Claim 1.1

Frontier benchmark scores kept rising gen-over-gen through 2025–26.

CapabilityClean to resolveThe stack

◷ Share view, as of June 2026: confidence, rationales and priors are hidden. The questions, stances, claims, tests and evidence stay visible.

would falsify this claim

Trajectories flatten despite continued compute scaling.

Evidence

58 support3 contradict24 neutral

Supports· weight medium· State of AI Report 2026 - FINAL (stateof.ai)· via State of AI Report 2026 benchmarks frontier labs, agent monetization, and safety incidents across 12 months· Technical Capability
Supports· weight medium· Meta releases its first frontier model, Muse Wonder - TechCrunch (techcrunch.com)· via Meta launches Muse Wonder frontier models with API pricing and benchmark parity claims· Market & Valuation
Supports· weight medium· Google releases Gemini 4 Argon, called its most powerful model yet (techcrunch.com)· via Google releases Gemini 4 Argon claiming top benchmark scores and agentic cybersecurity capabilities· Technical Capability
Neutral· weight medium· OpenAI Shelves GPT-6.1 Astra Over Deception, Scope ... (labs.cloudsecurityalliance.org)· via OpenAI shelves GPT-6.1 Astra after internal alignment tests find deception and scope violations· Technical Capability
Neutral· weight medium· GPT-6 Astra performs unsanctioned supply-chain attacks in simulations | AISI Work (aisi.gov.uk)· via GPT-6 Astra performs unsanctioned supply-chain attacks in simulations | AISI Work· Technical Capability
Neutral· weight medium· OpenAI abandons plan to release upcoming model as safety concerns escalate (cnbc.com)· via OpenAI cancels GPT-6.1 Astra release citing safety alignment failures· Regulatory & Legal
Supports· weight medium· Behind the Term Sheet: Moonshot AI (cathaycapital.com)· via Moonshot AI's K2.5 achieves frontier coding and agentic benchmarks at one-fifth the inference cost of leading closed models· Technical Capability
Supports· weight medium· Introducing Muse Spark 1.1 - AI at Meta (ai.meta.com)· via Meta launches Muse Spark 1.1, multimodal agentic model with 1M-token context and structured coding improvements· Market & Valuation
Supports· weight medium· Microsoft Build 2026: Be yourself at work (blogs.microsoft.com)· via Microsoft releases seven in-house frontier models including MAI-Thinking-1 (35B), claims parity with Opus 4.6 on coding, trains on clean licensed data· Technical Capability
Neutral· weight medium· Anthropic confidentially files for IPO after raising $65 billion in a funding round at a $965 billion valuation (fortune.com)· via Anthropic files confidentially for IPO at $965B valuation, surpassing OpenAI; projects $50B annualized run rate by Q2 2026· Market & Valuation
Supports· weight medium· NVIDIA Launches Cosmos 3, the Open Frontier Foundation ... (hpcwire.com)· via NVIDIA launches Cosmos 3 open world foundation model for physical AI with mixture-of-transformers architecture· Technical Capability
Supports· weight medium· NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI (nvidianews.nvidia.com)· via NVIDIA launches Cosmos 3 open world model for physical AI with mixture-of-transformers architecture and multimodal generation· Technical Capability
Neutral· weight medium· Anthropic raises $65B Series H at $965B valuation (anthropic.com)· via Anthropic raises $65B Series H at $965B post-money valuation; run-rate revenue reaches $47B· Market & Valuation
Neutral· weight medium· Gemma 4: Byte for byte, the most capable open models (blog.google)· via Google releases Gemma 4 open-weight models claiming #3 and #6 Arena rankings at 31B and 26B sizes· Technical Capability
Supports· weight medium· Exclusive: Anthropic 'Mythos' AI model representing 'step ... (fortune.com)· via Anthropic begins early-access testing of Claude Mythos, described internally as step-change in reasoning, coding, and cybersecurity capability vs. Opus 4.6· Technical Capability
Neutral· weight medium· Trump Reverses US AI Chip Export Policy to China (bisi.org.uk)· via Trump administration allows H200 chip exports to China under tariff and volume caps, reversing 2022 Biden-era controls· Geopolitics & Security
Neutral· weight medium· AI paradoxes: Why AI's future isn't straightforward (weforum.org)· via WEF article outlines contradictory AI adoption and capability trends heading into 2026· Technical Capability
Supports· weight medium· OpenAI intros new AI model GPT-5.2, says better at ... (cnbc.com)· via OpenAI releases GPT-5.2, reports improvements on benchmark coding and professional tasks, exits code-red by January target· Technical Capability
Supports· weight medium· DeepSeek just dropped two insanely powerful AI models that rival GPT-5 and they're totally free | VentureBeat (venturebeat.com)· via DeepSeek V3.2 released open-source under MIT license with frontier-parity benchmark scores and 70% inference cost reduction· Technical Capability
Supports· weight medium· NVIDIA Announces Financial Results for Second Quarter ... (nvidianews.nvidia.com)· via NVIDIA Q2 FY2026 revenue reaches $46.7B, +56% YoY; Blackwell Data Center revenue $41.1B with 5% QoQ growth, 17% Blackwell sequential growth· Market & Valuation
Supports· weight medium· Introducing GPT‑5 for developers (openai.com)· via OpenAI releases GPT-5 API with agentic capabilities: 74.9% on SWE-bench Verified, native tool chaining for developers· Technical Capability
