Will Gemini Argon debut with a Humanity's Last Exam score of 60% or higher?

- [2026-09-30] Google announced Gemini 4 Argon. It is rolling out only to trusted cyber defenders through the Fairwind Program. Google says it is in the US government's voluntary pre-release access process and will make Argon available to developers, enterprises and consumers "as soon as possible". The blog post gives no HLE score in its text, and I could not read the benchmark chart image. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ - [2026-10-05] There is still no public release date. Paid API customers and Google AI Ultra subscribers are next in line, with no date given. https://aiidelist.com/blog/gemini-4-argon-release-date - [2026-10-03] Secondary source: only Fairwind defenders could use Argon. Fairwind launched 2026-09-02. https://aiwiz.uk/blog/gemini-4-argon/ - [2026-10-07] Artificial Analysis (AA) measures Argon (High) at 57.1% on HLE, using 2,158 text-only questions. That puts it third on AA's HLE board, behind Claude Opus 5.5 (61.4%) and Claude Fable 5.1 (59.1%). Argon's AA Intelligence Index is 53, versus 58 for Opus 5.5. The AA page also says "no publicly available API", so AA's access route is unclear. https://benchlm.ai/benchmarks/aahle and https://artificialanalysis.ai/models/gemini-4-argon - [2026-10-12] The AA board is a different benchmark setup from agi.safe.ai. AA is third-party and text-only. Scale's board uses the full set and different judging. Anthropic reports full-set with-tools scores: Opus 5.5 at 67.7% (2026-09-22). https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-06] Scale's full-set HLE leaderboard had GPT-6 Astra first at 54.8%, with Fable 5.1 (xhigh) at 46.5% and gemini-3.1-pro-preview at 46.4%. Argon is not listed on it. https://iqmetrics.org/news/humanitys-last-exam-scores-what-the-leaderboards-show/ - [2026-09-22] The HLE site announced HLE-Diamond, a 1,000-question subset. Without tools, GPT-6 Astra scores 59.9% and Opus 5.5 scores 54.6%. With tools, GPT-6 Astra scores 82.9%. Argon is not in the figures I found. Which configuration the chart on agi.safe.ai displays is unclear. https://iqmetrics.org/news/humanitys-last-exam-scores-what-the-leaderboards-show/ - [2026-10-12] The agi.safe.ai page I fetched has its results table loaded by JavaScript. The static fallback table is stale, dated April 2025 with Gemini 3 Pro at 38.3% on top. I could not read the live "AI Progress on HLE" chart, and I could not confirm any Argon entry. The latest news line on the page is 09/22/2026 (HLE-Diamond). https://agi.safe.ai/ - [2026-10-02] Polymarket's description cites 57.1% as Argon's HLE score, calling it "recent leaderboards". That matches the AA figure, not an agi.safe.ai figure. The page's strike ladder includes 45%, 50%, 55% and 60%+ for this event. I did not report prices. https://polymarket.com/event/gemini-argon-humanitys-last-exam-debut-20261002 - [2026-10-03] Axios reported, per a secondary source, that Bloomberg said some Google staff found Argon's performance lacking in internal testing. Google disputed this. https://neuraltrust.ai/blog/gemini-4-argon - [2026-10-12] No scheduled public-release event is announced. A first appearance on agi.safe.ai probably depends on the Argon broader rollout, or on CAIS/Scale evaluating it through Fairwind or other access. Timing and access for that are unknown. BOTTOM LINE: As of 2026-10-12 Argon has no entry on agi.safe.ai that I could confirm, and there is no public release date. The only independent HLE figure is AA's 57.1%, below 60% and below Opus 5.5's 61.4% on the same board. The site's own scoring setup (full set vs HLE-Diamond, with or without tools) could produce a different number, so the displayed figure is uncertain. NOT FOUND: Argon's HLE score in Google's own benchmark table; a live view of the agi.safe.ai results chart; any CAIS/Scale HLE score for Argon; a public release or general-availability date.

