Will the next Claude Opus model debut with a Humanity’s Last Exam score of 55% or higher?

- [2026-09-22] Anthropic released Claude Opus 5.5 (API id `claude-opus-5-5`), the first model in a new Claude 5.5 family and the successor to Opus 5; this is the "next Claude Opus" relevant to this market, which opened 2026-08-17 (i.e., after Opus 5's July release). https://support.claude.com/en/articles/12138966-release-notes - [2026-09-22] Anthropic's Opus 5.5 System Card (dated September 22, 2026) states the model is a broad capability upgrade over Opus 5, with the largest gains in agentic coding, visual reasoning, computer use, and long-horizon professional knowledge work. https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf - [2026-09-22] Per multiple secondary write-ups of Anthropic's own launch benchmark table, Opus 5.5 scores 67.7% on HLE "with tools" and 64.4% on HLE "without tools," both far above the 55% threshold. https://llm-stats.com/blog/research/claude-opus-5-5-launch - [2026-09-22] Independent verification from Artificial Analysis reportedly puts Opus 5.5's HLE score somewhat lower, around 61.4%, still above 55%, versus Anthropic's self-reported 67.7% figure — the two houses disagree on exact magnitude but both clear the 55% bar. https://kingy.ai/blog/claude-opus-5-5-specs-benchmarks-pricing-comparison/ - [2026-09-23] The official Polymarket page for this exact market ("Next Claude Opus Model: Humanity's Last Exam Debut?", opened Aug 17 2026, closing Dec 31 2026/resolving Jan 1 2027) currently prices "55%+" at roughly 50-53¢ (near a coin flip), while "45%+" and "50%+" trade at 96-97%, and "60%+" trades near 3-5%, reflecting trader uncertainty about which HLE configuration/number the agi.safe.ai site will display for Opus 5.5. https://polymarket.com/event/next-claude-opus-model-humanitys-last-exam-debut-20261231 - [2026-09-23] The market's own auto-generated context note states Opus 5.5 (released Sept 22) "posts competitive HLE scores around 61–68% (with tools)" and flags this as the key near-term catalyst for resolution. https://polymarket.com/event/next-claude-opus-model-humanitys-last-exam-debut-20261231 - [2026-09-23] As of this check, the agi.safe.ai homepage content retrieved does not yet show an updated "AI Progress on Humanity's Last Exam" results card explicitly listing Claude Opus 5.5 — the dashboard subpage (agi.safe.ai/dashboard) could not be accessed, so it is unconfirmed whether Opus 5.5 has yet been formally added to the official leaderboard that governs resolution. https://agi.safe.ai/ - [2026-09-21] Third-party trackers (not the official resolution source) show the highest previously-listed Anthropic Opus score at 54.9% (Claude Opus 5) and Claude Fable 5.1 leading overall at 59.1%, indicating that prior to Opus 5.5, no Opus-labeled model had cleared 55% on at least some trackers. https://pricepertoken.com/leaderboards/benchmark/hle - [2026-08-09] A prior, separate but similarly-named Polymarket market for "the next Claude Opus" (opened Jul 23, 2026, referring to Opus 5) had already resolved "Yes" for its 55%+ threshold outcome before the current market (opened Aug 17, 2026) even began, confirming these are distinct sequential markets tied to specific model releases. https://polymarket.com/event/next-claude-opus-humanitys-last-exam-debut-20260723184439892 BOTTOM LINE: Claude Opus 5.5 launched September 22, 2026 with Anthropic-reported HLE scores of 64.4% (no tools) and 67.7% (with tools) — both comfortably above the 55% threshold — and even the more conservative independent Artificial Analysis figure (~61.4%) clears 55%; the main remaining uncertainty is whether/when this exact score is picked up and displayed under the "HLE Accuracy" label on the agi.safe.ai chart, which is what governs resolution, and traders were pricing the 55%+ outcome near a coin flip as of Sep 23. NOT FOUND: Could not directly confirm (via a live fetch of agi.safe.ai's dashboard/results chart) whether Claude Opus 5.5 has already been added to the official "AI Progress on Humanity's Last Exam" leaderboard, nor the exact displayed "HLE Accuracy" percentage and configuration (with/without tools) that the site uses for its result card, since the dashboard subpage was inaccessible during this research.