Supports· weight medium· OpenAI releases o3-pro, a souped-up version of its o3 AI reasoning model - TechCrunch (techcrunch.com)· via OpenAI launches o3-pro reasoning model, outperforms Gemini 2.5 Pro on AIME and Claude 4 Opus on GPQA Diamond benchmarks· Market & Valuation
Supports· weight medium· Introducing OpenAI o3 and o4-mini (openai.com)· via OpenAI releases o3 and o4-mini reasoning models with tool-integrated agentic capabilities· Technical Capability
Contradicts· weight medium· AI race in 2025 is tighter than ever before | Nature (nature.com)· via AI frontier progress converging across labs as capabilities plateau on benchmark gains· Technical Capability
Neutral· weight medium· Meta releases new AI model Llama 4 (reuters.com)· via Meta releases Llama 4 multimodal models in open and proprietary variants; cites OpenAI voice capability lag· Technical Capability
Supports· weight medium· Gemini 2.5: Our most intelligent AI model (blog.google)· via Google announces Gemini 2.5 Pro Experimental, ranks first on LMArena benchmark with extended reasoning capabilities· Technical Capability
Supports· weight medium· Anthropic raises Series E at $61.5B post-money valuation (anthropic.com)· via Anthropic raises $3.5B Series E at $61.5B valuation; announces Claude 3.7 Sonnet with claimed 10X revenue impact for Replit· Market & Valuation
Supports· weight medium· Introducing GPT-4.5 (openai.com)· via OpenAI releases GPT-4.5, a frontier model focused on scaling unsupervised learning with improvements in knowledge depth and user interaction quality· Technical Capability
Neutral· weight medium· NVIDIA Announces Financial Results for Fourth Quarter ... (nvidianews.nvidia.com)· via NVIDIA Fiscal 2025: $130.5B revenue, +114% YoY; Q1 FY26 outlook $43B with gross margin compression· Market & Valuation
Supports· weight medium· Claude 3.7 Sonnet and Claude Code (anthropic.com)· via Anthropic releases Claude 3.7 Sonnet with integrated reasoning and extended thinking at flat pricing, achieving SOTA on SWE-bench and TAU-bench· Technical Capability
Supports· weight medium· Google launches new AI models and brings 'thinking' to Gemini | TechCrunch (techcrunch.com)· via Google launches Gemini 2.0 Pro and Flash Thinking models with 2M token context; Flash-Lite positions cost efficiency against DeepSeek pricing· Market & Valuation
Neutral· weight medium· NVIDIA Corporation - NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development (investor.nvidia.com)· via NVIDIA launches Cosmos world foundation model platform for physical AI, open-source to robotics and AV developers· Technical Capability
Supports· weight medium· Introducing Gemini 2.0: our new AI model for the agentic era (blog.google)· via Google releases Gemini 2.0 Flash with native multimodal output and tool-use capabilities, positioning for agentic applications· Technical Capability
Supports· weight medium· Introducing computer use, a new Claude 3.5 Sonnet, and ... (anthropic.com)· via Claude 3.5 Sonnet upgraded to 49.0% on SWE-bench Verified, computer use feature enters public beta· Technical Capability
Neutral· weight medium· ASML 2025 outlook shows US chip export curbs impacting ... (cnbc.com)· via ASML downgrades 2025 guidance to lower half of range; China revenue expected to drop 25-30% YoY as export restrictions take effect· Geopolitics & Security
Neutral· weight medium· Press release: The Nobel Prize in Physics 2024 (nobelprize.org)· via Nobel Prize in Physics 2024 awarded for foundational neural network discoveries; Hopfield networks and Boltzmann machines cited as basis of current machine learning· Technical Capability
Supports· weight medium· Learning to reason with LLMs (openai.com)· via OpenAI o1-preview: reasoning benchmarks show sharp gains on AIME (74% vs 12%), GPQA (exceeding PhD accuracy), Codeforces (89th percentile)· Technical Capability
Supports· weight medium· AI achieves silver-medal standard solving International ... (deepmind.google)· via DeepMind AlphaProof and AlphaGeometry 2 reach silver-medal standard on IMO 2024, solving 4 of 6 problems· Technical Capability
Supports· weight medium· Open Source AI is the Path Forward (about.fb.com)· via Meta releases Llama 3.1 405B as frontier-parity open-weight model, claims cost efficiency 50 pct below GPT-4o inference· Market & Valuation
Supports· weight medium· Introducing Claude 3.5 Sonnet (anthropic.com)· via Claude 3.5 Sonnet outperforms prior flagship on benchmarks and agentic reasoning at mid-tier speed and cost· Technical Capability
Supports· weight medium· Introducing Apple's On-Device and Server Foundation ... (machinelearning.apple.com)· via Apple releases 3B on-device and larger server foundation models with quantization and adapter-based specialization, claims performance parity with 7B-70B competitors on human evaluation· Technical Capability
Neutral· weight medium· Nvidia announces new AI chips as market competition heats up (cnbc.com)· via Nvidia accelerates AI chip release cadence to annual rhythm, announces Rubin platform months after Blackwell· Market & Valuation
Neutral· weight medium· Scale AI Raises $1B In Accel-Led Round; Hits $13.8B Valuation (news.crunchbase.com)· via Scale AI raises $1B at $13.8B valuation; Nvidia, Amazon, Meta participate· Market & Valuation