- [2026-09-30] Google announced Gemini 4 Argon. Access is limited to trusted cyber defenders through the Fairwind Program. Google says it is gathering feedback and working through the U.S. government's voluntary pre-release access process "before making Argon available" more widely, with no date given. The launch blog's visible text does not quote an HLE score; its benchmark images could not be read. Timing risk: the HLE chart entry may come late or never. The market's cutoff for a qualifying addition is 2027-06-30. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ - [2026-10-12] The agi.safe.ai page, as served, loads its results dynamically. The static table is stale: Gemini 3 Pro 38.3%, GPT-5 25.3%, with a "Dataset Updated: April 3rd, 2025" label. The news ticker lists "HLE-Diamond released" on 09/22/2026. There is no Argon entry in what I could read. I could not read the live chart. https://agi.safe.ai/ - [2026-10-12] Artificial Analysis (AA), a third-party evaluator, reports Argon (High) at 57% on HLE, versus 61% for Claude Opus 5.5 and 37% for Gemini 3.8 Flash (Low). AA is not the resolution source. https://artificialanalysis.ai/models/comparisons/gemini-4-argon-vs-claude-opus-5-5 - [2026-10-01 to 2026-10-12] A secondary summary lists AA's HLE figures as Argon 57.1%, Opus 5.5 61.4% and GPT-6 Astra 54.7%. BenchLeader shows Argon at 57.1%, ranked #5 on that measure. These are mirrors of AA's number. https://neuraltrust.ai/blog/gemini-4-argon and https://www.benchleader.com/models/gemini-4-argon - [2026-09-30] Scale Labs' public HLE leaderboard, which is likely closer to the agi.safe.ai source, ran lower than AA for Gemini, according to a BenchLM mirror of the Scale snapshot. Its 2026-09-23 snapshot had GPT-6 Astra 54.8%, Fable 5.1 (xhigh) 46.5%, Gemini 3.1 Pro Preview 46.4%, Gemini 3.8 Flash 44.5%. The 2026-09-30 snapshot had GPT-6 Astra 60.6%, Opus 5.5 55%, Fable 5.1 51.3%, Gemini 3.8 Flash 34.3%. Scale's scores for the same model moved by about 10 points between snapshots, so the evaluation set or settings changed (HLE-Diamond was released 09/22). Argon has no Scale entry that I found. https://benchlm.ai/benchmarks/scale-hle-diamond and https://benchlm.ai/benchmarks/scale-humanitys-last-exam - [2026-10-12] HLE scores vary a lot by source. Claude Fable 5.1 is 59.1% on pricepertoken (AA-sourced, 2026-10-09). It is 65.0% on llm-stats and BenchLM, which appear to use vendor-reported or with-tools figures. The only 60%+ scores seen for any model are Scale's 60.6% for GPT-6 Astra and the 60–65% vendor-reported figures for Claude models and DeepSeek-V4-Pro (60.0%). https://pricepertoken.com/leaderboards/benchmark/hle and https://llm-stats.com/benchmarks/humanity's-last-exam - [2026-10-12] Reference class: Gemini models usually appear on the agi.safe.ai/Scale list. On the 2026-09-23 Scale snapshot, Gemini 3 Pro was 37.5% and Gemini 3.1 Pro Preview was 46.4%. The list was 53 models deep, with many older entries. Entry lag is uncertain and Google's HLE figures are often not the same as the board's. https://benchlm.ai/benchmarks/scale-humanitys-last-exam - [2026-10-12] Sibling strikes on the same Polymarket event (view-only page, Oct 12): 45%+ was 78%, 50%+ was 74%, 55%+ was 69%. The market's own price for this strike is excluded as instructed. The 45% to 55% prices appear to assume a debut in the mid-to-high 50s. An earlier search snippet (about Oct 5) showed 97% and 92% for the 45% and 50% strikes, so the 45–55% prices fell sharply in a week. The Polymarket page's AI-generated summary cites "around 57%", which is AA's number and not the site's. https://polymarket.com/event/gemini-argon-humanitys-last-exam-debut-20261002 - [2026-10-12] Argon is not generally available. Reports say there is no model ID, endpoint or GA date, and only one provider hosts it. The Fairwind cohort reportedly received it without cyber safeguards. If the chart entry waits for public API access, entry could be delayed. If it