nikitahudov/benchmark-lifetimes

Benchmark Lifetimes

Birth and death dates of 224 AI benchmarks that nine frontier labs reported in model release materials between June 2018 and September 2026, coded for survival analysis. Each date has a source URL; each maintainer statement has a quote. Cutoff: 2026-09-30.

Code, coder outputs and the full pipeline: github.com/nikitahudov/benchmark-lifetimes. Article (Russian, Habr): link will be added on publication.

Key results

Quantity	Value
Median lifetime, all 224 benchmarks (Kaplan–Meier)	23 months (95% CI 19–27)
Born 2016–2022 / born 2023–2026	38 (27–69) / 19 (15–23) months
Hazard ratio per birth year, Cox	1.22 (1.11–1.34)
Headroom at release, median, born ≤2022 / 2023+	2.02 / 2.14 logits to the threshold
Frontier speed on the same benchmarks (Epoch AI trajectories)	0.75 → 1.56 logits per year
Born 2023+, dead of own defects or a corrected version within 24 months	14% (Aalen–Johansen)

Configurations

Config	File	Rows	What it is
lifetimes (default)	data/halflife_dataset_v3.csv	235	One row per unit: 224 units of the main analysis and 11 excluded units with the reason
decisions	data/decisions_log_v3.csv	21	Reconciliation decisions: units, codebook rule, decision, source
release_matrix	data/d3_matrix_v3.csv	6,652	Release × benchmark × score rows from 230 releases of 9 labs, with the reconciled unit (unit_v3) and whether the row is the benchmark's own score (is_report)
cards	data/cards_v3.csv	224	Per unit: first and last report, dropout from model cards, reporting after death
headroom	data/headroom_v3.csv	203	Best result at release and headroom to the threshold, in logits and percentage points

analysis_v3.json holds every result of the analysis; docs/ holds the codebook and the coder instructions.

from datasets import load_dataset
d = load_dataset("nikitahudov/benchmark-lifetimes", "lifetimes", split="train").to_pandas()
main = d[d.stratum == "main"]          # 224 units; T_months and event are the survival time and the event flag

Unit and outcome

A unit is a public benchmark, or a version of it: a new version (> 10% of items replaced, or a changed metric or protocol), an exam year (AIME 2024, 2025, 2026) or a domain (τ²-bench airline, retail, telecom). A unit enters the population if its own score was reported in official release materials by at least two of nine labs (OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, Alibaba/Qwen, Moonshot, Mistral) or by one lab in at least three releases.

Birth is the earliest of: arXiv v1, the official announcement, the public dataset, a leaderboard launch with the items; for exams, the exam date. Death is the earliest of:

Code	Event	Applies when
A	Best counted result ≥ human baseline	baseline ≥ 85% of the scale, measured by humans on the same split with a described protocol
B	Best counted result ≥ 90% of the main metric	no usable human baseline, or baseline below 85%
C	Maintainers withdraw the set, freeze submissions, deprecate it, or name saturation or defects as the reason for a replacement	public statement by the creators only
D	Organizer-declared target reached	verified by the organizers
E	Independent compromise	answer leakage documented for most frontier models; a trivial method reaches the ceiling; an audit finds ≥ 20% defective items and at least one of the nine labs stops reporting

The date of death is the date of public disclosure of the result, and the rule is "first in time, not first found". Units alive at the cutoff are censored at 2026-09; units abandoned for 12 months without lab reports or leaderboard activity are censored at their last activity. In the main mode any publicly reported system counts (bare model, scaffold or agent, ensemble) if the official leaderboard accepted it or its developer published it with a protocol; the strict-mode columns keep only single models with independent or organizer-verified runs.

Main columns of lifetimes

Column	Meaning
unit, stratum, exclusion_reason	Unit name; main or the exclusion stratum; why it was excluded
release, release_source, birth_year, cohort	Birth month (YYYY-MM) with its source; cohort used in the analysis
class, format, modality, test_access, curation, n_items	Static test, agentic environment, LLM-judge or economic/time; item format; public, hidden labels, private held-out or rolling test
human_baseline, human_type, criterion_ab, threshold	Human baseline if measured, and which of A or B applies with its threshold
ab_date, ab_system, ab_score, ab_system_type, ab_source_type, ab_source	First counted result at or above the threshold; system type (single, scaffold, ensemble); source type (self-report, leaderboard, independent)
strict_*, T_strict, event_strict	The same in the strict mode
c_date, c_reason, c_kind, c_source, c_quote	Maintainer event: reason (saturation, flaws, administrative) and kind (version, successor, freeze, withdrawal)
d_date, d_source, e_date, e_type, e_source, e_quote	Organizer target; independent compromise (i leakage, ii trivial method, iii audit)
abandoned_date, abandoned_evidence	Last activity of an abandoned unit
primary_date, primary_criterion, cause, status_cutoff	Main event and its cause: threshold (A, B, D), declared_saturation (C, saturation) or defects (C with flaws or administrative reasons, E)
T_months, event, outcome_cr	Survival time in months, event flag, competing-risk code (0 censored, 1 models, 2 instrument, 3 abandoned)
T_replacement_censored, event_replacement_censored	Sensitivity: a replacing version is censoring, not death
score_at_release, score_at_release_system	Best result at release (for annual exams: best frontier result in the first two weeks)
first_report, last_report, entry_months, n_labs_v3, n_releases_v3	Reporting by the nine labs
best_by_cutoff, confidence, date_upper_bound, recon_flags, notes	Best result at the cutoff; coder confidence; date known only as an upper bound; ids of reconciliation decisions; coder notes

How it was made

The release-material matrix was extracted from 230 releases of nine labs and normalized to canonical names with 310 regular expressions. The population was then coded by Claude agents under a fixed codebook: coder A (Claude Opus) coded all units, coder B (Claude Sonnet) blindly coded a sample of 24 units stratified by outcome. Agreement on status at the cutoff: Cohen's κ = 0.71; birth month agreed in 24 of 24 units; where both found a death, the month agreed within one month in 11 of 12. Every change to a coder A row is recorded in decisions and applied in code.

Limitations

• The 90% threshold is a convention. Lifetimes from different studies are comparable only under the same threshold and the same rule on which results count.
• 51 of 82 deaths by criteria A and B rest on developers' self-reports. In the strict mode the overall median is 39 months instead of 23; the gap between eras remains.
• Only benchmarks reported by frontier labs are included.
• 11 rows have low coder confidence, and 8 death dates are known only as an upper bound. Units born in 2026 are observed for at most eight months.

Citation

@misc{khudov2026benchmarklifetimes,
  title  = {Benchmark Lifetimes: When AI Benchmarks Are Born and When They Die},
  author = {Khudov, Nikita},
  year   = {2026},
  url    = {https://huggingface.co/datasets/nikitahudov/benchmark-lifetimes}
}

License

CC BY-SA 4.0. Scores, dates and model names are facts from the cited public sources; fields ending in _quote are short excerpts kept for verification and remain the property of their authors.