# OpenAlex: 324 million papers given away free, run by a nonprofit that spends $4.2 million a year

OpenAlex indexes 323,964,135 works under CC0 and rivals Scopus and Web of Science. The API just moved to metered pricing. What it costs, and what it gets wrong.

Author: J.A. Watte
Published: August 12, 2026
Source: https://jwatte.com/blog/openalex-open-index-of-research/

---


Clarivate's Web of Science and Elsevier's Scopus are the two databases that decide, in practice, which research counts. Universities buy rankings from them. Funders read reports built on them. Neither is free, and neither will let you have the underlying data.

OpenAlex is the open replacement, and the scale of it is not a rounding error against the incumbents. When I measured it on 9 August 2026 it held 323,964,135 works and 124,697,641 authors. The whole dataset is CC0. You can download all of it.

The organisation behind it spent $4,209,666 in its most recent filed tax year, the one ending June 2025.

That ratio is the reason this post exists.

## What it is, and where it came from

OpenAlex is a catalogue of the scholarly record modelled as a graph. The documentation names eight entity types: works, authors, sources, institutions, topics, publishers, funders and countries. Awards and keywords have their own endpoints alongside them. Every one has an ID, an endpoint, and links to the others. The name is a reference to the Library of Alexandria.

It exists because Microsoft killed something. The Microsoft Academic Graph stopped being updated on 31 December 2021, taking with it the free bibliographic backbone that a lot of tools had quietly been built on. OurResearch announced a replacement and shipped it in January 2022.

The founding paper, [arXiv:2205.01833](https://arxiv.org/abs/2205.01833) by Jason Priem, Heather Piwowar and Richard Orr, describes what launched: "metadata for 209M works (journal articles, books, etc)" plus disambiguated authors, 124k venues, 109k institutions and 65k Wikidata concepts.

Four years later the works count has gone from 209 million to 324 million, the concepts have been deprecated in favour of a cleaner topic hierarchy, and the docs have moved to a new home at `developers.openalex.org`. If you have a bookmark at `docs.openalex.org`, it redirects.

The pitch in OpenAlex's own words has not softened:

> OpenAlex offers an open replacement for industry-standard scientific knowledge bases like Elsevier's Scopus and Clarivate's Web of Science. Compared to these paywalled services, OpenAlex offers significant advantages in terms of inclusivity, affordability, and availability.

They claim "about twice the coverage of the other services," with better coverage of non-English work and work from the Global South.

## Who runs it and who pays

OurResearch is the nonprofit behind OpenAlex, and it is the same group behind Unpaywall and Impactstory. Its legal name is still Impactstory. It is a 501(c)(3) registered in Sanford, North Carolina, EIN 46-1599252, with an IRS ruling date of February 2014.

The financials from its filings are worth looking at directly, because they explain both the ambition and the fragility:

| Tax year (ending June) | Revenue | Expenses | Contributions | Contributions as share of revenue |
|---|---|---|---|---|
| 2025 | $4,275,757 | $4,209,666 | $2,388,800 | 55.9% |
| 2024 | $3,545,495 | $3,726,775 | $2,330,256 | 65.7% |
| 2023 | $2,249,041 | $2,829,054 | $1,116,360 | 49.6% |
| 2022 | $3,025,253 | $2,585,228 | $1,803,990 | 59.6% |

Four consecutive years where between half and two thirds of income is contributions rather than earned revenue. The rest is genuinely earned: program services brought in $1,885,623 in the year ending June 2025, which is 44 percent of total revenue. That is the metered API and the paid snapshot tiers doing their job.

The balance sheet is tighter than the revenue line suggests. The year ending June 2024 closed with $2,004,058 of liabilities against $1,877,329 of assets, which is to say negative net assets. The following year recovered to $1,173,136. An organisation running the largest open bibliographic index in the world has been operating inside a margin most people would find uncomfortable. The largest identifiable contributor is Arcadia, and because Arcadia publishes its full grant data as a CC0 file, the picture is unusually clear. Three grants to Our Research:

- **May 2018**: $850,000 over two years, "Search engine for open access scholarly content"
- **June 2021**: $4,500,000 over three years, "Tools to steer academic systems to open research"
- **February 2024**: $7,500,000 over five years, "OpenAlex: a free index for the world's research"

That third grant runs to early 2029. It is the single most important fact about OpenAlex's durability, and it is also the reason the API changed this year.