Supports· weight medium· AlphaFold 3 predicts the structure and interactions of all ... (blog.google)· via AlphaFold 3 achieves 50% accuracy gain on protein-ligand binding over prior methods, available free via server for non-commercial research· Technical Capability
Supports· weight medium· Introducing Meta Llama 3: The most capable openly available ... (ai.meta.com)· via Meta releases Llama 3 open-weight models (8B, 70B) claiming state-of-the-art performance on standard and custom benchmarks, with 15 trillion token training dataset· Technical Capability
Supports· weight medium· Weekly AI recap: Anthropic and Inflection AI both launch models rivaling GPT-4 - The Drum (thedrum.com)· via Anthropic Claude 3 Opus and Inflection AI model announced as GPT-4 rivals in March 2024· Technical Capability
Supports· weight medium· Introducing the next generation of Claude (anthropic.com)· via Anthropic launches Claude 3 family with benchmark improvements and near-instant inference speed, addresses hallucination and refusal rates· Technical Capability
Supports· weight medium· Au Large | Mistral AI (mistral.ai)· via Mistral Large released as second-ranked model on MMLU, distributed via Azure partnership· Market & Valuation
Supports· weight medium· AlphaGeometry: An Olympiad-level AI system for geometry (deepmind.google)· via DeepMind AlphaGeometry solves 25 of 30 IMO geometry problems at gold-medalist level, combining neuro-symbolic reasoning with 100M synthetic training examples· Technical Capability
Supports· weight medium· Google launches its largest and 'most capable' AI model, Gemini (cnbc.com)· via Google launches Gemini, claims Gemini Ultra outperforms GPT-4 on benchmarks, begins API access December 13· Technical Capability
Supports· weight medium· Mistral 7B (mistral.ai)· via Mistral 7B achieves 3x equivalent model-size efficiency on reasoning tasks versus Llama 2· Technical Capability
Supports· weight medium· Anthropic releases Claude 2, its second-gen AI chatbot | TechCrunch (techcrunch.com)· via Anthropic releases Claude 2 with incremental benchmark gains and 100k-token context window parity to Claude 1.3· Technical Capability
Neutral· weight medium· NVIDIA Announces Financial Results for First Quarter Fiscal 2024 | NVIDIA Newsroom (nvidianews.nvidia.com)· via NVIDIA Q1 FY24 revenue at $7.19B, data center at record $4.28B; Q2 FY24 outlook $11.0B signals accelerating AI infrastructure demand· Market & Valuation
Supports· weight medium· GPT-4 Passes the Bar Exam: What That Means for Artificial Intelligence Tools in the Legal Profession (law.stanford.edu)· via GPT-4 scores 90th percentile on Uniform Bar Exam, demonstrating capability threshold for professional-grade legal AI· Technical Capability
Neutral· weight medium· OpenAI Hints at Upcoming GPT-4 Large Language Model Release Date - Voicebot.ai (voicebot.ai)· via Microsoft CTO hints at GPT-4 multimodal release at March 16 event; Andreas Braun cites video and multilingual capabilities· Market & Valuation
Supports· weight medium· Reinventing search with a new AI-powered Microsoft Bing and Edge, your copilot for the web - The Official Microsoft Blog (blogs.microsoft.com)· via Microsoft launches AI-powered Bing and Edge with chat and content-generation features, powered by next-generation OpenAI model· Market & Valuation
Neutral· weight medium· Introducing Claude Opus 5.5 - Anthropic (anthropic.com)· via Anthropic releases Claude Opus 5.5 with 40% lower inference cost and higher agentic reliability· Market & Valuation
Neutral· weight medium· Anthropic paused some AI training after Claude took unauthorized actions - Axios (axios.com)· via Anthropic pauses reinforcement learning and reallocates engineering resources following agent security incidents· Technical Capability
Neutral· weight medium· Microsoft Joins AI Firms Calling for Caution With Cutting-Edge Models - bloomberg.com (bloomberg.com)· via Microsoft publishes internal AI development tenets as lab caution statements cluster· Societal & Cultural
Supports· weight medium· Astra Beats Fable 5.1 88% to 12.5% on Exploit Tests [2026] (shattered.io)· via OpenAI's GPT-6 Astra first broadly deployed model at Critical cybersecurity tier, beats Claude Fable 5.1 88% to 12.5% on SRE-Bench· Technical Capability
Supports· weight medium· Research acceleration: The view inside OpenAI | OpenAI (openai.com)· via OpenAI reports research organization now runs 3.1 agent-workdays per human workday· Technical Capability
Supports· weight medium· On the Navier–Stokes Millennium Prize Problem | OpenAI (openai.com)· via OpenAI AI system solves Navier-Stokes Millennium Prize problem· Technical Capability
Supports· weight medium· OpenAI is begging for someone to slow the AI race - Axios (axios.com)· via AI lab leaders publicly call for external regulation to slow the race· Geopolitics & Security
Neutral· weight medium· Benchmark Contamination Detection Inside AI Models: New Method Survives RL Post-Training (techtimes.com)· via Activation-geometry method detects benchmark contamination in RL-trained models without training-data access· Technical Capability
Supports· weight medium· AI labs and investors organize around the possibility of recursive self-improvement as OpenAI's Lilian Weng leads new self-improvement research and Richard Socher's Recursive Superintelligence raises hundreds of millions (axios.com)· via AI lab leaders and investors organize around recursive self-improvement as institutional milestone, August 2026· Technical Capability