never appears by 2027-06-30, the market resolves No. https://orcarouter.ai/blog/gemini-4-argon - [2026-10-12] On the AA Intelligence Index Argon scores 53, against 58 for Claude Opus 5.5 and the same as GPT-6 Astra and Fable 5.1. That puts Argon at the frontier but not ahead of the best HLE performers. https://www.rdworldonline.com/googles-overdue-gemini-4-argon-reaches-the-frontier-with-mixed-results-for-science/ - [2026-10-12] Historically, Gemini 3 Pro led HLE at launch (38.3% on the site's stale table). Argon's AA score (57%) is below the best AA-measured score (Fable 5.1 at 59.1%). The agi.safe.ai chart's judge and settings differ from AA, so AA's 57% is only a loose proxy for the displayed figure. https://agi.safe.ai/ BOTTOM LINE: The only measured Argon HLE number is AA's 57%, below 60%. The resolution site uses a different Scale/CAIS methodology that has produced both lower (Gemini 3.8 Flash 44.5% vs AA 37%, a different direction) and much higher (GPT-6 Astra 60.6%) readings than AA. The outside view is therefore a wide spread around the high 50s, plus a real chance that Argon never appears on the chart before 2027-06-30. NOT FOUND: the live agi.safe.ai chart contents, any Argon entry on it, Google's own HLE figure for Argon (the blog's benchmark images could not be read), the historical lag between a Gemini release and its appearance on the chart, and any expert forecast of Argon's HLE score other than Polymarket's.

YES means: the next Argon-class Gemini model added to the agi.safe.ai "AI Progress on HLE" chart shows an HLE Accuracy of 60% or higher at 12:00 PM ET on the day after it first appears, and it appears by 2027-06-30. - [2026-10-12] A fresh render of agi.safe.ai returned only the contributor and affiliation text. The results chart was not in the output, so I could not confirm any Argon entry. Newest news item is still the 09/22 HLE-Diamond release. I only checked this one page. https://agi.safe.ai/ - [2026-10-12] Wikipedia's HLE results table, sourced to Artificial Analysis (AA) and dated 22 Sept, lists Argon at 57.1%. That is second on the table, behind Opus 5.5 at 61.4%. It is a text-only subset and not the resolution source. The 57.1 figure is a floor-style anchor: a methodology that scores Argon about 3 points higher would clear 60. https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam - [2026-10-12] YES path, step 1: the chart is run by CAIS and Scale, and Scale's own methodology scores models differently from AA. The 2026-09-30 Scale snapshot has GPT-6 Astra at 60.6%, against 54.7% on AA, so Scale's setup lifted that model by about 6 points. Argon at 57.1 on AA plus a similar uplift would land near 60–63. This is not a consistent pattern. Opus 5.5 scored 55% on Scale against 61.4% on AA. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-10-12] YES path, step 2: the chart may display a with-tools or best-configuration score. Vendor-reported with-tools HLE runs are well above 60% (Fable 5.1 at 65.0%, Opus 5 at 64.7%, DeepSeek-V4-Pro at 60.0%). GPT-6 Astra reached 82.9% with tools on HLE-Diamond (CAIS post, 2026-09-22). If the chart takes a tools-enabled run, Argon clears 60 easily. Which configuration the chart shows is unconfirmed. https://llm-stats.com/benchmarks/humanity's-last-exam and https://www.octagonai.co/news/ai-benchmark-humanity-s-last-exam-market-odds/ - [2026-10-12] Argon is a strong model on other measures. It leads or ties on 13 of 18 disclosed benchmarks in Google's table. It ranks #1 on arena.ai with 1525, against 1505 for Fable 5. A high-effort run by Google or Scale, using a harness that lets it search or run code, could plausibly reach 60+. https://tech.slashdot.org/story/26/09/30/2244216/google-unveils-gemini-4-argon-retaking-benchmark-lead-over-openai-and-anthropic - [2026-10-12] Google ran Argon at its highest thinking setting for its own table. The AA 57.1 figure is for "High" effort, and BenchLeader says Argon has been measured at 2 effort