## The API just became metered, and most write-ups are out of date

If you read almost anything published about OpenAlex before 2026, you will see some version of "100,000 calls a day, no key required, just put your email in a mailto parameter." That is no longer how it works.

The current model, from the [authentication and pricing guide](https://developers.openalex.org/guides/authentication):

> OpenAlex data is and will remain available at no cost. Our data snapshot is totally free for bulk download. The API is a freemium service with free daily usage, $0.10/day with no key, or 10× that ($1/day) with a free API key, and after that you pay for what you use. We sell services, not data.

Calls now carry a published price by type, and I confirmed each one against the live response headers:

| Call type | Documented cost | Credits I was charged |
|---|---|---|
| Single entity, e.g. `/works/W2741809807` | Free | 0 |
| List or filter, e.g. `/works?filter=...` | $0.0001 | 1 |
| Search, e.g. `/works?search=...` | $0.001 | 10 |
| Content download (PDF) | $0.01 | not tested |

Every response carries the meter in its headers. Unauthenticated, I saw `X-RateLimit-Limit: 1000`, `X-RateLimit-Limit-USD: 0.1`, and a `X-RateLimit-Remaining` that dropped by 1 on a filter call and by 10 on a search. That is a much better designed rate limit than most, because you can watch it rather than guess at it.

What the free key buys you per day, from the docs:

- Single-entity lookups: unlimited
- List and filter calls: 10,000 calls, up to 1,000,000 results
- Searches: 1,000 calls, up to 100,000 results
- Content downloads: 100 PDFs

For the overwhelming majority of real work, that is not a constraint. Ten thousand filter calls at 100 results each is a million records a day, free. But if you were running an unattended pipeline against the old limits with no key, it is now living on a tenth of that budget, and you should go get a key at `openalex.org/settings/api`.

One inconsistency worth flagging: the machine-readable summary at `developers.openalex.org/llms.txt` says the no-key limit is "$0.01/day." The human documentation says $0.10, and the live headers said `0.1`. Trust the headers.

## The bulk snapshot is still completely free

This is the part that did not change, and it is the part that matters most for anyone doing serious analysis.

The whole database sits in a public S3 bucket. No AWS account, no credentials, no requester-pays:

```bash
aws s3 sync "s3://openalex/data/jsonl" "openalex-snapshot" --no-sign-request
```

I confirmed the bucket lists anonymously and that `data/jsonl/` and `data/parquet/` both exist, each a complete copy of every entity. The licence file sitting in the bucket root is CC0 1.0 Universal, which is about as unambiguous as licensing gets.

Sizes, from the docs: the JSON Lines snapshot is around 330 GB compressed and roughly 1.6 TB decompressed. Pull both formats and you are over 660 GB of transfer. That transfer is free to you because the AWS Open Data program covers the egress, which the docs put at about $70 per download.

The catch is cadence. The free public snapshot is refreshed quarterly. Daily-refreshed snapshots and monthly updates need a paid plan, and the documentation does not publish a price for either. It says to contact sales. If your project needs fresh bulk data on a schedule, budget for a conversation rather than a number.

There is also an official CLI for pulling filtered subsets without taking the whole thing:

```bash
pip install openalex-official
openalex download --api-key YOUR_KEY --filter "publication_year:2024" --output ./results
```

And full text: PDFs and TEI XML are available for about 60 million works, at $0.01 per file, key required.

## Five queries that actually answer something

Everything below I ran. Counts are as of 9 August 2026.

**Where research is coming from.** Group 2025 works by author country:

```bash
curl "https://api.openalex.org/works?filter=publication_year:2025&group_by=authorships.countries"
```

China 1,440,258. United States 1,309,288. Japan 605,474. India 511,935. Indonesia 480,689. Whatever you think that means, it is a number you can now check yourself in one call rather than buying it in a report.

**How open the literature actually is.** Group 2025 works by open access status:

```bash
curl "https://api.openalex.org/works?filter=publication_year:2025&group_by=open_access.oa_status"
```

Of 15,060,496 works published in 2025: green 5,292,432, closed 4,794,078, gold 1,717,632, diamond 1,715,377, hybrid 1,030,954, bronze 510,023. So roughly 68 percent of 2025's output is open by some route. Try getting that figure out of Scopus without a licence.

**What language science is written in.** Group everything by language:

```bash
curl "https://api.openalex.org/works?group_by=language"
```

English 220,922,811 of 323,964,135. Japanese 12,730,111. German 11,508,362. Spanish 10,433,820. French 9,223,583. Portuguese 5,485,184. Chinese 5,058,555. That last number is the one that surprised me, and it is a decent argument for why specialist tools that go looking for non-English technical writing exist at all.