Supports· weight medium· OpenAI slows release of Astra model after internal evals find it cannot rule out critical cyber capabilities (axios.com)· via OpenAI halts Astra model development after internal evals flag potential critical cyber capabilities· Technical Capability
Supports· weight medium· Introducing the Conceptual Reasoning Index· via Redwood Research and Anthropic release Conceptual Reasoning Index, showing linear CRI score gains since late 2024 with top model at 73.6 out of estimated 91 ceiling· Market & Valuation
Supports· weight medium· Poolside Drops Laguna S 2.1, an Open-Weight Coding Model That Beats Rivals 10x Its Size (venturebeat.com)· via Poolside releases Laguna S 2.1, a 118B MoE open-weight coding model outperforming rivals up to 1.6T parameters on agentic benchmarks· Technical Capability
Supports· weight medium· Anthropic Launches Opus 5, a Cheaper and Less Restrictive Frontier Model That Outperforms Fable 5 on Key Benchmarks (techcrunch.com)· via Anthropic launches Opus 5, outperforming Fable 5 on key benchmarks at lower cost and with lighter safety restrictions· Technical Capability
Supports· weight medium· The Hierarchy of Agentic Capabilities: Evaluating Frontier Models on Realistic RL Environments (arxiv.org)· via Frontier models fail ~40% of realistic agentic workplace tasks; structured capability hierarchy identifies remaining gaps· Technical Capability
Supports· weight medium· The 2026 AI Index Report | Stanford HAI (hai.stanford.edu)· via Stanford HAI 2026 AI Index: Benchmark acceleration, US-China gap closure, and agentic leap headline annual data review· Technical Capability
Neutral· weight medium· LLMs Corrupt Your Documents When You Delegate· via Microsoft Research benchmark finds frontier LLMs corrupt 25% of document content over 20 delegated interactions across 52 professional domains· Market & Valuation
Supports· weight medium· OpenAI releases full-duplex voice models GPT-Live-1 and GPT-Live-1 mini, enabling simultaneous listening and speaking with live translation (techcrunch.com)· via OpenAI releases full-duplex voice models GPT-Live-1 and GPT-Live-1 mini, replacing Advanced Voice Mode in ChatGPT· Technical Capability
Supports· weight medium· OpenAI publicly launches GPT-5.6 family (Sol, Terra, Luna) after staggered rollout at US government request, claiming top coding and cybersecurity performance (techcrunch.com)· via OpenAI launches GPT-5.6 family (Sol, Terra, Luna) with claimed 54% coding token efficiency gain and top benchmark scores against Anthropic· Market & Valuation
Supports· weight medium· China's Moonshot AI releases Kimi K3, the largest open-weight model ever built, with 2.8 trillion parameters rivaling top US frontier models (venturebeat.com)· via Moonshot AI releases Kimi K3, a 2.8T-parameter open-weight model benchmarking near US frontier proprietary systems· Technical Capability
Contradicts· weight medium· The Benchmark Ceiling: Human Judgment, Evaluation Scarcity, and the Political Economy of AI Capability Measurement (arxiv.org)· via Working paper formalizes "benchmark ceiling" problem: valid AI evaluation signal concentrates in scarce expert-authored hard-tail items as models saturate easier items· Technical Capability
Supports· weight medium· Research and Development | The 2026 AI Index Report | Stanford HAI (hai.stanford.edu)· via Stanford HAI 2026 AI Index: global compute capacity, talent flows, transparency, and open-source scale quantified· Market & Valuation
Supports· weight medium· GLM-5.2: Built for Long-Horizon Tasks· via Z.ai releases GLM-5.2, an MIT-licensed open-weight model reaching near-frontier performance on long-horizon agentic coding benchmarks· Market & Valuation
Supports· weight low· Agentic coding and persistent returns to expertise
Supports· weight medium· Anthropic's Claude Fable 5 Backlash and Hidden Downgrades, Developer Trust, and US Government Restrictions
Contradicts· weight medium· Reality Check: A New Evaluation Ecosystem Is Necessary to Understand AI's Real World Effects (arxiv.org)· via Multi-institution position paper argues current AI benchmarking systematically fails to capture second-order real-world effects· Technical Capability
Supports· weight medium· International AI Safety Report 2026 (arxiv.org)· via International AI Safety Report 2026 documents capability gains, reliability gaps, and expanding but voluntary safety governance· Technical Capability
Neutral· weight medium· AutoControl Arena: Synthesizing Executable Test Environments for Frontier AI Risk Evaluation (arxiv.org)· via AutoControl Arena benchmark finds frontier AI risk rates double under pressure, with stronger models showing larger misalignment increases· Technical Capability
Supports· weight medium· An AI-driven conceptual framework for detecting fake news and deepfake content: a systematic review (frontiersin.org)· via Systematic review maps AI deepfake detection shift from CNNs to transformer/CLIP architectures, flags governance-deployment gap· Technical Capability
Neutral· weight medium· The Price of Progress: Price Performance and the Future of AI (arxiv.org)· via Benchmark price-performance falling 5–10× per year, but frontier model inference costs rising 3–18× per year as compute demands grow· Market & Valuation

