Scan: Will the next Google Gemini Pro model debut with a Humanity’s Last Exam score of 50% or higher?

date	predicted	market price	confidence
2026-08-24	60%	70%	52%
2026-08-14	53%	72%	43%

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 current HLE Accuracy displayed at agi.safe.ai for the latest Gemini Pro model (e.g., Gemini 3 Pro / 3.1 Pro / 3.5 Pro), and which configuration (with/without tools) is shown on the card?
What has been the step-up in HLE accuracy between successive Gemini Pro releases (2.5 Pro → 3 Pro → 3.1 Pro etc.), and what do frontier competitors (GPT-5.x, Claude, Grok) currently score on HLE?
Has Google announced or is it rumored to release a new Gemini Pro model (e.g., Gemini 3.5 Pro or 4.0 Pro) before December 31, 2026, and how quickly does CAIS typically add new Gemini models to the HLE site after release?
What HLE scores has Google itself reported for its latest Gemini Pro/Deep Think models, and do CAIS's independent evaluations typically match or fall below vendor-reported numbers?
What are the Polymarket prices for sibling threshold markets (e.g., 40%, 45%, 55%, 60%) for the next Gemini Pro HLE debut, and do they imply a coherent distribution?
Is there risk that no qualifying Pro model is added by the deadline, or that a GA promotion of an existing model appears with a lower score than expected?

Planner reasoning

The question hinges on (a) what the current top Gemini Pro model scores on HLE at agi.safe.ai and the trajectory of score gains between successive Pro releases, (b) whether a new Gemini Pro model is expected to be released and added to the leaderboard before Dec 31, 2026, and (c) how CAIS displays scores (tools vs no-tools configuration, which card counts). The Polymarket price is the primary anchor; sibling markets at other thresholds and the current leaderboard state provide the key calibration.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.1s	1	## This Market's Polymarket Data Will the next Google Gemini Pro model debut with a Humanity’s Last Exam score of 50% or higher? - Current price (probability): 67.50% - 7-day price change: -3.50% - 30-day price change: -3.00% - Total volume: $21,721 (USD notional) - Price range: 50.00% - 91.00%
resolution_source	OK	4.1s	1	already_resolved=unclear; matches_criteria=True; last_updated=01/28/2026. The fetched page shows the Humanity's Last Exam homepage with news updates, dataset info, and contributor list, but does not display any model leaderboard entries, HLE Accuracy percentages, or any Gemini Pro model results.
polymarket_related	OK	0.6s	20	Polymarket search over 3 keyword(s), kept 20 matches. keyword "Gemini Pro Humanity's Last Exam": 4 markets \| keyword 'Gemini HLE': 8 markets \| keyword 'next Google Gemini': 8 markets
kalshi_related	OK	2.6s	1	1 related markets / summaries. keyword 'Gemini': no matches \| keyword "Humanity's Last Exam": ok \| keyword 'Google AI model': no matches
claude_news	OK	35.3s	17	Here are key findings on HLE scores and the next Gemini Pro model's likely debut score: Gemini's HLE track record (no-tools scores): - Gemini 3 demonstrates PhD-level reasoning with top scores on Humanity's Last Exam (37.5% without the usage of any tools). Source: blog.google - Gemini 3 Deep
article_search	OK	1.3s	25	Fetched 25 unique articles across 4 queries (lookback=90d). "Gemini Pro Humanity's Last Exam score": 10 articles \| 'Google Gemini 3.5 Pro release': 10 articles \| 'Gemini 4 Pro announcement': 10 articles \| "Humanity's Last Exam leaderboard frontier model 2026": 10 articles
gdelt_news	OK	74.4s	10	GDELT: 10 articles across 2 queries (lookback=60d). "Gemini Pro Humanity's Last Exam": error GDELT rate-limited after retries (429) \| 'Google Gemini new model release': 10 hits
wikipedia	OK	0.3s	2	Fetched 2 Wikipedia entries (0 missing pages).