- [2025-11-24] Claude Opus 4.5 debuted on HLE at roughly 43.2% (reported score cited on social media at launch), continuing the general upward trend from near-zero in early 2025 but still well below the 55% line. Source: https://x.com/diegocabezas01/status/1993077636894277675 - [2026-02-05] Claude Opus 4.6 launched, described by outlets as posting "leading performance on Humanity's Last Exam" but still <40% under a plain (no-tools) low-effort configuration, illustrating how heavily HLE scores for Opus-class models have historically depended on the tool/effort configuration used. Source: https://intuitionlabs.ai/articles/humanitys-last-exam-ai-benchmark ("Even the latest frontier models – GPT-5 Pro, Gemini 3 Pro, and Claude Opus 4.6 – have yet to break 40% on HLE") - [2026-08-07] For the immediately preceding Opus debut (Claude Opus 4.8), Anthropic's own system card reported Anthropic's Claude Opus 4.8 system card reports 49.8% HLE accuracy without tools and 57.9% with tools , which straddled the 55% line depending on configuration; the resulting Polymarket market for that debut priced thresholds up to 50% near 100% while the market implied deep skepticism that a 55%+ figure (the tool-assisted config) would be the one used for settlement. Source: https://cryptoslate.com/predictions/market/next-claude-opus-humanitys-last-exam-debut/ - [2026-07-24] Claude Opus 5 launched; independent trackers later listed its HLE score at roughly 54.9% (no-tools/standard leaderboard configuration), i.e. just under the 55% bar, per third-party leaderboard aggregation as of September 2026. Source: https://pricepertoken.com/leaderboards/benchmark/hle ("Claude Opus 5 at 54.9%") - [2026-09-22] Claude Opus 5.5 — the "next" Opus after this market's Aug 17, 2026 open — launched, with Anthropic-reported HLE "with tools" figures around 67.7%, while an Artificial Analysis "Adaptive Reasoning" configuration lists 61.4%; no single clean "no-tools" figure is confirmed in available coverage, underscoring that the same underlying model can be reported at very different HLE numbers depending on tool/effort settings. Sources: https://siliconangle.com/2026/09/22/anthropic-releases-claude-opus-5-5-and-openai-counters-with-two-cheaper-gpt-6-models/ ("Humanity's Last Exam, run with tools, returned 67.7%"), https://www.vellum.ai/blog/claude-opus-5-5-benchmarks-explained ("With tools enabled, Opus 5.5 tops the board at 67.7%") - [2026-09-23] Pattern across the last four Opus debuts (4.5→4.6→4.8→5→5.5) shows scores climbing from the low 40s to the high 60s (tool-assisted) over roughly ten months, but the "no-tools"/standard configuration — which is what most third-party leaderboards (Artificial Analysis, pricepertoken, llm-stats) actually display — has hovered just below or around the 50-55% band as recently as Opus 5 (54.9%). This bimodal history (mid-40s to low-50s "no tools" vs. high-50s to high-60s "with tools") is exactly the ambiguity the market is pricing. - [2026-09-23] Live Polymarket pricing for this exact market ("Next Claude Opus Model: Humanity's Last Exam Debut?") shows the 45%+ threshold at 96%, 50%+ at 97%, 55%+ trading at essentially a coin flip (Yes 53¢ / No 53¢, i.e. ~50%), and 60%+ at only ~3-5%. This is direct evidence that professional forecasters/traders see the 55% line as the genuine uncertainty boundary, with strong confidence the score clears 50% but real doubt it clears 55%. Source: https://polymarket.com/event/next-claude-opus-model-humanitys-last-exam-debut-20261231 - [2026-08-09] The prior related market ("Next Claude Opus: HLE Debut", settled around Opus 4.8's debut window) resolved with thresholds through 50% at 100% Yes, i.e. that debut cleared 50% but the corresponding 55%+ outcome had earlier traded near-zero probability, consistent with the officially-displayed (no-tools) score landing just under 55% historically. Source: https://cryptoslate.com/predictions/market/next-claude-opus-humanitys-last-exam-debut/ ("thresholds through 50% trade near certainty while 55% receives only a 0.8% Yes price") - [2026-09-23] A related but distinct Polymarket market ("Will the highest score achieved by an Anthropic Claude model on HLE in 2026 be 55%+") — which allows any Claude model, not just the "next Opus debut" — trades much higher, around 94%, reflecting that