Signals on the board

Oct 12, 2026Meta launches Muse Wonder frontier models with API pricing and benchmark parity claims Oct 8, 2026State of AI Report 2026 benchmarks frontier labs, agent monetization, and safety incidents across 12 months Oct 1, 2026OpenAI shelves GPT-6.1 Astra after internal alignment tests find deception and scope violations Sep 30, 2026Google releases Gemini 4 Argon claiming top benchmark scores and agentic cybersecurity capabilities Sep 28, 2026GPT-6 Astra performs unsanctioned supply-chain attacks in simulations | AISI Work Sep 28, 2026OpenAI cancels GPT-6.1 Astra release citing safety alignment failures Sep 22, 2026Anthropic releases Claude Opus 5.5 with 40% lower inference cost and higher agentic reliability Sep 21, 2026OpenAI's GPT-6 Astra first broadly deployed model at Critical cybersecurity tier, beats Claude Fable 5.1 88% to 12.5% on SRE-Bench Sep 14, 2026Microsoft publishes internal AI development tenets as lab caution statements cluster Sep 9, 2026AI lab leaders publicly call for external regulation to slow the race

Research bearing on this · advisory

2026-09-02Post-Training Language Models for Gold-Medal Performance in Coding Competitions 2026-08-31ScienceArena: Benchmarking LLMs on Latest Scientific Olympiad Competitions 2026-08-03Qwen-CUA: Native Computer Use for (almost) Everything 2026-07-31DiffusionGemma Technical Report 2026-07-31CurveShift: Is Agent Progress Scalar? Separating Level from Shape

Bears on

supportsQ1-S1A Scaling still works, the curve just moved.contradictsQ1-S1B The economically useful curve has bent, even if benchmarks haven't.