METR researchers estimate that the length of software tasks, measured in human time, that AI agents complete with 50% success doubled about every 207 days (95% interval 166–240) from 2019 to 2025, while performance was “much lower on less structured, ‘messier’ tasks.”

SocioLens project,

Claim LD01-SL-C23 (v1.0)

Provenance and status

Claim type: Estimate
Publication status: Eligible for publication, with a qualifier
Verification state: Checked against the primary source; the source is a preprint. Who checked this claim is not recorded for it; claim checks so far were AI-led, and no person has checked claims against their sources.
Does not license: A finding that systems act as agents in organizations, set goals or take initiative; or that the trend holds for messy real-world work
Limitations: Read in arXiv version 4 (10 July 2026), marked “Preprint. Under review.”; no venue is confirmed in the file read; tasks are self-contained and automatically scored; the result depends on the human baseliner population; one author works for Anthropic, the developer of the model that drafted this debate, and Anthropic models, including Claude 3.7 Sonnet, are among those measured
Source (as cited): Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” arXiv:2503.14499v4, 2026
Source type: Preprint (recorded as “Preprint (arXiv version 4, marked “Under review”)”)
Institutional position: Research nonprofit (METR), which works with AI developers; one author is at Anthropic (work done at METR)
Maturity of the evidence: Emerging evidence
Recorded date: 16 October 2026
Currentness / correction: This is the current version. No correction or concern is recorded on its sources.

Sources this version cites

Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” arXiv:2503.14499v4, 2026 (Preprint)
Read the original

Version

This page shows version v1.0. Its record never changes, even when newer evidence arrives; the site may add a note to the page or withhold a figure not cleared for public use.

This pinned version is also the current version.

Versions: v1.0 · current

Raw record: every recorded field, with its coded values

All 25 canonical Claim fields, verbatim, with the coded values derived from them. The plain-language labels above are read from these.

Claim type (coded): ESTIMATE (as printed: “ESTIMATE”)
Publication status (coded): ELIGIBLE (QUALIFIED) (as printed: “ELIGIBLE WITH QUALIFIER”)
Verification state (coded): CONTENT_VERIFIED_PRIMARY (PREPRINT)
Source type (coded): PREPRINT
Maturity (coded): EMERGING

#	Canonical field	Canonical value (verbatim)	Normalized (Claim Value Vocabulary v1.0.2)
1	CLAIM ID	LD01-SL-C23	—
2	PUBLIC-FACING CLAIM	METR researchers estimate that the length of software tasks, measured in human time, that AI agents complete with 50% success doubled about every 207 days (95% interval 166–240) from 2019 to 2025, while performance was “much lower on less structured, ‘messier’ tasks.”	—
3	CLAIM TYPE	ESTIMATE	base_types: ESTIMATE
4	RESEARCH STRAND	B — Evidence that could tell the accounts apart	—
5	MATURITY	EMERGING	states: EMERGING
6	SOURCE	Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” arXiv:2503.14499v4, 2026	source_ids: SRC-SL-KWA-ET-AL-METR-ARXIV-2503-14499
7	SOURCE TYPE	Preprint (arXiv version 4, marked “Under review”)	classes: PREPRINT
8	INSTITUTIONAL POSITION	Research nonprofit (METR), which works with AI developers; one author is at Anthropic (work done at METR)	—
9	SCOPE	Software and research-engineering tasks with timed human baselines; models released 2019–2025	—
10	DESIGN	Task suite with human completion times; logistic fit of model success against task length; trend over model release dates	—
11	SAMPLE / DATASET	170 tasks; 11–12 models	—
12	MODEL / TOOL / VERSION	Frontier models 2019–2025 with agent scaffolding	—
13	COMPARATOR	Human baseliners’ completion times	—
14	MAIN RESULT	Doubling time 207 days (166–240); 80% time horizons several times shorter; the authors extrapolate a one-month horizon between mid-2028 and mid-2031 “if these results generalize” (the conclusion says mid-2030)	—
15	INFERENTIAL TARGET	How long a task AI agents can carry out without help	—
16	DOES NOT LICENSE	A finding that systems act as agents in organizations, set goals or take initiative; or that the trend holds for messy real-world work	—
17	VERIFICATION STATE	CONTENT VERIFIED PREPRINT	state: CONTENT_VERIFIED_PRIMARY · qualifiers: PREPRINT · verified_by: UNKNOWN
18	LIMITATIONS	Read in arXiv version 4 (10 July 2026), marked “Preprint. Under review.”; no venue is confirmed in the file read; tasks are self-contained and automatically scored; the result depends on the human baseliner population; one author works for Anthropic, the developer of the model that drafted this debate, and Anthropic models, including Claude 3.7 Sonnet, are among those measured	—
19	CONTRADICTORY / SUPPORTING SOURCES	The same measure revised: LD01-SL-C24; agency as a graded property: LD01-SL-C10	—
20	GEOGRAPHY / LANGUAGE	Not applicable / English	—
21	DISTRIBUTIONAL LEDGER	Not analysed in this claim	—
22	R01 / FIELD-MAP LINKS	LD-01, T4, D4b	—
23	WHAT WOULD CHANGE THIS CLAIM?	A trend that breaks on realistic tasks, or one that holds on them	—
24	PUBLICATION STATUS	ELIGIBLE WITH QUALIFIER	state: ELIGIBLE · qualifiers: QUALIFIED
25	RECORDED DATE	2026-10-16	—

