Benchmarks

Benchmarks is a data page on the AI desk of Glassbox Auditing. Every public AI benchmark: its best score now and a year ago, how much of its room is closed, U.S. against Chinese models, and how long a task AI can finish on its own (METR). The capability index, the lab rankings and the prediction markets on AI are all built from these results.

Last updated: 8 Nov 2026.

What is the latest reading?

EBR-bench: each new best score: 78.1% as of 26 Oct 2026, +70.2 pts since Aug 2025. Source: Epoch AI Benchmarking Hub (CC BY 4.0).

What is on this page today?

The benchmark registry holds 97 benchmarks and 106,524 results in 11 categories; General reasoning has the most tests (33) — 2,974 model entries in the biggest category · results are ranked only against others run the same way (Epoch AI Benchmarking Hub, LMArena, SWE-bench and other public leaderboards, 8 Nov 2026)
Anthropic's models hold the best score under 57 of 202 live benchmark set-ups, OpenAI's 55 — 202 of 332 set-ups had a result in the last 12 months · most are agents tests (81) · a set-up is one benchmark run one way, never mixed with another (Epoch AI Benchmarking Hub, LMArena, SWE-bench and other public leaderboards, 6 Nov 2026)
18,130 benchmark results were logged for the first time in the last 7 days, BFCL the most (5,341) — 15,839 in the 7 days before · 48 benchmarks added results · snapshot leaderboards add a row only when a score changes (Stanford HELM, LMArena, Berkeley Function Calling Leaderboard, 6 Nov 2026)
GPQA diamond: Gemini 4 Argon leads 203 models tested at 96.4% — set-up: Epoch AI evaluation · higher is better · best score so far +10.2 on a year earlier · dated by the tested models' release dates (Epoch AI Benchmarking Hub, 5 Nov 2026)
EBR-bench: the best model closes 78% of the room, up 65 points in 12 months — best score 78.1 by GPT-6.1 Terra (OpenAI), released 26 Oct 2026 · 13% a year earlier (Epoch AI Benchmarking Hub, 25 Oct 2026)
APEX-Agents: the best model closes 82% of the room, up 78 points in 12 months — best score 82.4 by Gemini 4 Argon (Google), released 30 Sep 2026 · 4% a year earlier (Epoch AI Benchmarking Hub, 29 Sep 2026) GOOGL
MirrorCode: the best model closes 79% of the room, up 74 points in 12 months — best score 78.9 by Claude Opus 5.5 (Anthropic), released 22 Sep 2026 · 4% a year earlier (Epoch AI Benchmarking Hub, 21 Sep 2026)
Mystery Game Puzzles: the best model closes 82% of the room, up 61 points in 12 months — best score 84.0 by GPT-6 Astra (OpenAI), released 3 Sep 2026 · 22% a year earlier (Epoch AI Benchmarking Hub, 2 Sep 2026)
GSO-Bench: the best model closes 88% of the room, up 74 points in 12 months — best score 88.2 by Claude Fable 5.1 (Anthropic), released 1 Sep 2026 · 15% a year earlier (Epoch AI Benchmarking Hub, 31 Aug 2026)
The best AI METR has measured finishes half of tasks that take experts 17.4 hours (60 minutes for models released by April 2025): Claude Mythos Preview (Anthropic, released April 2026) — doubling every 2.9 months · METR times each model months after its release (METR, 6 Apr 2026)

Where does the data come from?

Items on this page come from Epoch AI Benchmarking Hub, LMArena, SWE-bench and other public leaderboards; Stanford HELM, LMArena, Berkeley Function Calling Leaderboard; Epoch AI Benchmarking Hub; METR. Each item links to its own source. All sources, how items are checked.

What does this page lead to?

Progress Level (AI): Epoch's Capabilities Index is built from dozens of these benchmarks, so new best scores feed straight into the frontier level and its pace. (moves the same way; weeks; sector scale) Source: Epoch AI Capabilities Index (CC BY 4.0).
Labs & Countries (AI): Best scores decide which lab and which country hold the lead; a new top result widens or narrows the gaps the Labs page measures, depending on who posts it. (either way, depending on the cause; same day; sector scale) Source: Epoch AI Benchmarking Hub (CC BY 4.0).
AI Trajectory (AI): Coding-test results are an input to the trajectory's coding vector, so a lab's new coding score changes its gap to the leader there. (moves the same way; days; company scale) Source: SWE-bench leaderboards.
Labor Market (Macro): METR measures how long a task, in human working time, AI can finish on its own; as that grows, more software and office work can be automated, which can cut some roles and create others. (either way, depending on the cause; years; economy scale) Source: METR time horizons (via Epoch AI).

What feeds this page?

AI Models (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).