Independent benchmark evidence

The frontier moved. Here is the evidence.

GPT-5.6 Sol (pro, max) currently leads the composite dataset at 162.1 ECI. Individual benchmarks show uneven progress: hard reasoning, autonomous task duration, and evaluation durability remain distinct constraints rather than one finish line.

Current composite frontier: 162.1 ECI
Observed leader: GPT-5.6 Sol (pro, max)
Evidence basis: 6 public measures

Evidence-weighted, not a countdown. See how the observatory separates measurement from speculation.

Epoch Capabilities Index

Frontier index

162.1 +1.1 ECI frontier change

A stitched general-capability scale built from more than 50 benchmarks.

Explore the observed frontier by year 2019 0.0 GPT-2 (1.5B)

Useful for comparing the frontier over long periods, but not a probability or countdown to AGI.

Inspect source data

Optional interpretation

Ask GPT-OSS to explain this signal

The model receives the selected public record, not an open web prompt. It cannot alter the score.

01 / Public record

Signals shaping the frontier

Each row is the latest observed leader for one measure. Expand it to see the prior-frontier comparison and source.

Epoch Capabilities Index Jul 09, 2026 GPT-5.6 Sol (pro, max) defines the observed frontier index frontier 162.1 on Epoch Capabilities Index. Useful for comparing the frontier over long periods, but not a probability or countdown to AGI. +1.1 ECI High confidence

GPT-5.6 Sol (pro, max), reported by OpenAI, improved the observed frontier by +1.1 ECI relative to the previous dated leader in this dataset. Scores can change as benchmark maintainers add or revise evaluations.

Open Epoch AI data

ARC-AGI-2 Jul 09, 2026 GPT-5.6 Sol (Max) defines the observed generalization frontier 92.5% on ARC-AGI-2. Strong performance can reveal flexible reasoning, while benchmark saturation can weaken the signal. +2.5 pp Medium confidence

GPT-5.6 Sol (Max), reported by OpenAI, improved the observed frontier by +2.5 pp relative to the previous dated leader in this dataset. Scores can change as benchmark maintainers add or revise evaluations.

Open Epoch AI data

FrontierMath Apr 23, 2026 GPT-5.5 Pro pre-release (high) defines the observed reasoning frontier 52.4% on FrontierMath. Higher scores show stronger hard-problem solving; they do not establish broad reliability. +0.7 pp Medium confidence

GPT-5.5 Pro pre-release (high), reported by OpenAI, improved the observed frontier by +0.7 pp relative to the previous dated leader in this dataset. Scores can change as benchmark maintainers add or revise evaluations.

Open Epoch AI data

SWE-bench Verified Apr 16, 2026 Claude Opus 4.7 (max) defines the observed software work frontier 83.5% on SWE-bench Verified. Measures bounded repository tasks, not unattended ownership of production systems. +4.8 pp Medium confidence

Claude Opus 4.7 (max), reported by Anthropic, improved the observed frontier by +4.8 pp relative to the previous dated leader in this dataset. Scores can change as benchmark maintainers add or revise evaluations.

Open Epoch AI data

METR task horizon (80%) Feb 19, 2026 Gemini 3.1 Pro Preview defines the observed autonomy frontier 1.5 hr on METR task horizon (80%). Longer horizons matter, but a benchmark task is not the same as safe, persistent agency. +20 min Medium confidence

Gemini 3.1 Pro Preview, reported by Google DeepMind, improved the observed frontier by +20 min relative to the previous dated leader in this dataset. Scores can change as benchmark maintainers add or revise evaluations.

Open Epoch AI data

Video-MME Jun 18, 2025 video-SALMONN 2+ defines the observed multimodal frontier 79.7% on Video-MME. Captures one form of visual-language understanding, not embodied world modelling. +4.7 pp Medium confidence

video-SALMONN 2+, reported by ByteDance, improved the observed frontier by +4.7 pp relative to the previous dated leader in this dataset. Scores can change as benchmark maintainers add or revise evaluations.

Open Epoch AI data

02 / Counterweight

Still unsolved

High scores are evidence of capability. They are not evidence that every hard problem disappeared.

Reliability under change

Not captured by one frontier score

A model can lead a benchmark and still fail under distribution shift, adversarial inputs, or changed tools.

Does the capability survive outside the evaluation setup?
Inspect supporting methodology
Long-horizon autonomy

1.5 hr at an estimated 80% success threshold

Longer task horizons are improving, but reliable multi-hour work remains different from open-ended, self-correcting agency.

Can the system recover from its own mistakes without hidden human repair?
Inspect supporting methodology
Evaluation durability

A moving measurement target

Benchmarks lose information as they saturate, leak into training data, or reward narrow strategies.

Would the result hold on a fresh, private, contamination-resistant test?
Inspect supporting methodology

03 / Method

Evidence-weighted, not a countdown.

Road to AGI displays the published Epoch Capabilities Index alongside individual benchmark frontiers. It does not convert those measurements into an AGI probability, arrival date, or hidden proprietary score.

What the index does

01 Use dated benchmark records and preserve the original unit.
02 Show the frontier and the previous frontier so movement is inspectable.
03 Keep capability evidence separate from unresolved reliability questions.
04 Link every displayed signal to the underlying public source.

Update cycle

The public archive is checked every day at 04:17 UTC. Automation rebuilds the committed evidence snapshot, and a connected Cloudflare Pages project deploys the refreshed result from the production branch.

Where the model stops

Groq-hosted GPT-OSS is optional and only explains already-selected evidence. It does not choose sources, alter scores, or determine the frontier.

The frontier moved. Here is the evidence.

Frontier index

Ask GPT-OSS to explain this signal

Evidence-weighted, not a countdown.

What the index does

Update cycle

Where the model stops

Primary references