**One institution's entire footprint.** Resolve by ROR identifier rather than by name:

```bash
curl "https://api.openalex.org/institutions/ror:03vek6s52"
```

Harvard University, OpenAlex ID I136199984, 708,731 works, 145,505,007 citations.

**Anything, by any identifier you already have.** `/works/doi:10.1234/example`, `/works/pmid:29456894`, `/authors/https://orcid.org/0000-0001-6187-6610`, `/sources/issn:2167-8359`. The identifier translation layer alone justifies keeping the API in your toolkit.

One rule the documentation is emphatic about, and it is the correct rule: never filter on a display name. Resolve the name to an ID first, then filter on the ID. Names are ambiguous, IDs are not.

## What it gets wrong

This is where I want to be careful, because the honest answer is not "nothing" and OpenAlex does not pretend otherwise.

**Author disambiguation has been genuinely rough, and is being actively repaired.** The snapshot release notes are unusually candid. The 22 May 2026 release describes "continued cleanup of overmerged authors: 2.6M splinter profiles merged across 7 phases, affecting ~5M works." The 25 June 2026 release describes overhauling corresponding-author detection and "raising accuracy from ~60% to ~90%." An author profile you pulled last year may not match the one you pull today, and for a stretch, roughly four in ten corresponding-author flags were wrong. If you built a metric on that field, rebuild it.

**The year buckets are not what you think.** Group all works by publication year and 2026 comes back with 28,947,494, nearly triple 2024's 10,781,357. That is not a publishing boom. Break it down by type and 16,107,050 of those 2026 records are datasets, against 553,097 datasets in 2024. A large tranche of dataset records has landed with recent dates. If you are plotting research output over time, filter to `type:article` or you will publish a chart of an ingestion event.

**Deprecated fields will silently change your results.** `host_venue`, `alternate_host_venues`, `grants` and the `has_ngrams` filter have all been removed outright. Concepts still work but are unmaintained and being replaced by topics, which are a much smaller and cleaner set: about 4,500 topics in a four-level hierarchy of 4 domains, 26 fields and 254 subfields, versus roughly 65,000 legacy concepts inherited from Microsoft. The docs say 254 subfields and the live `/subfields` endpoint returned 252 when I checked, so the taxonomy is not quite frozen either. Code written against a 2023 tutorial will either error or quietly return something different.

**Upstream errors flow downstream.** OpenAlex assembles from Crossref, DOAJ, ORCID, Unpaywall, PubMed and repository feeds. Bad metadata entered upstream arrives here too. That is a structural property of any aggregator, not a bug in this one, but it means OpenAlex is not a source of truth about any single record. It is a very good map, and you still check the territory before you cite it.

**Basic paging stops at 10,000 results.** Past that you need cursor pagination. `per_page` maxes at 100, `sample` at 10,000, and an OR filter takes at most 100 values.

## Who is actually using it

The adoption story is not hypothetical, and the strongest example is documented by the institution itself rather than by OpenAlex.

Sorbonne University [announced in December 2023](https://www.sorbonne-universite.fr/en/news/sorbonne-university-unsubscribes-web-science) that it was discontinuing its subscription to Web of Science and Clarivate's bibliometric tools in 2024. From the announcement:

> Sorbonne University has therefore decided to discontinue its subscription to the Web of Science database and Clarivate's bibliometric tools in 2024. [...] This decision is in line with the University's overall policy of openness, and it is now working to consolidate a sustainable, international alternative, in particular by using OpenAlex.

The same page notes that Leiden's CWTS, which produces one of the more widely watched university rankings, is working on a fully transparent version of its rankings based on open data from Crossref and OpenAlex.

There is a second, quieter kind of adoption worth knowing about. Georgetown's Center for Security and Emerging Technology publishes a dataset of its own classifier outputs keyed directly to OpenAlex work IDs, [on Zenodo](https://zenodo.org/records/20370823), currently at version 5.17.0 and about 2.8 GB. When a policy research shop builds its public artefact on your identifiers rather than its own, that is a real vote about which identifier system it expects to still exist in ten years.

## The thing worth sitting with

An index of 324 million scholarly works, updated continuously, given away under CC0, with a bulk download that costs the user nothing because a cloud provider absorbed the bandwidth, produced by an organisation whose entire annual expenditure would not cover a mid-sized company's software licences.

