Scan: Will the highest score achieved by a Moonshot Kimi model on Humanity’s Last Exam in 2026 be 50% or higher?

date	predicted	market price	confidence
2026-09-01	37%	45%	50%

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

Sub-questions (Fermi decomposition)

What is the highest HLE accuracy currently listed for any Moonshot Kimi model on the official agi.safe.ai leaderboard, and does the leaderboard include tool-augmented results?
What HLE scores have Kimi K2 Thinking / K2.5 or newer Kimi models achieved (self-reported vs. independently evaluated), and how do self-reported numbers compare to leaderboard entries?
What is the current top HLE score across all labs (e.g., GPT-5.x, Gemini 3, Grok 4 Heavy) and how fast have top scores risen over the past 12 months?
What Kimi model releases (K3, K2.5 updates) does Moonshot plan or is expected to ship before end of 2026, and how does Moonshot typically lag frontier US labs on reasoning benchmarks?
How frequently does the agi.safe.ai leaderboard get updated with new/Chinese models, and is there a risk a qualifying Kimi score is never posted there?
What do sibling Polymarket markets (other thresholds for Kimi, and same thresholds for other labs) imply about the probability distribution?

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 primary anchors. Key uncertainties are Kimi's current HLE score (Kimi K2 Thinking reportedly ~44.9% with tools), whether the official leaderboard lists Kimi models (and which config — with/without tools), the rate of HLE score improvement among frontier labs, and Moonshot's expected 2026 release cadence.

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 Moonshot Kimi model on Humanity’s Last Exam in 2026 be 50% or higher? - Current price (probability): 36.50% - 7-day price change: +2.50% - 30-day price change: -12.50% - Total volume: $23,165 (USD notional) - Price range: 24.
resolution_source	OK	4.6s	1	already_resolved=unclear; matches_criteria=True; last_updated=01/28/2026 (latest news item shown). The fetched page content does not display any leaderboard scores or model-specific HLE accuracy figures, including for Moonshot Kimi models; it only shows news updates, dataset info, and contributor li
polymarket_related	OK	0.8s	16	Polymarket search over 3 keyword(s), kept 16 matches. keyword "Humanity's Last Exam": 8 markets \| keyword 'Kimi': 8 markets \| keyword 'HLE 2026': 0 markets
kalshi_related	OK	5.5s	1	1 related markets / summaries. keyword "Humanity's Last Exam": ok \| keyword 'Kimi': no matches \| keyword 'HLE': no matches
claude_news	OK	25.4s	10	Based on research, here are key findings on Moonshot Kimi models' HLE performance and the broader 2026 landscape: Kimi K2 Thinking (released Nov 2025): - Moonshot self-reported K2 Thinking achieved a state-of-the-art score of 44.9%, with search, python, and web-browsing tools, establishing new
article_search	OK	46.5s	29	Fetched 29 unique articles across 3 queries (lookback=120d). "Kimi K2 Thinking Humanity's Last Exam score": 10 articles \| 'Moonshot AI Kimi new model release benchmark': 10 articles \| "Humanity's Last Exam leaderboard top score 2026": 10 articles
gdelt_news	OK	85.5s	20	GDELT: 20 articles across 2 queries (lookback=90d). "Moonshot Kimi Humanity's Last Exam": 10 hits \| 'Kimi K3 release': 10 hits
wikipedia	OK	0.4s	3	Fetched 3 Wikipedia entries (0 missing pages).