Field 22 references resolved against Field Map v0.2: LD-01 (LIVING_DEBATE); T4 (TENSION); D4b (SUB_AREA)

The sources as recorded, with their record IDs and coded source types:

Pinned Sources (field 6, version-specific)

“Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” arXiv:2503.14499v4, 2026” → Kwa et al. (2026), “Measuring AI Ability to Complete Long…” (SRC-SL-KWA-ET-AL-METR-ARXIV-2503-14499--v1.0) — Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” arXiv:2503.14499v4, 2026; class PREPRINT (explicit public provenance pin)

Cited by Living Debate versions

What kind of thing is generative AI, and does the answer predict anything? (v0.1) LD-01

Cite this page

SocioLens project. (2026, October 16). METR researchers estimate that the length of software tasks, measured in human time, that AI agents complete with 50% success doubled about every 207 days (95% interval 166–240) from 2019 to 2025, while performance was “much lower on less structured, ‘messier’ tasks.” [Claim LD01-SL-C23, version v1.0]. SocioLens. https://sociolens-web.pages.dev/claims/ld01-sl-c23/v1.0/

@misc{sociolens_ld01_sl_c23_v1_0,
  author = {{SocioLens project}},
  title = {METR researchers estimate that the length of software tasks, measured in human time, that AI agents complete with 50\% success doubled about every 207 days (95\% interval 166–240) from 2019 to 2025, while performance was “much lower on less structured, ‘messier’ tasks.”},
  year = {2026},
  date = {2026-10-16},
  howpublished = {\url{https://sociolens-web.pages.dev/claims/ld01-sl-c23/v1.0/}},
  note = {Claim LD01-SL-C23, version v1.0}
}

The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.

Research-of-record provenance

Materialized by sociolens-materializer 0.6.0 at 2026-09-28T20:00:00.000Z; normalization contract SocioLens — Claim Value Vocabulary v1.0.2 — Pre-Spike Materialization Contract (v1.0.2). Content seal SHA-256 63ea65a1b00ef56f6dd7f61c48c3faf9870a1406eeba3cb1cc4f1ea9771e88df.

Repository-authored inputs (not Drive exports)

SocioLens claim — LD01-SL-C23 v1.0 (v1.0) — written in the SocioLens repository by Project Lead (Claude, AI) on 2026-10-16; no Drive identity exists. Authority: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 16 October 2026 to build the LD-01 debate; a new Claim, first version, written from the project's LD-01 source notes, checked against full texts, official web pages or official abstracts. Block “ENTRY 23 — LD01-SL-C23” SHA-256 549f79a50121f4a7424a73f7cf444d63436a5336c02254592cf22a5c1831204c; fixture fixtures/authored/claims/LD01-SL-C23--v1.0.

Builder decisions (not Drive data)

VERSION = v1.0 — basis: Claim version taken from the repository-authored claim fixture version (SocioLens claim — LD01-SL-C23 v1.0); the first version of a new Claim of the LD-01 debate
CURRENT_POINTER = true — basis: repository-authored claim "SocioLens claim — LD01-SL-C23 v1.0" (Project Lead (Claude, AI), 2026-10-16); basis: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 16 October 2026 to build the LD-01 debate; a new Claim, first version, written from the project's LD-01 source notes, checked against full texts, official web pages or official abstracts
FIELD_6_SOURCE_REFERENCE = SRC-SL-KWA-ET-AL-METR-ARXIV-2503-14499--v1.0 — basis: field 6 of ENTRY 23 — LD01-SL-C23 cites "Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” arXiv:2503.14499v4, 2026"; the only field-6 segment of this ClaimVersion's repository-authored ledger block.
SOURCE_PROVENANCE_PIN = SRC-SL-KWA-ET-AL-METR-ARXIV-2503-14499--v1.0 — basis: explicit version-pinned public provenance of this ClaimVersion (Amendment v1.2 §6): the pinned field-6 Source record "Kwa et al., “Measuring AI Ability to Complete Long Software Tasks,” arXiv:2503.14499v4, 2026".