settings. A max or xhigh setting is not shown in the AA data I found. Anthropic's top HLE entries on AA are Max settings, so headroom may exist. https://www.eesel.ai/blog/gemini-4-argon and https://www.benchleader.com/models/gemini-4-argon - [2026-10-12] The resolution rule takes the highest-scoring qualifying model if several appear on the same date. A Pro, Deep Think or other Argon-class variant, or the public release counterpart, would count. Only the best variant added that day matters. https://llm-stats.com/blog/research/gemini-4-argon-launch - [2026-10-12] YES path, step 3: timing. Google says it will expand access "as soon as possible", starting with paid API customers and Ultra subscribers. The Argon price is already published at $2/$10 per 1M tokens, which suggests the commercial release is being planned. The window runs to 2027-06-30, about 8.5 months. A public API entry would let Scale and CAIS run it. https://www.therundown.ai/tools/gemini-4-argon and https://www.online-tech-tips.com/google-gemini-4-argon-announcement/ - [2026-10-12] Third-party HLE figures for the leading models span a wide range. Fable 5.1 is 59.1 on AA and 65.0 on vendor-reported boards. The same model can differ by 6 points by method. For Argon, a vendor-style or tools-on figure is the likeliest route above 60. https://pricepertoken.com/leaderboards/benchmark/hle and https://benchlm.ai/benchmarks/hle BOTTOM LINE: The most likely YES path is that Argon is publicly released and added to the chart, and the chart's methodology (Scale/CAIS, possibly tools-enabled or at max effort) reads at least about 3 points above AA's 57.1. The weak links are that the chart configuration is unconfirmed, Argon is not on the 2026-09-30 Scale snapshot, and there is no release date. NOT FOUND: the live agi.safe.ai chart contents or any Argon entry; which configuration the chart displays; Google's own HLE number for Argon (the benchmark images in its post were unreadable); any Scale/CAIS score for Argon; a public release date.

YES means: the next Gemini Argon model added to the agi.safe.ai HLE chart shows an "HLE Accuracy" of 60% or higher at 12:00 PM ET on the day after it first appears. - [2026-10-12] Wikipedia's HLE results table, sourced to Artificial Analysis (AA) on the text-only subset and dated 22 Sept 2026, lists Argon at 57.1%. That is behind Opus 5.5 at 61.4%. The table is AA data, not agi.safe.ai data, and is another mirror of the same AA number. https://en.wikipedia.org/wiki/Humanity%27s_Last_Exam - [2026-10-12] The Scale Labs HLE leaderboard is the likely feeder for the agi.safe.ai chart. The rendered page I fetched is built on the 2,500-question set. It carries a "September 17, 2026" HLE-Rolling note and reports score ± CI with calibration error. The visible entries are old: the Gemini entries shown top out at gemini-2.5-pro, and gpt-5.2 sits at 27.8%. Scale adds models slowly and does not follow launch dates. The visible rows had no Argon, and the top rows were truncated in my fetch. https://labs.scale.com/leaderboard/humanitys_last_exam - [2026-10-12] A Scale-style, no-tools full-set number would very likely fall below AA's. My colleagues' dated snapshots show Gemini 3.8 Flash at 34–44% on Scale. The Scale HLE-Diamond snapshots had GPT-6 Astra at 54.8% on 09-23 and 60.6% on 09-30, so the same model's score moved about 6 points in a week. That is a high-variance measurement, but the best frontier non-tool number on the site's own methodology is only just around 60%. Argon is rated below the leaders on AA. It would need to beat the best Scale-measured model to reach 60%. (Source: the fetched Scale page plus colleague snapshots.) https://labs.scale.com/leaderboard/humanitys_last_exam - [2026-10-11] Argon is still gated, with no API model ID. A felloai.com update dated 2026-10-11 says only vetted Fairwind defenders can use it and Google has given no date or API model ID. Scale and CAIS evaluate via public APIs, so a chart