Will Gemini Argon debut with a Humanity's Last Exam score of 45% or higher?

- [2026-09-30] Google announced Gemini 4 Argon. It is "rolling out soon" and was first released only to trusted cyber defenders through the Fairwind Program. Google said it will reach developers, enterprises and consumers "as soon as possible," with no date given. The post publishes no HLE score (its table covers DeepSWE, Vals Index, AutomationBench and similar). https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ - [2026-10-13] Per Tech Insider's summary of Ars Technica, 9to5Google and Engadget reports, Argon began rolling out in the Gemini app (AI Pro and Ultra) and in Antigravity (Ultra only). The API intro price is $2/$10 per million tokens until Nov 30, 2026. Broader developer/enterprise availability is still undated. This is secondary reporting; I did not open the primary articles. https://tech-insider.org/gemini-4-argon-rollout-paywall-2026/ - [2026-10-07] llm-stats said Argon had no public API model code and was not listed on the Gemini API models or pricing pages as of Oct 7. This predates the Oct 13 app rollout. https://llm-stats.com/blog/research/gemini-4-argon-launch - [2026-10-14] Artificial Analysis lists Gemini 4 Argon (High) at 57% on HLE, on its text-only subset of about 2,158 questions. It lists Claude Opus 5.5 at 61% (Max) and 56% (High). This is a different harness from agi.safe.ai/Scale, so it is not the resolution figure. https://artificialanalysis.ai/models/comparisons/gemini-4-argon-vs-claude-opus-5-5-high - [2026-10-06] On Scale's full-set board, GPT-6 Astra leads at 54.80%. On HLE-Diamond it scores 59.9% without tools and 82.9% with tools. Artificial Analysis has Claude Opus 5.5 at 61.4%. Scale-style full-set scores run lower than Artificial Analysis text-only scores. https://iqmetrics.org/news/humanitys-last-exam-scores-what-the-leaderboards-show/ - [2026-09-30] BenchLM mirrored Scale Labs' public HLE board. It showed GPT-6 Astra at 60.6%, Claude Opus 5.5 at 55% and Claude Fable 5.1 at 51.3%. Gemini 3.8 Flash was the best Google entry at 34.3%, and Claude Opus 5 was at 38.6%. This is a different snapshot from the 54.80% figure above, and BenchLM's figures conflict with the Oct 6 reading. Argon was not yet on the board. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-10-14] The agi.safe.ai page was fetched. Its news ticker shows HLE-Diamond released on 09/22/2026 and HLE published in Nature on 01/28/2026. The results chart loads dynamically ("Loading HLE results...") and I could not read it. The static table in the page text is stale (Gemini 3 Pro 38.3% is the top row) and has no Argon entry. I could not confirm whether Argon is on the chart yet. https://agi.safe.ai/ - [2026-10-09] The Gemini 4 Argon model card (DeepMind) lists "results as of October, 2026," but the visible snippet shows only Agent's Last Exam (39.5%), which is a different benchmark from HLE. https://deepmind.google/models/model-cards/gemini-4-argon/ - [2026-10-12] Other trackers show Anthropic models leading HLE at about 65% (Claude Fable 5.1), using various harnesses. These are not agi.safe.ai figures. https://benchlm.ai/benchmarks/hle - [2026-10-14] The Polymarket page for this event lists strikes at 45%+, 50%+, 55%+ and 60%+. Its description states Argon had 57.1% on HLE "in recent leaderboards." That figure matches Artificial Analysis, not Scale/CAIS. https://polymarket.com/event/gemini-argon-humanitys-last-exam-debut-20261002 - [2026-10-04] Bloomberg reported (via Axios, cited by NeuralTrust) that some Google staff found Argon's performance lacking in internal testing. Google disputed this. This is a report, not confirmed. https://neuraltrust.ai/blog/gemini-4-argon BOTTOM LINE: Argon has been announced and is partly available (Gemini app, Antigravity, Fairwind), so Scale/CAIS can evaluate it. As of Oct 14 I could not confirm it is on the agi.safe.ai chart. Third-party HLE scores for Argon are about 57% (text-only, Artificial Analysis), but the Scale full-set scores that agi.safe.ai probably uses run several points lower, with the leader at 54.8% to 60.6% depending on snapshot. So a 45% threshold is plausible, but the harness gap and the unknown timing of its appearance on the chart are the main uncertainties. NOT FOUND: the rendered agi.safe.ai chart data (it loads dynamically and I could not read it); any Scale/CAIS-run HLE score for Gemini 4 Argon; the date Argon was or will be added to agi.safe.ai; the full Argon model-card HLE row; a date for broad API availability.

