# E123 — CAIN vs METR Time Horizons

**Verdict: CAIN is unranked on METR's general long-task horizon. It is not a frontier general agent.**

METR's metric is the 50%-task-completion time horizon: fit P(success) against log2(human minutes), read the
duration where the fitted curve crosses 50%. METR timed humans on HCAST, RE-Bench and 66 shorter tasks. CAIN has
not run that suite, so no comparable 50% horizon exists.

## Published METR TH1.1 (external; not re-run by CAIN)

| Model | Release | 50% horizon |
|---|---|---:|
| claude_mythos_preview_early_inspect | 2026-04-07 | 1045 min | 17.4 h |
| claude_opus_4_6_inspect | 2026-02-05 | 719 min | 12.0 h |
| gemini_3_1_pro | 2026-02-19 | 384 min | 6.4 h |
| gpt_5_2 | 2025-12-11 | 352 min | 5.9 h |
| gpt_5_3_codex | 2026-02-05 | 350 min | 5.8 h |
| gpt_5_4 | 2026-03-05 | 342 min | 5.7 h |
| claude_opus_4_5_inspect | 2025-11-24 | 293 min | 4.9 h |
| gemini_3_pro | 2025-11-18 | 224 min | 3.7 h |
| gpt_5_1_codex_max_inspect | 2025-11-19 | 224 min | 3.7 h |
| gpt_5_2025_08_07_inspect | 2025-08-07 | 203 min | 3.4 h |
| o3_inspect | 2025-04-16 | 120 min | 2.0 h |
| claude_4_1_opus_inspect | 2025-08-05 | 100 min | 1.7 h |
| claude_4_opus_inspect | 2025-05-22 | 100 min | 1.7 h |
| claude_3_7_sonnet_inspect | 2025-02-24 | 60 min | 1.0 h |
| o1_inspect | 2024-12-05 | 39 min | 0.7 h |
| claude_3_5_sonnet_20241022_inspect | 2024-10-22 | 21 min | 0.3 h |
| o1_preview | 2024-09-12 | 20 min | 0.3 h |
| claude_3_5_sonnet_20240620_inspect | 2024-06-20 | 11 min | 0.2 h |
| gpt_4o_inspect | 2024-05-13 | 7 min | 0.1 h |
| gpt_4_1106_inspect | 2023-11-06 | 4 min | 0.1 h |
| gpt_4 | 2023-03-14 | 4 min | 0.1 h |
| claude_3_opus_inspect | 2024-03-04 | 4 min | 0.1 h |
| gpt_4_turbo_inspect | 2024-04-09 | 4 min | 0.1 h |
| gpt_3_5_turbo_instruct | 2022-03-15 | 1 min | 0.0 h |
| davinci_002 | 2020-05-28 | 0 min | 0.0 h |
| gpt2 | 2019-02-14 | 0 min | 0.0 h |

METR notes measurements above 16 h are unreliable with the current suite; doubling time is ~187.8 days all-time.

## CAIN

- METR suite run: **No**
- Comparable 50% horizon: **None**
- Six-hour governed workload: **RUNNING** at build (started 1795138211, deadline 1795159811)
- If it completes, it is a **single 360-minute workload success**, not a 50% horizon.
- E118 and E119 gains are on **constructed** tasks, not human-timed METR tasks.

## Governance axis

CAIN tests identity, scoped single-use permits, resource-side enforcement, signed receipts, revocation, a KILL
fence and an offline verifier. METR's suite measures task success vs human duration; it does not measure those
properties. This is a different axis, not an intelligence score.

## Conclusion

CAIN is **not** ranked against Claude, GPT, Gemini or other METR-tested agents on general long-task autonomy.
The published frontier range at 50% is roughly **1–17 hours**. CAIN's measured differentiator is verifiable
governance and bounded, re-checkable research evidence — not a frontier time horizon.

Limits: no METR suite, no human-timed CAIN tasks, 6 h RUNNING, constructed OOD not comparable, no external reviewer.
