AI Models
AI Models is a data page on the AI desk of Glassbox Auditing. One page per AI model: its lab, every public benchmark it was measured on, its history, cost and news, beside the corporate footprint's record of who built it.
Last updated: .
What is on this page today?
- The major labs published 867 new model families on Hugging Face in 2026 to 8 Nov, -5% on the same date last year — most active this year by repositories published: nvidia (681), LiquidAI (343), allenai (263), inclusionAI (217), microsoft (208) · 910 by this date last year · a family counts once however many sizes and quantized… (Hugging Face, 7 Nov 2026)
- 19 of 333 models on OpenRouter's price list cut their output price since 26 Sep, 40 raised it — biggest cut: OpenAI: gpt-oss-120b from $0.6 to $0.079 per million output tokens · OpenRouter's public price list, logged daily · prices are not adjusted for capability (OpenRouter, 6 Nov 2026)
- LMArena's top text rating is 1539 (Gemini 4 Carbon (High)), +33 points in 90 days — the best rating among today's ten highest models, on 25 dated snapshots since 27 Jul 2026 · blind pairwise votes, Elo scale (LMArena, 5 Nov 2026)
- New model scored: Mistral Large 4 Preview, 152.6 on the index — Mistral AI · France · open weights · released 6 Oct · 16.4 points behind the leader (Epoch AI Capabilities Index, 5 Nov 2026)
- New model scored: GPT-6.1 Terra, 167.8 on the index — OpenAI · United States of America · released 26 Oct · 1.2 points behind the leader (Epoch AI Capabilities Index, 29 Oct 2026)
- New model scored: DeepSeek V4.1 Flash 1017, 159.3 on the index — DeepSeek · China · open weights · released 17 Oct · 9.8 points behind the leader (Epoch AI Capabilities Index, 29 Oct 2026)
- New model scored: Beam 1.0, 150.8 on the index — Reflection · United States of America · open weights · released 5 Oct · 18.3 points behind the leader (Epoch AI Capabilities Index, 29 Oct 2026)
- New model scored: Muse Wonder, 162.0 on the index — Meta AI · United States of America · released 12 Oct · 7.0 points behind the leader (Epoch AI Capabilities Index, 20 Oct 2026)
- New model scored: Muse Spark Open, 155.4 on the index — Meta AI · United States of America · open weights · released 15 Oct · 13.7 points behind the leader (Epoch AI Capabilities Index, 20 Oct 2026)
- Epoch AI catalogued 931 AI models released in 2024, +82% on 2023 (index 554, 2019 = 100) — 104 notable, 12 frontier, 394 open-weights · 2025: 587 so far; 2026: 232 so far (the catalogue fills in over 12-18 months) (Epoch AI model catalogue, 30 Dec 2024)
Where does the data come from?
Items on this page come from Hugging Face; OpenRouter; LMArena; Epoch AI Capabilities Index; Epoch AI model catalogue. Each item links to its own source. All sources, how items are checked.
What does this page lead to?
- Products & Share (Companies): New models are sold through the developers' and their partners' cloud lines (Azure, Google Cloud, AWS), the product revenue these companies report. (moves the same way; quarters; company scale) Source: SEC EDGAR XBRL company facts (10-Q/10-K).
- Business & Industry (AI): Cheaper, more capable models lower the cost of putting AI into a business process; Census's survey shows how many firms do. (moves the same way; months; economy scale) Source: OpenRouter model prices.
- Regulatory Exposure (Companies): Models trained with very large compute (above 10^25 operations) are presumed to carry systemic risk under the EU AI Act, which adds testing, reporting and security duties for their developers. (moves the same way; months; company scale) Source: Official texts: EUR-Lex, state legislatures, the Federal Register and national gazettes (each law links its own).
- AI Supply Chain (AI): Training compute for notable models has grown several-fold a year (Epoch AI); each larger training run is spending on chips, power and data-center space. (moves the same way; quarters; sector scale) Source: Epoch AI notable models catalogue (CC BY 4.0).
- Competitors (Companies): A capable open or low-priced model can undercut software firms that name AI as a competitive risk in the competition section of their annual reports. (either way, depending on the cause; months; company scale) Source: SEC EDGAR filings (10-Q/10-K text).
- Nation (AI): Frontier training runs and the data centers behind them are investment in computers, software and structures, the spending the Nation page reads in GDP. (moves the same way; years; economy scale) Source: Epoch AI notable models catalogue (CC BY 4.0).
- Benchmarks (AI): New models are run on the public benchmarks once released, so a stronger model raises the best scores and closes more of each benchmark's remaining room. (moves the same way; weeks; sector scale) Source: Epoch AI Benchmarking Hub (CC BY 4.0).
What feeds this page?
- AI Supply Chain (AI): Labs train on the clusters the supply chain delivers, so more accelerators and data-center power let them run larger training runs; Epoch AI records each notable model's training compute. (moves the same way; quarters; sector scale) Source: Epoch AI notable models catalogue (CC BY 4.0).