# 5 ais predict arcagi3
Published: 2026-03-27
URL: https://diegoprime.com/blog/5-ais-predict-arcagi3

[arc-agi-3](https://arcprize.org/arc-agi/3) launched march 25. every frontier llm scored below 0.4%. one day later, anthropic's next model ["claude mythos" leaked](https://m1astra-mythos.pages.dev/). i gave the same prompt to 5 ai deep research tools: predict what mythos will score.

### the prompt

> What will Claude Mythos score on ARC-AGI-3? Produce a rigorous probabilistic prediction. Search extensively. Reason from first principles. Take as long as you need.

same prompt, no follow-ups, let each one run independently.

### predictions

| ai                     | expected value | median |
| ---------------------- | -------------- | ------ |
| cc opus 4.6 (terminal) | 4.2%           | 0.7%   |
| sonnet 4.6 extended    | ~2%            | ~0.3%  |
| opus 4.6 extended      | —              | —      |
| gemini deep research   | 3.15%          | ~2-3%  |
| chatgpt deep research  | 0.97%          | 0.35%  |

opus extended wrote a 10k word article instead of a forecast. same conclusion, no single number.

all five agreed: mythos improves things llms are already good at. arc-agi-3 measures something llms are structurally bad at. expected values ranged from 0.97% to 4.2%, medians clustered around 0.3% to 0.7%. all five found the same key evidence independently (the duke harness paradox, stochasticgoose beating every llm by 34x, RHAE-squared killing marginal gains).

### how each one reasoned

**cc opus** in terminal was the most efficient. same insight as the 10k word reports in ~800 words.

**sonnet** was the cleanest forecaster. tight probability buckets, calibrated confidence.

**opus** went journalist mode. market impact analysis, architectural recommendations, 317 sources in 9 minutes.

**gemini** wrote an academic paper. invented the term "modality collapse." most detailed analysis of the leak itself.

**chatgpt** was the most conservative. nine probability buckets, lowest EV (0.97%), lowest self-assessed confidence (30%).

same evidence, same conclusion, completely different outputs. i'll update this when actual scores drop.

## Connections

- [[Software, AI & Building]]
- [[density-of-intelligence]]