It is not charity and it is not permanent. It runs on a five-year grant that ends in 2029, plus whatever the paid tiers bring in, which is exactly why the API started metering. "We sell services, not data" is a sustainability strategy, stated out loud, and it is a more honest one than most.

If you use it, the useful things you can do are: get a free key so your traffic is attributable, keep a snapshot copy of whatever your work depends on, and cite it. CC0 does not require attribution. Infrastructure that survives generally does.

## Fact-check notes and sources

All entity counts and query results were measured against the live API on 9 August 2026 and will have moved since.

- **323,964,135 works, 124,697,641 authors, 255,280 sources, 134,448 institutions, 4,516 topics, 10,703 publishers, 45,639 funders**: measured via `https://api.openalex.org/<entity>?per-page=1` and reading `meta.count`
- **Pricing model, the $0.10 and $1 per day figures, and the free daily allowance table**: [OpenAlex Developers, Authentication and Pricing](https://developers.openalex.org/guides/authentication)
- **Per-call prices by endpoint type**: [OpenAlex LLM quick reference](https://developers.openalex.org/llms.txt), cross-checked against the `X-RateLimit-Cost-USD` and `X-RateLimit-Credits-Used` headers returned on live calls
- **Snapshot size, quarterly free cadence, paid daily refresh, AWS Open Data covering roughly $70 of transfer per download, and the `--no-sign-request` command**: [OpenAlex Developers, Data Downloads Overview](https://developers.openalex.org/download/overview) and [Download to your machine](https://developers.openalex.org/download/download-to-machine)
- **CC0 1.0 Universal licence**: `LICENSE.txt` in the root of the public `openalex` S3 bucket
- **Author overmerge cleanup and corresponding-author accuracy going from about 60 percent to about 90 percent**: `RELEASE_NOTES.txt` in the same bucket, releases dated 22 May 2026 and 25 June 2026
- **Deprecated and removed fields, and the topics versus concepts comparison**: [OpenAlex Developers, Deprecations](https://developers.openalex.org/guides/deprecations) and [Key Concepts](https://developers.openalex.org/guides/key-concepts)
- **Launch history, 209 million works at launch, and the founding authors**: [Priem, Piwowar and Orr, arXiv:2205.01833](https://arxiv.org/abs/2205.01833)
- **Microsoft Academic Graph stopping updates on 31 December 2021, the January 2022 launch, and the Library of Alexandria naming**: [OpenAlex, Wikipedia](https://en.wikipedia.org/wiki/OpenAlex)
- **OurResearch legal name Impactstory, EIN 46-1599252, and the revenue, expense, contribution and asset figures for tax years 2022 through 2025**: IRS Form 990 data via [ProPublica Nonprofit Explorer](https://projects.propublica.org/nonprofits/organizations/461599252)
- **The three Arcadia grants, their dates, amounts and terms**: Arcadia's own 360Giving grant file, published under CC0 and linked from [Arcadia's grants awarded page](https://arcadiafund.org.uk/grants-awarded), records 360G-ArcadiaFund-4087, 4558 and 5147
- **Sorbonne University dropping Web of Science and adopting OpenAlex, and the reference to Leiden CWTS**: [Sorbonne University unsubscribes from the Web of Science](https://www.sorbonne-universite.fr/en/news/sorbonne-university-unsubscribes-web-science), published 8 December 2023
- **CSET metadata over OpenAlex works, version 5.17.0**: [Zenodo record 20370823](https://zenodo.org/records/20370823)

The 2026 dataset-record anomaly, the `llms.txt` inconsistency on the no-key daily limit, and the per-call credit costs observed in response headers are my own measurements from August 2026, not OpenAlex statements.

## Related reading

- [Harvard Dataverse: 302,000 datasets, no account needed](/blog/harvard-dataverse-research-data-repository/)
- [GDELT: the quarter-billion-record open dataset that lets a small newsroom watch the whole world](/blog/gdelt-open-data-kalev-leetaru/)
- [What Hugging Face actually is: the hub, the libraries, and the $0.10 free tier](/blog/hugging-face-explained-how-to-use/)
- [How to diversify your AI API spend before your vendor cuts you off](/blog/diversify-ai-api-spend-across-providers/)


---

Canonical HTML: https://jwatte.com/blog/openalex-open-index-of-research/
RSS: https://jwatte.com/feed.xml
JSON Feed: https://jwatte.com/feed.json
Hero image: https://jwatte.com/images/blog-openalex-open-index-of-research.webp