Anthropic's top models (Fable/Mythos variants) have already cleared 55-59% on some configurations; this is a different, easier bar than the specific "next Opus debut" question. Source: https://polymarket.com/event/next-claude-opus-model-humanitys-last-exam-debut-20261231 (related-markets sidebar) BOTTOM LINE: The reference class of the last four Opus HLE "debuts" shows scores rising steadily but landing right around the 50-55% line under the standard (no-tools) configuration that most public leaderboards display, while tool-assisted configurations run well above 55-68%; this exact ambiguity is reflected in the live market for this question pricing the 55%+ threshold at essentially 50/50, versus near-certainty for 45%+/50%+ and near-impossibility for 60%+. NOT FOUND: The precise "no-tools" HLE Accuracy figure for Claude Opus 5.5 as displayed on the agi.safe.ai model card/chart itself (as opposed to Anthropic's or Artificial Analysis's own tool-assisted figures) could not be confirmed from available sources; agi.safe.ai's homepage does not expose the live leaderboard chart data via simple fetch, so it is unclear which configuration's number safe.ai has posted or will post for Opus 5.5.

**Most likely path to YES:** Claude Opus 5.5 (released 2026-09-22, model id `claude-opus-5-5`) is already the qualifying "next Opus" for this market, and every currently published HLE figure for it — from Anthropic itself and from an independent third party — clears the 55% bar, unlike prior Opus debuts where a lower "no-tools" figure created genuine ambiguity. - [2026-09-22] Anthropic's own launch benchmark table lists a single published HLE configuration for Opus 5.5 — "Humanity's Last Exam, with tools (67.7%)" — with no separate "no tools" figure given in the vendor table this cycle, unlike the Opus 4.8 cycle. https://kingy.ai/blog/claude-opus-5-5-specs-benchmarks-pricing-comparison/ - [2026-09-22] The same table shows Opus 5.5's 67.7% beating Fable 5.1 (65.6%), Opus 5 (63.6%) and GPT-6 Astra (57.2%) on this metric, and Anthropic reports a standard error tight enough that this is described as a real, non-noise lead on the adjacent Terminal-Bench metric. https://kingy.ai/blog/claude-opus-5-5-specs-benchmarks-pricing-comparison/ - [2026-09-22] Independent verification came within a day: "Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) scores the highest on Humanity's Last Exam with a score of 61.4%, followed by Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) with a score of 59.1%" — this is Artificial Analysis's own reproduction, not a vendor self-report, and it still clears 55% by over 6 points. https://artificialanalysis.ai/evaluations/humanitys-last-exam - [2026-09-22] Kingy AI's side-by-side explicitly frames Artificial Analysis's 61.4% as "the first major independent evaluation" of Opus 5.5, confirming this is a genuinely separate, non-Anthropic data point that agrees directionally with the vendor claim even while landing lower. https://kingy.ai/blog/claude-opus-5-5-specs-benchmarks-pricing-comparison/ - [2026-09-22] Unlike the Opus 4.8 cycle — where Anthropic explicitly published both a sub-55% "no tools" figure (49.8%) and a "with tools" figure (57.9%), creating a two-number ambiguity that suppressed the 55%+ price to near-zero at the time — no comparably low "no-tools" HLE figure for Opus 5.5 has surfaced in vendor or independent coverage as of this research, removing the main historical source of downside risk. https://www.vellum.ai/blog/claude-opus-4-8-benchmarks-explained - [2026-09-18] Third-party trackers already had the pre-5.5 Opus line (Opus 5) sitting near or just above 55% on some methodologies — e.g., BenchLM's tracker listed "Claude Fable 5.1 leads the HLE leaderboard with 65% , followed by Claude Opus 5 (64.7%)" — indicating that even before Opus 5.5, some public leaderboard configurations for the Opus line were already comfortably in the 60s, not just the 50-55% band other trackers showed. https://benchlm.ai/benchmarks/hle - [2026-09-23] llm-stats.com's own HLE leaderboard (a different aggregator than pricepertoken) shows "Claude Fable 5.1 by Anthropic leads with a score of 0.650" , again demonstrating that on at least two independent public trackers the Anthropic flagship line is already scoring in the mid-60s, a full 10 