3. Evidence Brief Sonnet 5 · 6628 chars

# Current state The resolution hinges on the official agi.safe.ai HLE leaderboard showing a Kimi model at ≥50% accuracy at any point through Dec 31, 2026. As of the latest verified leaderboard data (independent, no-tools basis), Moonshot's best Kimi model (K2.5) scores only 24.4% — roughly half the required threshold — while even the current frontier proprietary models (Claude Fable 5.1, GPT-6 Astra, Gemini 3.1 Pro) are only just crossing into the 50-59% range in late 2026. # Timeline of key events - 2025-07: Kimi K2 (1T params) released. (confirmed, Wikipedia) - 2025-11: Kimi K2 Thinking released; Moonshot self-reports 44.9% HLE with search/python/web tools; Artificial Analysis independently measures 22.3% (no tools) — best open-weights score at the time. (confirmed self-report + independent eval, kimi.com/artificialanalysis.ai) - 2026-01: Kimi K2.5 released; verified as top open model on HLE at 24.4%, ranked #14 overall across all models. (confirmed, llmrun.dev/airank.dev) - 2026-01-28: agi.safe.ai leaderboard page last shows news update as of this date; actual score table not visible in fetch. (resolution_source tool) - 2026-04: Kimi K2.6 released. (confirmed, Wikipedia) - 2026-07-16/27: Kimi K3 (2.8T params, largest open-weights model) announced/weights released; Moonshot's own launch materials state it still trails top proprietary models; no confirmed HLE score reported. (confirmed release, reported capability gap — TechCrunch, dev.to, Wikipedia) - 2026-08-16: Scale AI leaderboard shows top scores in low-to-mid 40s (Gemini 3.1 Pro 46.44%, GPT-5.4 Pro 44.32%); no Kimi entry near top. (confirmed, per claude_news synthesis) - 2026-09: Frontier leaders (Claude Fable 5, Opus 5, GPT-6 Astra, Claude Fable 5.1) reach 54.7–59.1% HLE; Kimi K2.5 remains the top open-weights model at 24.4%, far below frontier. (confirmed via AA/pricepertoken) # Event Will any Moonshot Kimi model score ≥50% on Humanity's Last Exam (per agi.safe.ai) at any point in 2026? # Outcomes to forecast Yes / No # Kalshi market anchor No direct Kalshi ticker data was returned for this specific market; Polymarket is used as the primary cross-market anchor: **36.5% YES** (current), up +2.5% over 7 days, down -12.5% over 30 days, range 24.5–60.5%, volume $23,165. No related Kalshi markets exist for Kimi/HLE specifically (kalshi_related found none). # Resolution source agi.safe.ai fetched (200, 40,690 chars); page shows news items, dataset info, contributor lists — no visible leaderboard scores in this fetch. Last news item dated 01/28/2026. Cannot confirm current Kimi HLE score directly from this fetch; independent trackers (Artificial Analysis, llmrun.dev, Scale AI) are used as proxies. Condition is NOT confirmed met. # Sub-question answers 1. **Current Kimi leaderboard score / tool-augmented inclusion?** No direct visibility into agi.safe.ai's table in this fetch; independent trackers show Kimi K2.5 at 24.4% (no tools, verified) — the resolution criteria specify "HLE Accuracy" metric, likely no-tools basis. (llmrun.dev, airank.dev) 2. **K2 Thinking/K2.5+ scores, self-reported vs. independent?** K2 Thinking self-reported 44.9% (tool-augmented); Artificial Analysis independently measured only 22.3% (no tools) — a ~22pp gap. K2.5 independently verified at 24.4%. (kimi.com, artificialanalysis.ai) 3. **Current frontier top score & rate of rise?** As of Sept 2026, top scores are Claude Fable 5.1 at 59.1% (AA) / Claude Fable 5 at 55.5%, Opus 5 54.9%, GPT-6 Astra 54.7% (pricepertoken.com). Scale AI's Aug 2026 leaderboard had leaders around 44-46%, implying rapid ~10pp rise within a month via new releases. (claude_news synthesis) 4. **Future Kimi releases & lag vs. frontier?** Kimi K3 (2.8T params) shipped July 2026; Moonshot itself states K3 still trails top proprietary models. No K3 HLE score confirmed. Historically Kimi trails ~10-30pp behind frontier leaders on HLE. (TechCrunch, dev.to, benchlm.ai) 5. **Leaderboard update frequency/risk of no posting?** Not directly answered; agi.safe.ai fetch shows recent news activity (Jan 2026), suggesting active maintenance, but no confirmation of update cadence for Chinese models specifically. 6. **Sibling Polymarket markets implied distribution?** Kimi 45%+ trades at 85%, 50%+ at 36.5%, 55%+ at 17.5%, 60%+ at 10.5%, 65%+ at 6.75% — implying market expects Kimi's ceiling most likely lands in the 45-50% band. Comparable Gemini Pro debut markets are priced much higher (81-91% for 40-50% thresholds), reflecting Kimi's known lag vs. top US labs. # Key facts 1. [llmrun.dev] Kimi K2.5 verified at 24.4% HLE, top open model, rank #14 overall. 2. [artificialanalysis.ai] K2 Thinking's no-tools score is 22.3%; tool-augmented self-report is 44.9%. 3. [Wikipedia/TechCrunch] Kimi K3 (July 2026) is largest open-weights model but Moonshot admits it trails proprietary frontier models. 4. [pricepertoken/AA] Frontier ceiling now 54.7-59.1% (Sept 2026), still climbing. # Cross-market signals - Polymarket (same market): 36.5% YES, declining 30-day trend (-12.5%) despite recent 7-day uptick. - Polymarket siblings: Kimi 45%+ at 85%, 55%+ at 17.5% — steep drop-off between 45% and 55% thresholds, consistent with 50% being a genuine coin-flip-ish pivot point. - No Kalshi-specific comparables found. # Analyst opinions and speculation - TechCrunch (reported): K3 "expected to perform at par with or surpass Opus 4.8" — general capability claim, not HLE-specific. - Moonshot's own materials (reported): K3 still trails top proprietary models overall. # Directional lean per outcome - **Yes**: Rapid Kimi release cadence (K2→K2.5→K2.6→K3 within ~12 months); tool-augmented scores already reached 44.9%; frontier ceiling rising fast, narrowing the gap if Moonshot ships another major model (K3.x/K4) by year-end with tool use counted. - **No**: Independently verified (likely resolution-relevant) scores remain at 24.4%, less than half the threshold; even with tools K2 Thinking only reached 44.9%, still under 50%; Moonshot self-admits K3 trails top labs; historical lag pattern suggests Kimi unlikely to leapfrog into 50%+ territory within the remaining 2026 window. # Gaps / unknowns - No confirmed K3 HLE score. - Unclear whether agi.safe.ai leaderboard counts tool-augmented scores (critical for whether 44.9% self-report could ever qualify). - Uncertain leaderboard update frequency for Chinese labs. # Calibration anchors - Polymarket YES ≈ 36.5% (current anchor). - Verified Kimi ceiling (24.4%) vs. threshold (50%) — a 25.6pp gap remains, larger than the ~15-20pp/12-month rate of the fastest movers, but K3 unscored and could be a wildcard.