entry before general availability looks unlikely. Kavukcuoglu said before launch that he hoped for a release "much earlier" than the end of 2026. That is a hope, not a date. https://felloai.com/all-we-know-about-google-gemini-4/ - [2026-10-07] LLM Stats confirms no public API as of 10-07. Argon is not on the Gemini API models or pricing pages. Paid API and AI Ultra come next with "no date". Argon also goes through the US government's voluntary pre-release access process. https://llm-stats.com/blog/research/gemini-4-argon-launch - [2026-10-01] An EvoLink check found no Argon entry or model ID in the public Gemini API catalog or changelog. This is a mechanism for delay. Any independent chart entry needs callable access, and Fairwind participants must limit use to cybersecurity work, which makes outside HLE evaluation harder. https://evolink.ai/blog/gemini-4-argon-release-date - [2026-10-01] Fairwind eligibility is restricted to cybersecurity, incident-response and penetration-testing employees, with MFA required, and there is no self-service path. This blocks CAIS or Scale from running HLE through the Fairwind route. https://shattered.io/gemini-4-argon-650-partners-fairwind-2026/ - [2026-09-30] Google's own comparison table has 19 rows, and HLE is not among the scores listed by the aggregators. The listed rows include DeepSWE, Vals Index, AutomationBench and LVBench. Google did not choose HLE as a headline, which suggests the vendor figure is not a showpiece. The lack of a vendor HLE figure means there is no reason to expect a 60%+ vendor-reported number to be carried onto the chart. https://toolscout.ai/news/gemini-4-argon-benchmarks - [2026-10-12] AA ranks Argon 53 on its Intelligence Index, level with GPT-6 Astra and 1 point ahead of GPT-6.1 Sol. AA says gains come from lower hallucinations and agentic capabilities, not from knowledge-heavy reasoning. Argon trails Opus 5.5 (58) overall. A model that is not the HLE leader on the most comparable independent board does not usually top the Scale board. https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs - [2026-10-12] Resolution technicalities that favour NO: (a) the strike needs "HLE Accuracy" shown on the chart card at noon ET the day after first appearance; (b) if the first added Argon model is a lower-effort or preview variant, it still counts, and only same-day additions are compared, so a later, stronger variant does not help; (c) if Argon never appears by 2027-06-30, the contract resolves No; (d) a mid-2026 chart refresh that adds HLE-Diamond or rolling-set scores could re-baseline the displayed numbers. https://agi.safe.ai/ - [2026-10-03] Reported internal doubt: neuraltrust.ai, citing Axios, says Bloomberg reported some Google staff found Argon's performance lacking in internal testing. Google disputed this. This is a report, not a confirmed fact, but it fits Argon being a "breadth" model rather than an HLE leader. https://neuraltrust.ai/blog/gemini-4-argon - [2026-10-12] Gemini delays: meetcody.ai notes that Gemini 3.5 Pro was promised for June 2026 and missed, so Google's schedule record is poor. A long delay to general availability, and so to a chart entry, is plausible within the 8.5-month window. https://meetcody.ai/blog/gemini-4-argon-gemini-pro-delay/ BOTTOM LINE: The NO case rests on three points. The only independent measure (57.1%, text-only) is below 60% and trails Opus 5.5. Argon has no public API, so a chart entry may be late or may never come before 2027-06-30. And the Scale and CAIS full-set numbers for Gemini models have run well under AA's. Only a high-variance reading from the site's own method, or a stronger later variant that appears first, would give 60%+. NOT FOUND: a live read of the agi.safe.ai chart; any Argon entry on the Scale or CAIS board; Google's own HLE figure for Argon; any CAIS or Scale statement on when Argon will be evaluated; the usual lag between a Gemini release and its appearance on the chart.