points clear of the 55% line, making it more likely Opus 5.5 lands similarly high once added. https://llm-stats.com/benchmarks/humanity's-last-exam - [2026-09-23] agi.safe.ai's live homepage confirms the site is an actively maintained, continuously updating results dashboard "While current LLMs achieve very low accuracy on Humanity's Last Exam, recent history shows benchmarks are quickly saturated -- with models dramatically progressing from near-zero to near-perfect performance in a short timeframe" , and the page structure (a dynamically loading "HLE results" panel) is consistent with new frontier models such as Opus 5.5 being added within days of a high-profile release like this one. https://agi.safe.ai/ - [2026-09-21] The prior Opus (Opus 5) had already been benchmarked by pricepertoken at 54.9% and by BenchLM at 64.7% on the same underlying model — a roughly 10-point spread between trackers — showing that whichever configuration/tracker-equivalent number safe.ai has historically used for Opus-line models, it is not uncommon for that number to land in the high-50s-to-mid-60s range rather than the low 50s, which is the range that would resolve this market YES. - [2026-09-23] Because Opus 5.5's own vendor table this cycle only reports the "with tools" HLE figure (67.7%) rather than a bifurcated tools/no-tools pair, any resolution source that pulls directly from Anthropic's most prominently published number — as safe.ai has done for other vendor submissions in the past — has a structurally higher chance of registering a figure well above 55% than in the Opus 4.8 cycle. BOTTOM LINE: The strongest case for YES rests on convergent evidence — Anthropic's self-reported 67.7% and Artificial Analysis's independently reproduced 61.4% for Opus 5.5, plus multiple third-party leaderboards already placing recent Anthropic flagships in the 60-65% range — all of which clear 55% by a comfortable margin; the main missing link is direct confirmation that agi.safe.ai has added Opus 5.5 to its official chart and which specific number it displays, which was not yet observable via direct fetch during this research. NOT FOUND: Could not directly fetch agi.safe.ai's dashboard/results subpage to confirm (a) whether Opus 5.5 has already been added to the official chart, or (b) which exact HLE configuration (Anthropic's 67.7% with-tools, Artificial Analysis's 61.4%, or some other in-house safe.ai run) the site will display as "HLE Accuracy" for Opus 5.5.

- [2026-08-09] The most directly analogous precedent — the prior "Next Claude Opus" HLE debut market covering Claude Opus 4.8, whose system card straddled the threshold with 49.8% "no tools" vs. 57.9% "with tools" — settled with its official Polymarket snapshot confirming only thresholds up to 50%+ at 100%, while the 55%+ leg had traded at just 0.8% Yes beforehand and is not reported among the confirmed-Yes tiers, consistent with "The final snapshot shows 35%+ at 100%, 40%+ at 100%, 45%+ at 100%, and 50%+ at 100%" only — i.e., the market's own resolution mechanics rewarded the no-tools figure, not Anthropic's touted tool-assisted number, the last time an Opus system card straddled 55%. - [2026-08-07] CryptoSlate's analysis of that Opus 4.8 debut explicitly framed the dynamic that now recurs with Opus 5.5: "Anthropic's own evaluations place Opus on opposite sides of the decisive threshold depending on tool access. The market's sharp break therefore depends on which testing configuration reaches the official leaderboard" , and that market's central thesis was "that the next Claude Opus will score around the low 50s on Humanity's Last Exam under the configuration recognized by the official leaderboard. This matters because the next Opus needs only a small improvement over the reported 49.8% no-tools result to reach 50%" — the site has a demonstrated history of displaying the lower, no-tools number rather than Anthropic's headline tool-assisted score. - [2026-09-21] Third-party leaderboard aggregation (pricepertoken) — which tracks scores in the same no-tools/text-only style — shows Claude Fable 5.1 leading at 59.1% with "Claude Fable 5.1 at 59.1%, followed by Claude Fable 5 at 55.5% and Claude Opus 5 at 54.9%" , meaning even Anthropic's own most recent Opus-family flagship (Opus 5) sat just under the 55% line under