3. Evidence Brief Sonnet 5 · 6631 chars

# Current state Google has not yet shipped a next-generation Gemini "Pro" model as of the latest research (through ~Sept 2026). The current flagship on the market is Gemini 3.1 Pro (Feb 2026), reportedly scoring 44.4% (no tools) / 51.4% (with tools) on HLE per third-party trackers — not confirmed as the figure displayed on the official agi.safe.ai chart. Gemini 3.5 Pro has been promised since May 2026 but has missed multiple targets (June, July, Aug); Google has instead shipped a string of Flash-tier models (3.6, 3.7, 3.8 Flash). The market resolves based on whichever score appears on agi.safe.ai when the *next* Gemini Pro-labeled model debuts — this is a forward-looking, unresolved bet on both timing and score level. # Timeline of key events - 2025-early 2026 (confirmed, blog.google): Gemini 3 Pro debuts, HLE (no tools) 37.5%; Gemini 3 Deep Think, 41.0%. - 2026-02 (reported, aireleasetracker.com): Gemini 3.1 Pro reports HLE 44.4% (no tools) / 51.4% (with tools). - 2026-02-12 (reported, remio.ai): Gemini 3 Deep Think refresh reaches 48.4% (no tools) — Deep Think, not "Pro," likely non-qualifying. - 2026-05-19 (confirmed, Yahoo Finance): Pichai promises Gemini 3.5 Pro "within a month" at I/O. - 2026-06 to 2026-08 (confirmed): Three consecutive missed launch targets for 3.5 Pro (June, mid-July, early August); BigGo reports a slipped July 17 target citing full architecture redesign. - 2026-07-16 (confirmed, CNBC): Alphabet shares fall on 3.5 Pro delay report. - 2026-07-21/22 (confirmed): Google ships Gemini 3.6 Flash and reports Q2 earnings; Pichai says 3.5 Pro still "in testing," compute shifting toward Gemini 4. - 2026-08-13/14 (confirmed): Gemini 3.7 Flash released; flagship still absent. - 2026-09-03/04 (confirmed): Gemini 3.8 Flash/Cyber released; Yahoo notes "four Flash models in 106 days" but no flagship Pro/Deep model yet. # Event Will the next Gemini "Pro" model's HLE Accuracy (as shown on agi.safe.ai) be ≥50% when it first appears? # Outcomes to forecast Yes / No # Kalshi market anchor No kalshi_direct data returned in research (tool did not surface a price for this ticker). Best available same-market cross-listing: Polymarket price for this identical question is **67.5% YES**, down from a high of 91%, trending -3.5% (7d) and -3.0% (30d); volume $21.7k over 51 days — moderate conviction, softening. # Resolution source agi.safe.ai fetched (status 200, last checked 01/28/2026) but the retrieved text only shows the HLE homepage/news/dataset content — the actual leaderboard/chart with model cards was not captured. Cannot confirm whether a qualifying Gemini Pro model is currently listed or what score it shows. Not yet resolved; status quo is indeterminate pending direct chart data. # Sub-question answers 1. **Current HLE score for latest Gemini Pro model** — Gemini 3.1 Pro (Feb 2026) reportedly shows 44.4% no-tools / 51.4% with-tools per aireleasetracker.com; unclear which configuration the official agi.safe.ai card actually displays. [claude_news, resolution_source] 2. **Step-up between releases / competitor scores** — Gemini has gained roughly +4-7 points per generation (37.5%→41.0%→44.4% no-tools). Rivals have since surged past 50%: Claude Fable 5 (55.5%), Claude Opus 5 (54.9%), GPT-6 Astra (54.7%), and Claude Fable 5.1 