- [2026-10-14] The agi.safe.ai page, as fetched, shows a static table with only Gemini 3 Pro (38.3% accuracy, 57.2 calibration error). It also shows a "Loading HLE results…" placeholder, so the live chart is JavaScript-rendered and I could not read its contents. The latest-news list ends with HLE-Diamond released on 2026-09-22. No Argon entry was visible. https://agi.safe.ai/ - [2026-09-30] Google announced Gemini 4 Argon, available only to trusted cyber defenders via the Fairwind Program. Broad availability is "as soon as possible" after safeguards work and the US government's voluntary pre-release access process. The blog's named benchmark claims are DeepSWE, Vals Index, AutomationBench and LVBench. The text I fetched does not state an HLE score. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ - [2026-10-03 to 10-12] Artificial Analysis (independent, text-only subset of 2,158 questions) gives Argon (High) an HLE of 57%, reported as 57.1% by secondary sites. That compares with Claude Opus 5.5 (Max) at 61.4%, Opus 5.5 (High) at 56%, and GPT-6 Astra (Low) at 49%. This is not the agi.safe.ai methodology. https://artificialanalysis.ai/models/comparisons/gemini-4-argon-vs-claude-opus-5-5 - [2026-09-30 snapshot] Scale Labs' full-set leaderboard (the likely source for agi.safe.ai) had GPT-6 Astra at 60.6%, Claude Opus 5.5 at 55%, Claude Fable 5.1 at 51.3%, Claude Opus 5 at 38.6% and Gemini 3.8 Flash at 34.3%. Argon was not listed. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-10-06] A separate report put Scale's board at GPT-6 Astra 54.80 (±1.94). Artificial Analysis scores were 61.4% for Opus 5.5, versus 55% for the same model on Scale's snapshot (above). The report says one model on Scale's board and at least five on Artificial Analysis are above 50%. So Scale's full-set numbers tend to be about 0–7 points below Artificial Analysis for the same model, though the comparison is thin and the two Astra readings (54.8 vs 60.6) are inconsistent. https://iqmetrics.org/news/humanitys-last-exam-scores-what-the-leaderboards-show/ - [2026-10-06] There is no single HLE leaderboard in 2026. Anthropic, Artificial Analysis, Scale (full set) and CAIS/Scale (HLE-Diamond, a cleaned 1,000-question subset) each publish their own. This creates ambiguity about which number the agi.safe.ai card will show. https://klu.ai/benchmarks/humanitys-last-exam - [Historical, Nov 2025–Feb 2026] Gemini 3 Pro scored 37.5% self-reported on the full set and 38.3% on agi.safe.ai, a small gap in that case. It was the sole Google entry on the agi.safe.ai table. Gemini 3 Deep Think later reported 48.4%. https://blog.google/products-and-platforms/products/gemini/gemini-3/ - [2026-08-16] A secondary source said no model on the official public leaderboard had reached 50%. This was before the Astra, Opus 5.5 and Fable results above appeared on Scale's board. https://intuitionlabs.ai/articles/humanitys-last-exam-ai-benchmark - [Historical] The site table cited by news in Jan–Feb 2026 and the Feb 2026 snapshot had few entries. Gemini 3 Pro stayed the displayed Google model, and Gemini 3.1 Pro and Gemini 3.8 Flash (34.3% on Scale's board) were not on the agi.safe.ai table I fetched. That table has not been updated since the early-2026 coverage cited above. Whether Scale's leaderboard entries propagate to the agi.safe.ai chart is unconfirmed. https://agi.safe.ai/ - [2026-10-14] Public availability is the main timing constraint. Argon is Fairwind-only with no public API date, so third-party evaluators with API access (Artificial Analysis) already have it, but Scale and CAIS may not. Gemini 3.8 Flash was evaluated by Scale after release. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-10-05 approx] A Polymarket page for this event shows other thresholds. Its snippet put 45%+ at 97% and 50%+ at 92%. These are stale snapshots; the live prices are given to forecasters separately. https://polymarket.com/event/gemini-argon-humanitys-last-exam-debut-20261002 BOTTOM LINE: If Argon appears on agi.safe.ai, the outside view is that a score of 45% or more is likely, since Artificial Analysis shows about 57% and Scale's board scores comparable frontier models at 51–61%. The larger risk is that no Argon entry is added before 2027-06-30, because the site's displayed table has lagged badly (only Gemini 3 Pro was shown) and Argon is not publicly released. A secondary risk is that Scale's full-set score runs materially below Artificial Analysis's. NOT FOUND: the live agi.safe.ai chart contents (JavaScript-rendered); any Argon HLE score from Scale or CAIS; any history of how many days new models took to appear on agi.safe.ai; a Google-reported Argon HLE figure; a public release date for Argon.