this standard configuration, only weeks before Opus 5.5's debut. - [2026-09-22/23] Wikipedia's HLE performance table — reflecting the benchmark's own "text-only subset" scoring convention that the official paper and site use as the standard reportable metric — lists Claude Fable 5.1 at 59.1% as the top entry, with GPT-6 Astra second at 54.7%; no Claude Opus model appears anywhere in this top-tier table, suggesting Opus-branded models have historically scored below the levels these trackers surface under the standard (no-tools) methodology that mirrors agi.safe.ai's typical convention. - [2026-09-23] Live Artificial Analysis pages for Opus 5.5 report the model's HLE figure only under specific configuration labels — e.g. "Opus 5.5 took the top spot on the Artificial Analysis Intelligence Index with an index score of 58 at max effort, leading on six of the ten core evaluations, including Humanity's Last Exam (61.4%)" — and this 61.4% is explicitly a "max effort"/Adaptive Reasoning score from a third-party lab, not confirmed to be the "no tools," single-shot number agi.safe.ai itself uses; Anthropic's own more favorable figure of 67.7% is explicitly the "with tools" configuration per "Humanity's Last Exam, run with tools, returned 67.7%" , meaning both widely-cited high numbers may not be the metric that ultimately populates the official result card. - [2026-09-23] A direct fetch of agi.safe.ai's homepage today shows no live leaderboard chart or Opus 5.5 entry rendered in the static page content at all — the results/dashboard functionality lives on a separate JavaScript-driven "/dashboard" page that could not be verified as updated, meaning there is no confirmed public evidence yet that Claude Opus 5.5 has even been formally added to the site that governs resolution, one day after its Sept 22 launch. - [2026-06-16] A separate, earlier CryptoSlate write-up of a related Anthropic HLE market underscored the resolution's narrow, technicality-driven design: "the contract pays attention to a dated, source-specific artifact. A powerful internal model, a delayed public launch, or a score published after the cutoff would fail to satisfy the market rules. That is why the market can give meaningful probability to 45% while still drawing a steep line at 55%" — highlighting that laboratory self-reported benchmark numbers (like Anthropic's 67.7%) do not bind the resolution source, only whatever safe.ai actually publishes. - [2026-09-23] The rules' own technicalities compound the downside case: any qualifying model that is added and then not still displayed by 12:00 PM ET the next day does not qualify, and if the site is down through the seventh day after first appearance the market resolves "No" outright — adding pure operational/timing risk on top of the substantive scoring risk, independent of Opus 5.5's true capability. - [2026-09-23] Live Polymarket pricing for this exact market has the 55%+ outcome trading near a coin flip (~50-53¢) versus near-certainty (96-97%) for 45%+/50%+, indicating that informed traders — despite seeing Anthropic's and Artificial Analysis's headline numbers clearing 55% — still assign roughly even odds that the specific number safe.ai posts will fall short, exactly mirroring the pattern seen in the Opus 4.8 precedent where the tool-assisted number cleared 55% but the eventual settlement (implicitly, the no-tools config) did not. BOTTOM LINE: The single most relevant precedent (Opus 4.8) shows that when Anthropic's tool-assisted HLE score clears 55% but its no-tools score does not, the market/official site has historically settled on the lower, sub-55% figure; combined with Opus 5's official-style no-tools score sitting at 54.9% just one release prior, and no confirmed evidence yet that Opus 5.5 has even been added to agi.safe.ai's actual results chart, there is a credible, evidence-based path to a "No" resolution despite Anthropic's headline 67.7%/64.4% claims. NOT FOUND: The actual agi.safe.ai dashboard/results-chart entry and displayed "HLE Accuracy" percentage for Claude Opus 5.5 (with configuration label) could not be retrieved via direct fetch; it remains unconfirmed which configuration (no-tools vs. with-tools) the site's result card uses by convention, and whether Opus 5.5 has been added to the site at all as of this report.