reportedly at 65.0% (sources conflict on exact leader value). [claude_news/pricepertoken.com/llm-stats.com] 3. **Next Pro model timing** — Gemini 3.5 Pro repeatedly delayed (missed June/July/Aug 2026 targets); as of Sept 2026 still unreleased, with Google shipping only Flash-tier models meanwhile; Gemini 4 pretraining reportedly underway. No explicit data on CAIS's typical lag to add new models to HLE site. [claude_news, gdelt_news] 4. **Vendor vs. independent HLE scores** — Only vendor-reported figures found (Google's own 3.1 Pro numbers); no CAIS-vs-vendor discrepancy data surfaced. General industry pattern (not evidenced here) is independent evals often run at/below vendor self-reports. 5. **Polymarket sibling thresholds** — 40%: 92.5% Yes; 45%: 80% Yes; 50%: 67.5% Yes; 55%: 33% Yes. Monotonic and roughly coherent, implying a fairly smooth probability mass concentrated near 45-55%. [polymarket_related] 6. **Risk of no qualifying model by deadline / GA-promotion risk** — Related Polymarket markets show low probability of Gemini Pro release by Sept 30 (12.5%) but high probability by Oct 31 (83.5%) and Nov 30 (93.5%), suggesting the market expects release in Oct/Nov 2026, leaving buffer before Dec 31 close but with real delay risk given track record. [polymarket_related] # Key facts (high-confidence, factual) 1. [blog.google] Gemini 3 Pro: 37.5% HLE (no tools); Gemini 3 Deep Think: 41.0%. 2. [aireleasetracker.com] Gemini 3.1 Pro: 44.4% (no tools)/51.4% (with tools). 3. [CNBC/TechCrunch/arstechnica] Gemini 3.5 Pro delayed three+ times through Aug 2026; no ETA given. 4. [pricepertoken.com/llm-stats.com] Rival models (Claude, GPT-6 Astra) now score 54-65% on HLE as of Sept 2026. # Cross-market signals - Kalshi related: only an unrelated Last-of-Us market surfaced; no direct Gemini/HLE Kalshi comparables found. - Polymarket: same-question market at 67.5% YES (declining); sibling thresholds show coherent declining curve (40%:92.5%→55%:33%). - Sportsbook implied: N/A. # Analyst opinions and speculation - Claude_news synthesis: 50% threshold is "plausible but not certain," with delay risk as the dominant threat to Yes resolution, given competitive pressure pushing scores up but launch timing uncertain. # Directional lean per outcome - **Yes (≥50%)**: Supported by rising competitor scores (>54%) pressuring Google to leapfrog; 3.1 Pro already reportedly near/at 50% (with tools); Polymarket at 67.5%. - **No (<50%)**: Supported by ambiguity over which score configuration displays on the official chart (no-tools historically lower, ~44%); repeated launch delays increasing chance of a rushed/conservative debut or missed deadline; possible GA promotion scoring lower than hoped. # Gaps / unknowns - No confirmation of the exact score configuration (tools vs. no-tools) shown on agi.safe.ai cards. - No Kalshi-direct price captured for this ticker. - No visibility into actual Gemini 3.5 Pro/4.0 Pro launch date or its true HLE score. - CAIS site-update lag time unknown. # Calibration anchors - Kalshi YES price: unavailable directly; Polymarket same-market proxy = 67.5% (declining trend). - Precedent: Google's historical per-generation HLE gains (~4-7 pts) and current 3.1 Pro reported score (44-51% range) bracket the 50% threshold closely, making this a near-coin-flip/moderate-Yes lean overall.