YES means: Argon (or a qualifying Argon-class or public-release model) is newly added to the agi.safe.ai HLE chart before 2027-06-30, and its displayed "HLE Accuracy" is 45% or higher at 12:00 PM ET on the next calendar day. - [2026-10-13] Argon is no longer limited to Fairwind. Google's developer post says developers can build with it "starting today". It is on the paid Gemini API tier, in AI Studio and in Android Studio, and Ultra subscribers get AI Studio and Antigravity. A third-party hands-on test says the GA model ID `gemini-4-argon` (no preview suffix) was live on launch day. This removes the biggest YES prerequisite, because Scale and CAIS can now run the model through the public API. It also supersedes the Oct 7 "no API code" note. https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/ - [2026-10-13] The DeepMind model card was published Oct 13 and is dated October 2026. It calls Argon a frontier model that is not a fine-tune of a prior model. It gives Argon a high-effort headline for Agent's Last Exam (39.5%) but shows no HLE row. Google's own table also lacks HLE, so no self-reported number sits above or below the 45% line. Third-party scores are the only evidence. https://computingforgeeks.com/gemini-4-argon-released-features-benchmarks/ - [2026-10-14] Artificial Analysis puts Argon (High) at 57% on HLE. Its main page says it was measured only at high effort, "the only configuration Google has made available". That is 12 points above the 45% threshold. Argon scores 20 points higher than Gemini 3.8 Flash (Low), which has 37% on the same harness. On Artificial Analysis, Argon is ahead of GPT-6 Astra (Low) at 49% and roughly level with Opus 5.5 (High) at 56%. https://artificialanalysis.ai/models/comparisons/gemini-4-argon-vs-gemini-3-8-flash-low - [2026-10-14] Artificial Analysis lists a second, multi-effort comparison. Argon (High) at 57% sits above Opus 5.5 (High) at 56%, and below Opus 5.5 (Max) at 61%. Even if Scale's harness takes several points off, as on the Oct 6 and Sept 30 snapshots, a 12-point cushion over 45% is large. The Sept 30 Scale-board scores for Opus 5.5 (55%) and Fable 5.1 (51.3%) are the same peer class as Argon. https://artificialanalysis.ai/models/comparisons/gemini-4-argon-vs-claude-opus-5-5-high - [2026-10-13] Google's table puts Argon above Opus 5.5 on 13 of 19 rows, on knowledge-work and long-context benchmarks. Per DeepLearning.AI's Oct 10 coverage, it has the lowest hallucination rate of any model scoring 45 or higher on the Artificial Analysis Intelligence Index (15%). HLE penalises confident wrong answers through its calibration error. That does not affect the displayed accuracy, but it points to a knowledge-heavy model. https://www.deeplearning.ai/the-batch/argon-a-low-hallucination-model-build-for-knowledge-work - [2026-10-12] A benchmark aggregator ranks Argon's HLE at 57.1%, 98th percentile, rank 11 of 478 tracked models. Argon at 45% would still sit in the top few percent of the whole field. Only a harness failure, such as a heavy refusal rate or truncated reasoning, could drop it below that. Its default thinking level is medium, not high. A Scale run at default settings could therefore score lower, but 45% would still be about 12 points under the AA high-effort score. https://benchmarklist.com/models/google-gemini-4-argon/ - [2026-10-13] Argon has a large output budget: 1M window, 65K per request, plus a "long decode" mode. This is less likely to cut off long reasoning than earlier Gemini models. A cut-off would otherwise lower HLE accuracy. https://computingforgeeks.com/gemini-4-argon-released-features-benchmarks/ - [2026-09-22] CAIS/Scale released HLE-Diamond, and the agi.safe.ai news ticker shows the HLE site is actively maintained. This is the best sign the chart is being updated now. The earlier static table was