benchmark

Task-completion time horizons

Human-task duration associated with 50% agent success.

Research autonomy

What this measure tells us

Supporting measure of research and software-task autonomy.

Human task duration is not uninterrupted AI runtime.
Task distribution and scaffold revisions change estimates.

Results by version

Benchmark version

Metric

Results to compare

Protocol

TH1.1, GPT-5.6 Sol June 2026 assessment

2026-06-26

METR's June 2026 TH1.1 software-task assessment uses the ReAct harness family; exact revision/configuration and task count are not stated. Cheating attempts scored as failures. Approximate 50%-horizon estimate, which METR does not consider robust. Source hours retained.

Metric definitions

50% task-completion horizon (hours): A task-specific measurement; not a percentage of RSI achieved.

Published results · TH1.1, GPT-5.6 Sol June 2026 assessment
Key	System / organization	Metric	Reported result	Date	Protocol	Verification	Evidence
—	GPT-5.6 Sol / METR horizon evaluation	50% task-completion horizon	around 11.3 hoursConfidence interval: 95% CI: 5–40 hoursCheating counted as failure; METR considers this estimate not reliable.	2026-06-26	th-1-1-sol-p	independent evaluation	Summary of METR's predeployment evaluation of GPT-5.6 Sol ↗

TH1.1

2026-01-29 · 228 tasks

Appendix estimates published together on January 29; evaluation dates not supplied.

Metric definitions

50% task-completion horizon (minutes): A task-specific measurement; not a percentage of RSI achieved.

Task-completion time horizons · 50% task-completion horizon

limited comparison · Point estimates; uncertainty is shown in the table.

Configurations and budgets not fully captured in this extraction. No automatic deltas or joined calendar trend.

Published results · TH1.1
Key	System / organization	Metric	Reported result	Date	Protocol	Verification	Evidence
6	Claude Opus 4.5 / METR horizon evaluation	50% task-completion horizon	320 minutesConfidence interval: [170, 729] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-1-p	independent evaluation	Time Horizon 1.1 ↗
3	GPT-5 / METR horizon evaluation	50% task-completion horizon	214 minutesConfidence interval: [117, 480] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-1-p	independent evaluation	Time Horizon 1.1 ↗
5	Claude Opus 4 / METR horizon evaluation	50% task-completion horizon	101 minutesConfidence interval: [58, 170] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-1-p	independent evaluation	Time Horizon 1.1 ↗
4	o3 / METR horizon evaluation	50% task-completion horizon	121 minutesConfidence interval: [74, 201] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-1-p	independent evaluation	Time Horizon 1.1 ↗
7	Claude 3.7 Sonnet / METR horizon evaluation	50% task-completion horizon	60 minutesConfidence interval: [32, 106] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-1-p	independent evaluation	Time Horizon 1.1 ↗
2	GPT-4 1106 / METR horizon evaluation	50% task-completion horizon	3.6 minutesConfidence interval: [1.6,7.5] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-1-p	independent evaluation	Time Horizon 1.1 ↗
1	GPT-4 0314 / METR horizon evaluation	50% task-completion horizon	3.5 minutesConfidence interval: [1.6,6.9] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-1-p	independent evaluation	Time Horizon 1.1 ↗

TH1 (historical estimates)

Release date not reported · 170 tasks

Appendix estimates published together on January 29; evaluation dates not supplied.

Metric definitions

50% task-completion horizon (minutes): A task-specific measurement; not a percentage of RSI achieved.

Task-completion time horizons · 50% task-completion horizon

limited comparison · Point estimates; uncertainty is shown in the table.

Configurations and budgets not fully captured in this extraction. No automatic deltas or joined calendar trend.

Published results · TH1 (historical estimates)
Key	System / organization	Metric	Reported result	Date	Protocol	Verification	Evidence
9	Claude Opus 4.5 / METR horizon evaluation	50% task-completion horizon	289 minutesConfidence interval: [110, 1268] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
4	GPT-5 / METR horizon evaluation	50% task-completion horizon	138 minutesConfidence interval: [68, 281] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
7	Claude Opus 4 / METR horizon evaluation	50% task-completion horizon	86 minutesConfidence interval: [44, 144] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
6	o3 / METR horizon evaluation	50% task-completion horizon	94 minutesConfidence interval: [48, 165] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
10	Claude 3.7 Sonnet / METR horizon evaluation	50% task-completion horizon	56 minutesConfidence interval: [28, 94] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
1	GPT-5.1-codex-max / METR horizon evaluation	50% task-completion horizon	173 minutesConfidence interval: [81,411] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
12	Claude Sonnet 4.5 / METR horizon evaluation	50% task-completion horizon	122 minutesConfidence interval: [59,252] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
8	Claude Opus 4.1 / METR horizon evaluation	50% task-completion horizon	114 minutesConfidence interval: [56,215] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
5	Grok 4 / METR horizon evaluation	50% task-completion horizon	109 minutesConfidence interval: [48,235] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
11	Claude Sonnet 4 / METR horizon evaluation	50% task-completion horizon	75 minutesConfidence interval: [38,132] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
3	GPT-4 1106 / METR horizon evaluation	50% task-completion horizon	8.5 minutesConfidence interval: [4.0,16.1] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗
2	GPT-4 0314 / METR horizon evaluation	50% task-completion horizon	5.4 minutesConfidence interval: [2.5,10.3] minutes · Source-reported bootstrap interval; level not specified in this table	2026-01-29	th-1-0-p	independent evaluation	Time Horizon 1.1 ↗

Version lineage

TH1.1, GPT-5.6 Sol June 2026 assessmentTH1 (historical estimates) ← TH1.1TH1 (historical estimates) → TH1.1

Reference points

No reference points are indexed for this measure.

Availability

What is publicly available
Resource	Status
public description	yes
public results	yes
public tasks	yes
public code	yes
public evaluation service	unknown

Official sources

Time Horizon 1.1Task-suite changes; Appendix: Changes to Model Horizon EstimatesOriginal source ↗