Methods · Model

50%/80% horizon ratio

METR 50% horizon divided by 80% horizon for the same model (derived metric horizon_ratio_80_50). ~7-10x recently, up from ~5x.

Reason through this

The reading

The gap is widening, which is pro-normal. It is the paper's reliability argument made measurable.

Every readingeach dot links to its record

a readingnormal rangefast range

measured

25 readings drawn from the derived metric horizon_ratio_80_50; 50 observations in the evidence.

Derived (horizon_ratio_80_50) from: METR time horizons (v1.1). METR, Measuring AI Ability to Complete Long Tasks (Horizon v1.1).

The numbers

as of	reading	value	record
2026-03-05	gpt 5 4	6.34×	measured
2026-02-19	gemini 3 1 pro	4.28×	measured
2026-02-05	gpt 5 3 codex	6.39×	measured
2026-02-05	claude opus 4 6 inspect	10.3×	measured
2025-12-11	gpt 5 2	5.34×	measured
2025-11-24	claude opus 4 5 inspect	5.93×	measured
2025-11-19	gpt 5 1 codex max inspect	4.42×	measured
2025-11-18	gemini 3 pro	4.14×	measured
2025-08-07	gpt 5 2025 08 07 inspect	5.30×	measured
2025-08-05	claude 4 1 opus inspect	4.28×	measured
2025-05-22	claude 4 opus inspect	4.91×	measured
2025-04-16	o3 inspect	3.99×	measured
2025-02-24	claude 3 7 sonnet inspect	4.99×	measured
2024-12-05	o1 inspect	5.48×	measured
2024-10-22	claude 3 5 sonnet 20241022 inspect	7.91×	measured
2024-09-12	o1 preview	4.60×	measured
2024-06-20	claude 3 5 sonnet 20240620 inspect	6.82×	measured
2024-05-13	gpt 4o inspect	5.52×	measured
2024-04-09	gpt 4 turbo inspect	4.02×	measured
2024-03-04	claude 3 opus inspect	6.19×	measured
2023-11-06	gpt 4 1106 inspect	5.17×	measured
2023-03-14	gpt 4	4.48×	measured
2022-03-15	gpt 3 5 turbo instruct	2.35×	measured
2020-05-28	davinci 002	2.56×	measured
2019-02-14	gpt2	4.20×	measured

Why it matters

The ratio is the reliability gap in one number. If it collapses toward 2x while the 80% horizon exceeds eight hours, the normal-technology thesis loses its productisation argument.

What cuts against it

Ratio is sensitive to the task mix and to the exclusion of models beyond the suite ceiling; a single new model can move it.

How it is measured

derived metric horizon_ratio_80_50 (formula in the semantic layer)
source METR time horizons (v1.1) · default tier 1 · METR; analysis code MIT at https://github.com/METR/eval-analysis-public
proxy benchmark · cadence per release · valve invention to product

Normal = ratio ≥ 5 (gap holding or widening); fast = ratio ≤ 2. The falsifying condition (≤ 2 AND 80% horizon > 8 h) is compound and lives in the thesis monitor, not the band.

Applied to metric:horizon_ratio_80_50.

Model comparison (lower AIC fits better; both on ln(horizon) residuals)

fit	estimate	95% interval	n	R²	AIC	inputs
`ord_half_life_misfit`	2.04×	—–—				sources1 2

Derived rows (25)

as of	dims	value	inputs
2026-03-05	gpt_5_4	6.34×	sources1 2
2026-02-19	gemini_3_1_pro	4.28×	sources1 2
2026-02-05	gpt_5_3_codex	6.39×	sources1 2
2026-02-05	claude_opus_4_6_inspect	10.3×	sources1 2
2025-12-11	gpt_5_2	5.34×	sources1 2
2025-11-24	claude_opus_4_5_inspect	5.93×	sources1 2
2025-11-19	gpt_5_1_codex_max_inspect	4.42×	sources1 2
2025-11-18	gemini_3_pro	4.14×	sources1 2
2025-08-07	gpt_5_2025_08_07_inspect	5.30×	sources1 2
2025-08-05	claude_4_1_opus_inspect	4.28×	sources1 2
2025-05-22	claude_4_opus_inspect	4.91×	sources1 2
2025-04-16	o3_inspect	3.99×	sources1 2
2025-02-24	claude_3_7_sonnet_inspect	4.99×	sources1 2
2024-12-05	o1_inspect	5.48×	sources1 2
2024-10-22	claude_3_5_sonnet_20241022_inspect	7.91×	sources1 2
2024-09-12	o1_preview	4.60×	sources1 2
2024-06-20	claude_3_5_sonnet_20240620_inspect	6.82×	sources1 2
2024-05-13	gpt_4o_inspect	5.52×	sources1 2
2024-04-09	gpt_4_turbo_inspect	4.02×	sources1 2
2024-03-04	claude_3_opus_inspect	6.19×	sources1 2
2023-11-06	gpt_4_1106_inspect	5.17×	sources1 2
2023-03-14	gpt_4	4.48×	sources1 2
2022-03-15	gpt_3_5_turbo_instruct	2.35×	sources1 2
2020-05-28	davinci_002	2.56×	sources1 2
2019-02-14	gpt2	4.20×	sources1 2

Latest readings

2026-03-05 gpt_5_4 · 6.34×as of 2026-03-05(2 obs)
2026-02-19 gemini_3_1_pro · 4.28×as of 2026-02-19(2 obs)
2026-02-05 gpt_5_3_codex · 6.39×as of 2026-02-05(2 obs)
2026-02-05 claude_opus_4_6_inspect · 10.3×as of 2026-02-05(2 obs)
2025-12-11 gpt_5_2 · 5.34×as of 2025-12-11(2 obs)
2025-11-24 claude_opus_4_5_inspect · 5.93×as of 2025-11-24(2 obs)

Why this status, and its history

Latest non-disputed model ratio is 6.3x (>= 5 normal band); the Mythos point is excluded as beyond the 16h suite ceiling. Matches the brief's expectation. Initial seed; pending Alex's review.

2026-09-10unmeasured to consistent with normalconf — → 70 · initial seed (model-drafted)
Latest non-disputed model ratio is 6.3x (>= 5 normal band); the Mythos point is excluded as beyond the 16h suite ceiling. Matches the brief's expectation. Initial seed; pending Alex's review.
sources1 2

Why this confidence

Rests on 50 observations from METR time horizons (v1.1); best evidence tier 1, grade A; stale since 2026-03-05.

Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.

Predictions that lean on this

Happening
Ajeya Cotra's median for the longest METR fifty-percent time horizon reported by the end of 2026 was about twenty-four hours of expert work. · Ajeya Cotra
Happening
Each test AI passes shows the test's limits, not that AI has arrived. · Narayanan & Kapoor
Slower than said
If the trend speeds up, by March 2027 AI could reliably do software tasks that take people years. · AI Futures Project (Kokotajlo, Lifland, Jurkovic et al.)
Slower than said
Models might do most of a software engineer's work end to end within six to twelve months. · Dario Amodei
Too early to tell
AI improving itself cannot get past limits that sit outside the lab. · Narayanan & Kapoor

More predictions that lean on this

Not happening
AI is not an ordinary technology · This site