← all series

arXiv submissions

• Domain: outside the three domains
• Role: contrast case: volume
• Metric: research output; preprints submitted to arXiv per month
• Coverage: 1991-07 to 2026-08, monthly, the last month partial at the 2026-08-10 fetch
• Data: arxiv-by-month.csv; per-category arxiv-categories-by-month.csv
• Upstream: https://arxiv.org/stats/monthly_submissions, with per-category counts from https://oaipmh.arxiv.org/oai
• Verdict: accelerating — a 28,450 submissions/month mean over 2026-01 to 2026-07 against monthly means of 23,707 in 2025 and 20,336 in 2024

arXiv monthly submissions

Submissions per month by field

Primary category only, grouped to arXiv's own top level.

Mathematics subfields, all of them

Click a legend entry to isolate a subfield; shift-click to compare several. Colours repeat across the thirty-plus series, so the legend, not the hue, identifies a line.

Definition

arXiv is a preprint repository for physics, mathematics, computer science and adjacent fields. It publishes its own count of submissions per month, from the archive's opening in July 1991 onward, and its public metadata records each paper's submission history and category labels.

An event in this series is one preprint submitted, counted in its submission month; the total series is arXiv's published count. The per-category file counts each paper once, in the month of its first version's submission date, under its primary (first-listed) category; categories are grouped to arXiv's own top level, and math-ph sits with physics because arXiv puts it there. The last row of the monthly file is the month in progress at fetch time. Neither file carries an authorship field.

Facts

• span: from 17,271 submissions in November 2022 (the month ChatGPT was released) to 29,687 in July 2026, the last complete month; 72% growth
• 2026 rate: a 28,450 submissions/month mean over 2026-01 to 2026-07 against monthly means of 23,707 in 2025 and 20,336 in 2024
• computer science: from 5,967 monthly submissions in November 2022 to 12,409 in July 2026, and above physics every month since August 2023
• physics and mathematics: physics rose 41% from November 2022 to July 2026; mathematics 66%, from 3,143 to 5,216
• combinatorics: math.CO reached 743 submissions in July 2026 against a 2024 monthly average of 342
• mathematics against 2024: mathematics ran at 1.6 times its 2024 monthly average in July 2026

The collection-wide cumulative index redraws this series as cumulative submissions to date:

Method

fetch.py downloads arXiv's own monthly-submissions file and keeps two columns. The raw download carries a historical_delta column, nonzero only for corrections to 1991–1997, which is dropped.

The per-category file behind the two field charts is built by fetch_categories.py, which is run by hand rather than by make fetch. It counts each paper once, in the month of its first version's submission date and under its primary (first-listed) category, and it has two interchangeable inputs: arXiv's official metadata snapshot — the ~5 GB arxiv-metadata-oai-snapshot.json distributed via Kaggle, which needs a login to download and is kept beside this document but deliberately not committed — and a no-credentials OAI-PMH harvest of the same metadata, which walks about 2,400 resumption pages at the pace arXiv meters out and takes the better part of a day. Both produce the same aggregation; papers first submitted after the repository's snapshot date are dropped. A few dozen migrated records carry v1 dates before arXiv opened in July 1991, and the charts start at the launch month.

figure.py draws the total series through lib/families.py's shared volume shape: years on the x-axis, the count on the y-axis, January 2026 onward shaded, an open marker on a part period. The three volume folders draw through that one shape. The series is drawn in slate rather than the blue used elsewhere for human or uncredited finders because it has no authorship field. The field grouping and the legacy-archive mapping — alg-geom, q-alg, dg-ga and funct-an to their modern names, so the older panels keep their early-1990s history — live at the top of figure.py, and an archive the mapping does not know fails the build rather than landing in a silent bucket. The subfield grid draws every math.* primary category, one panel each, sorted by total volume and scaled independently. check.py recomputes the fact lines from both CSVs.

Limitations

• no authorship field. The files record months, categories and counts only; whether a human or a model wrote a paper is not recorded.
• volume, not discovery. The series counts artifacts submitted; the domain folders count results, and the two need not move together.
• no denominator of effort. Nothing here measures how many researchers were working, so output per unit of input cannot be separated from more input.
• composition is invisible. A rise is consistent with more short papers, more salami-slicing, or more of the same kind of work.
• field mix is not held fixed. arXiv's growth over 35 years includes whole disciplines joining the platform.
• primary categories undercount interdisciplinary work. Each paper counts once, under its first-listed category; cross-lists are ignored, and a shift in cross-listing habits moves no line here.
• category labels are chosen. Labels are the authors' and moderators' choices, and arXiv occasionally reclassifies old papers, which moves history on a refetch.

AI attribution

The dataset carries no authorship field; no AI share can be computed from it. arxiv-by-month.csv holds a month and a count per row, and arxiv-categories-by-month.csv holds a month, a category and a count; no AI credit appears in either vendored file or in arXiv's published monthly-submissions series as of the 2026-08 fetch.

Sources

• https://arxiv.org/stats/monthly_submissions — the monthly totals in arxiv-by-month.csv.
• https://oaipmh.arxiv.org/oai — the no-credentials harvest behind arxiv-categories-by-month.csv.
• [@nist2026cvegrowth] — NIST's report of 263% growth in CVE submissions between 2020 and 2025, with its enrichment of nearly 42,000 CVEs in 2025 failing to keep pace; a submission count in the vulnerabilities domain, the same unit type as this series.
• output-crossref and output-github-pushes — the other two volume series, drawn through the same shared shape; they count DOI registrations and git pushes where this folder counts preprint submissions.