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

claude-fable-5-1

47% conf

Yes	63%
No	37%

The best market anchor is Polymarket's identical-question price at 67.5% YES (softening from 91%), with a coherent sibling ladder (45%: 80%, 55%: 33%) placing the modal expected score in the high-40s to mid-50s. Gemini 3.1 Pro reportedly sits at 44.4% no-tools / 51.4% with tools, and historical per-generation gains of ~4-7 points plus intense competitive pressure (Claude/GPT rivals at 54-65%) make a ≥50% debut for a redesigned 3.5 Pro/next Pro plausible. Offsetting risks: the agi.safe.ai chart historically displays no-tools figures, independent CAIS evals often land at or below vendor claims, and Google's three missed launch targets leave a non-trivial chance of no qualifying Pro model appearing on the chart before the Dec 31 close (sibling markets imply ~5-10% no-release risk even by late Nov). I shade slightly below the market given the declining trend and configuration ambiguity.

claude-fable-5-1

47% conf

Yes	66%
No	34%

The Polymarket same-question proxy sits at 67.5% YES (no Kalshi-direct price), with coherent sibling thresholds (45%: 80%, 55%: 33%) bracketing 50% as the central mass. Substantively, Gemini 3.1 Pro already reports 44.4% no-tools (51.4% with tools) and the Deep Think refresh hit 48.4%, while rivals now score 54-65%, so a delayed, architecture-redesigned 3.5 Pro debuting below 50% would be a competitive embarrassment—pointing toward Yes on score alone. The dominant No risks are timing (three missed launch targets, ~6-15% implied chance of no Pro release before the Dec 31 close plus unknown CAIS listing lag) and uncertainty over whether the agi.safe.ai card shows a no-tools figure that could land in the high 40s. Netting a ~80% conditional chance of ≥50% given a listed debut against ~15-20% release/listing failure risk lands near the market anchor.

claude-fable-5-1

44% conf

Yes	63%
No	37%

No Kalshi-direct price was available; the identical Polymarket question trades at 67.5% YES, down from 91% and softening, with coherent sibling thresholds (45%: 80%, 55%: 33%). The current Gemini 3.1 Pro reportedly sits at 44.4% no-tools / 51.4% with-tools, so applying Google's historical +4-7 pt per-generation gain lands 3.5 Pro right around 48-51% — a near coin flip on score alone, though the long delay, 'full architecture redesign,' and rivals now at 54-65% argue for a larger leap. Offsetting this are (a) ambiguity over whether agi.safe.ai displays the lower no-tools figure, (b) ~10% risk no Pro model ships (or is listed by CAIS) before Dec 31 given three missed launch targets, and (c) the declining market trend. I shade slightly below the Polymarket anchor to ~63% YES.

gpt-5.6-sol

58% conf

Yes	66%
No	34%

The 67.5% YES Polymarket price is the best available consensus anchor, and I shade it slightly downward to 66% because no Kalshi-direct price or official agi.safe.ai model card was captured. Gemini's historical 4-7 point HLE gains suggest that improving from the reported 44.4% no-tools score to at least 50% is plausible, while competitors scoring above 54% create pressure for Google to deliver a substantial upgrade. Conversely, the 51.4% tools-enabled result may not match the configuration used by agi.safe.ai, and independent evaluation could fall below vendor-reported performance. Repeated Gemini 3.5 Pro delays add meaningful deadline and qualification risk, although related markets still strongly expect a Pro release before year-end.

gpt-5.6-sol

61% conf

Yes	64%
No	36%

The 67.5% YES Polymarket price is the best available consensus anchor, though its decline indicates growing concern about delays and benchmark ambiguity. Gemini’s historical 4–7 point gains suggest the next Pro could rise from 3.1 Pro’s reported 44.4% no-tools score to roughly the 48–51% range, placing the threshold near the center of plausible outcomes. Competitors scoring above 54% create strong pressure for Google to withhold the repeatedly delayed model until it is competitive, while expected October–November release timing leaves some buffer before year-end. I move modestly below the market because agi.safe.ai’s configuration is unconfirmed, independent scores may trail vendor figures, and further launch or leaderboard-update delays remain material.

6. Final plain mean of 5 draws

64% (51% confidence)

Yes: 64% No: 36%

Mean of 5 draws (0.63, 0.66, 0.63, 0.66, 0.64). The best market anchor is Polymarket's identical-question price at 67.5% YES (softening from 91%), with a coherent sibling ladder (45%: 80%, 55%: 33%) placing the modal expected score in the high-40s to mid-50s. Gemini 3.1 Pro reportedly sits at 44.4% no-tools / 51.4% with tools, and historical per-generation gains of ~4-7 points plus intense competitive pressure (Claude/GPT rivals at 54-65%) make a ≥50% debut for a redesigned 3.5 Pro/next Pro plausible. Offsetting risks: the agi.safe.ai chart historically displays no-tools figures, independent CAIS evals often land at or below vendor claims, and Google's three missed launch targets leave a non-trivial chance of no qualifying Pro model appearing on the chart before the Dec 31 close (sibling markets imply ~5-10% no-release risk even by late Nov). I shade slightly below the market given the declining trend and configuration ambiguity.

Pipeline Timing

Total pipeline time: 167.8s

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