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

claude-fable-5-1

47% conf

Yes	32%
No	68%

The Polymarket anchor sits at 36.5% YES with a declining 30-day trend, and the verified no-tools Kimi ceiling (K2.5 at 24.4%) is barely half the 50% threshold. Even Moonshot's tool-augmented self-report for K2 Thinking (44.9%) fell short, and Moonshot itself states K3 still trails proprietary frontier models, which only recently crossed 50-59%. The main path to YES is a late-2026 K3.x/K4 release scored with tool use and posted to agi.safe.ai, which the sibling markets (45%+ at 85%) suggest traders consider plausible, but the unresolved question of whether tool-augmented scores count plus the historical 10-30pp lag argues for landing modestly below the market anchor.

claude-fable-5-1

45% conf

Yes	30%
No	70%

The Polymarket anchor sits at 36.5% YES with a declining 30-day trend, and the brief's evidence tilts further toward No: the best independently verified Kimi score (K2.5) is 24.4% no-tools, less than half the 50% threshold, and even Moonshot's tool-augmented self-report for K2 Thinking (44.9%) fell short. Kimi K3 (July 2026) is an unscored wildcard, but Moonshot itself concedes it trails top proprietary models, which only just reached 54-59% in Sept 2026, and Kimi has historically lagged the frontier by 10-30pp on HLE. The sibling markets (45%+ at 85%) suggest traders expect a strong K3/K4 or tool-inclusive counting, so I keep meaningful YES mass for a late-2026 release plus possible tool-augmented leaderboard entries, but shade below the market given the verified-score gap and resolution-source ambiguity.

claude-fable-5-1

40% conf

Yes	36%
No	64%

Independently verified Kimi scores (K2.5 at 24.4% no-tools) sit far below 50%, and Moonshot itself concedes K3 trails frontier proprietary models, which only reached 54-59% in Sept 2026. However, the Polymarket sibling ladder (45%+ at 85%, 50%+ at 36.5%, 55%+ at 17.5%) strongly implies the resolving leaderboard counts tool-augmented self-reports like K2 Thinking's 44.9%, meaning a K3-Thinking-style release with tools by year-end could plausibly cross 50%. With no confirmed K3 HLE score and no visibility into the agi.safe.ai table, I anchor close to the 36.5% cross-market price, nudging slightly down for the declining 30-day trend and the large verified gap.

gpt-5.6-sol

67% conf

Yes	30%
No	70%

The 24.4% independently verified Kimi K2.5 score points toward No because it leaves a 25.6-point gap to the resolution threshold. K2 Thinking’s 44.9% tool-assisted self-report offers some upside, but its comparable no-tools result was only 22.3%, and it is unclear whether tool-assisted scores qualify on agi.safe.ai. K3’s unreported score and Moonshot’s rapid release cadence preserve a meaningful chance of a late breakthrough, although Moonshot acknowledges K3 still trails proprietary frontier models. I place Yes below the 36.5% Polymarket anchor because the remaining gap, missing K3 evidence, and leaderboard-posting uncertainty outweigh the recent rapid rise in frontier scores.

gpt-5.6-sol

67% conf

Yes	32%
No	68%

The 24.4% independently verified K2.5 score points toward No because it leaves a 25.6-point gap, while K2 Thinking’s 44.9% was tool-augmented and only 22.3% under independent no-tools evaluation. Rapid frontier improvement and Moonshot’s release cadence leave a credible Yes path, but K3 has no confirmed HLE score and reportedly still trails leading proprietary models. Leaderboard-update and tool-inclusion uncertainty further weaken the chance that a qualifying score appears on agi.safe.ai before year-end. I therefore move modestly below the 36.5% Polymarket YES anchor to 32%.

6. Final plain mean of 5 draws

32% (53% confidence)

Yes: 32% No: 68%

Mean of 5 draws (0.32, 0.30, 0.36, 0.30, 0.32). The Polymarket anchor sits at 36.5% YES with a declining 30-day trend, and the verified no-tools Kimi ceiling (K2.5 at 24.4%) is barely half the 50% threshold. Even Moonshot's tool-augmented self-report for K2 Thinking (44.9%) fell short, and Moonshot itself states K3 still trails proprietary frontier models, which only recently crossed 50-59%. The main path to YES is a late-2026 K3.x/K4 release scored with tool use and posted to agi.safe.ai, which the sibling markets (45%+ at 85%) suggest traders consider plausible, but the unresolved question of whether tool-augmented scores count plus the historical 10-30pp lag argues for landing modestly below the market anchor.

Pipeline Timing

Total pipeline time: 162.5s

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