stale, but the live chart is what matters. Scale and CAIS evaluate new frontier releases when API access exists, as the Gemini 3.8 Flash entry on Scale's board shows. https://agi.safe.ai/ - [2026-10-02] On Arena, Argon is #1 in Text (1525) and leads Hard Prompts, which is the kind of broad capability lead that usually carries over to knowledge-heavy exams like HLE. These are preliminary votes (4,932 overall), so treat this as weak support. https://dev.to/eristoddle/gemini-4-argon-won-every-arena-category-you-cant-use-it-4fc9 - [Path to YES] Argon is now public on the API (Oct 13). Scale and CAIS run it and add a card to the agi.safe.ai chart, within weeks to a few months and well before June 30, 2027. The displayed score, on the full set, is probably in the low-to-mid 50s, from AA's 57% minus the usual few-point harness gap. That clears 45% with margin. The main residual risk is no chart addition, not a low score. BOTTOM LINE: Argon's public API launch on Oct 13 removes the "not publicly accessible" obstacle. The 57% AA score and the 51–61% scores of peer models on Scale's board give a 6–12 point buffer over 45%, so YES mostly depends on Argon being added to the chart at all, not on its score. NOT FOUND: The live agi.safe.ai chart contents (JavaScript-rendered). Any Scale or CAIS HLE score for Argon. A date for Argon's addition to the chart. A Google-reported HLE figure for Argon. How quickly past Google models were added to agi.safe.ai after release.

YES means: the next Gemini Argon model added to the agi.safe.ai HLE results shows an "HLE Accuracy" of 45% or higher at 12:00 PM ET on the day after it first appears. If no Argon model is added by 2027-06-30, the market resolves No. - [2026-10-14] I fetched agi.safe.ai (plain and rendered). Both returned only the news ticker, the paper links, the "Submission Dashboard" link and contributor lists. The latest item is HLE-Diamond on 09/22/2026. Neither fetch showed any Argon entry or any chart data. The page's own metadata gives lastexam.ai as the canonical URL. So the resolution surface is a JavaScript chart that I could not read, and no visible evidence puts Argon on it. Status quo is "not listed". https://agi.safe.ai/ - [2026-10-14] The Argon availability gap that earlier reports flagged as a reason for a delayed listing has closed, and this removes part of the NO case. Sundar Pichai said on Oct 13 that Argon is available globally through the paid Gemini API, AI Studio, Antigravity (Ultra) and the Gemini app and Search (Pro and Ultra). Digital Trends reported this, and the stable model ID is `gemini-4-argon`. So "Scale and CAIS cannot access it" no longer works as an excuse. NO now depends on listing lag, not access. https://www.digitaltrends.com/computing/gemini-4-argon-is-finally-available-for-developers-businesses-and-paid-users/ and https://ai.google.dev/gemini-api/docs/models/gemini-4-argon - [2026-10-14] The "57.1%" figure in the market description and Wikipedia's table comes from Artificial Analysis, as the BenchLM AA-HLE page (Oct 7) shows. It is not a CAIS or Scale measurement. Wikipedia lists GPT-6 Astra at 54.7, a different number from both Scale snapshots in the earlier reports. Nothing in the search results shows an Argon figure measured under the agi.safe.ai methodology. https://benchlm.ai/benchmarks/aahle and https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam - [2026-10-13] Google's Argon comparison table has no HLE row. The table covers 19 rows, and a Oct 13 review said it has no SWE-bench Verified, GPQA or HLE score. Google chose not to headline HLE for Argon, so no vendor-reported number backs the Scale or CAIS score. A weak Scale or CAIS run would be a surprise to nobody at Google. Argon trails Opus 5.5 on several rows, including Terminal-Bench. https://computingforgeeks.com/gemini-4-argon-released-features-benchmarks/ - [2026-10-13] Artificial Analysis lists Argon at only one configuration (High). Google exposes low, medium and high thinking in the API. Scale or CAIS may run a default or lower-effort setting, or run without tools or with a different judge, and those choices can lower the score. Scale's 55% for Opus 5.5 against 61.4% at Artificial Analysis shows a gap of about 6 points. Applying a similar gap to Argon's 57.1 gives roughly 51. That is still above 45, so a score miss is a tail risk, not the main NO route. https://www.orcarouter.ai/blog/gemini-4-argon and https://benchlm.ai/benchmarks/aahle - [2026-10-09] The newest Scale entries on the public board that the earlier reports could see are GPT-6 Astra, Opus 5.5 and Fable 5.1, plus Google's Gemini 3.8 Flash at 34.3. Gemini 3.1 Pro (44.4% self-reported, Feb 2026) was not on the agi.safe.ai static table I saw. The only Google entry the site has ever shown is Gemini 3 Pro at 38.3%, the number news outlets quoted in Jan–Feb 2026. That is precedent for a long, unexplained lag between a Google model shipping and appearing on this specific chart. The Google lineage on the site is Gemini 3 Pro only, with no Gemini 3.1 Pro, 3.5 or 3.7 visible. Pricepertoken's Oct 9 HLE board still lists Gemini 3 Flash at 37.5 as an example row. https://pricepertoken.com/leaderboards/benchmark/hle and https://www.biggo.com/ (BigGo, Gemini 3 Pro 38.3%) at https://finance.biggo.com/news/g6nJB5wBpPlmmn5lqhfT - [2026-10-14] The market's strict "first appears" mechanic favours NO. The market needs a new Argon card on the chart with its date of first appearance. The chart is only checked at 12:00 PM ET the next day, so the card must persist. If Scale shows Argon only on the SEAL leaderboard, or on HLE-Diamond or "HLE-Rolling" but not the main agi.safe.ai chart, it does not count. The site now has HLE-Diamond (Sept 22) and a rolling fork (Oct 2025) beside the original 2,500-question set. That raises the chance the displayed chart is rebuilt, repointed or frozen. The resolver would also need to read a dynamic chart, which adds ambiguity. https://agi.safe.ai/ - [2026-10-14] The site's own news ticker has had no model-update posts for a year: 04/03/2025, 10/08/2025, 01/28/2026 and 09/22/2026 are all dataset or paper events. Nothing says the chart is updated on a schedule. The Quantitative Results section points readers to the CAIS Dashboard and SEAL leaderboards. That suggests the agi.safe.ai chart is a curated snapshot that may lag, or may never add Argon, while SEAL gets the newer results. I found no source that states an update cadence. https://agi.safe.ai/ - [2026-10-12] Critics and earlier reports describe Argon's strongest results as breadth results (13 of 19 rows). Its HLE score is third on Artificial Analysis, behind Opus 5.5 at 61.4 and Fable 5.1 at 59.1. A "not a leader" model is less likely to be fast-tracked by third parties. Argon also follows a rocky pre-release story, with reports of internal doubts (Bloomberg, Oct 4, disputed by Google). https://benchlm.ai/benchmarks/aahle BOTTOM LINE: The best NO route is not a low score, because 45% is well below the roughly 51–57% implied by third-party figures. The NO route is that Argon never appears on this chart as a new card by 2027-06-30, or appears in a form the criteria exclude, because the chart looks stale (only Gemini 3 Pro visible) and updates have no stated cadence. The access barrier is gone as of Oct 13, so the timing risk now rests on Scale/CAIS behaviour alone. NOT FOUND: the rendered chart data on agi.safe.ai (both fetches showed no chart); any Scale or CAIS HLE score for Argon; any record of how long Gemini 3.1 Pro or Gemini 3.8 Flash took to reach the agi.safe.ai chart, or whether they ever did; any stated update schedule for the chart; a Google-reported Argon HLE figure.