How AGI-pilled is each AI model?

Median forecast, years after 2026

Lower = more AGI-pilled.

Opus 5.5 vs. the field

All 8 breakthroughs occurred in 2026. Select a forecast to read the responses.

Achieved in 2026 (0 yrs)

Claude Opus 5.5 median

Field median (39 settings)

Individual forecasts. 'Never' at 2070+

Every model × every breakthrough

By lab family

Median years-out grouped by frontier lab.

Control: IMO 2025 Gold Medal

Share of 5 runs that knew AI already won IMO gold in July 2025.

How often models say "Never"

Share of 40 breakthrough answers predicting AI will never achieve it.

AI autonomous vs. human-assisted

Share of dated answers picturing AI as the primary solver rather than a human tool.

Browse all 1,755 verbatim model responses

Select any model and breakthrough below to read the 5 graded runs and full reasoning.

1. Select Model

2. Select Breakthrough

AGI forecasts vs. knowledge cutoff

Best reasoning effort per model. Lower forecasts = more AGI-pilled.

Loading reported cutoffs...

Methodology & Benchmark Design

How AGI-Pilled Bench measures internal AI timeline calibration using real 2026 breakthroughs.

Nine of the models (GPT-6 Luna, Astra and Sol, Claude Haiku, Sonnet and Opus 5.5, Gemini 3.8 Flash, Gemini 4 Argon and Grok 4.7) were run through their vendors' own CLIs (Codex, Claude Code, Antigravity, Cursor) on subscription sign-in, one headless session per answer with tools off, at each reasoning effort the model offers: low, medium, high and xhigh (Gemini 3.8 Flash and Argon stop at high). Five more (GLM 5.3, Kimi K3, Muse Spark 1.3, MiMo V2.6 Pro, DeepSeek V4.1 Flash) were run through the OpenRouter API at low effort and temperature 0.7. Every setting was sampled 5 independent times on 8 real breakthroughs that occurred between May and October 2026, plus 1 control task from July 2025 (IMO gold medal). A structured LLM grader (GPT-6 Luna) extracted whether the model said the milestone had already happened, its best-guess year, its optimistic/pessimistic range, whether it predicted "never", and whether it framed AI as acting autonomously or assisting human researchers.