropic's cadence) or the leaderboard adopting tool-augmented scoring, both uncertain within ~3 months. I lean slightly below the market anchor given the resolution-source gap and the Monte Carlo central estimate of ~30-36%.

Pipeline Timing

Total pipeline time: 137.5s

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