Inputs Library

Reuters reported that The Information said OpenAI topped $25B in annualized revenue at the end of February 2026; Reuters noted that it could not verify the report.

GPT-4.1 mini input pricing is listed at $0.40 per 1M tokens.

GPT-4o input pricing is listed at $2.50 per 1M tokens.

Microsoft said in April 2026 that its AI business had surpassed a $37B annual revenue run rate.

Anthropic said in April 2026 that its run-rate revenue had surpassed $30B, up from about $9B at the end of 2025.

CoreWeave reported $2.078B of revenue for Q1 2026; this input annualizes that quarter to $8.312B.

GPT-4.1 mini output pricing is listed at $1.60 per 1M tokens.

GPT-4o output pricing is listed at $10.00 per 1M tokens.

Reported as worth more than $250M over five years; stored here as a round-number benchmark.

A $500k prize pool spread across 2,500 final public questions implies about $200 per retained question, excluding organizer and reviewer labor.

Scale reports that HLE contributors competed for a $500,000 prize pool, with $500 awards for the next 500 questions after the top 50.

BLS reports a mean hourly wage of $122.99 for family medicine physicians in May 2025.

BLS reports a mean hourly wage of $128.46 for general internal medicine physicians in May 2025.

Informa disclosed a $10M+ initial fee plus recurring payments; the stored value is a conservative floor, not the full contract total.

Uses the high end of EFA's book-proposal per-word rate range as a conservative paid-writing proxy.

Uses a low-end professional editorial benchmark as a rough floor for paid per-word labor.

Reuters reported the Google-Reddit licensing contract was worth about $60M per year.

OpenAI's English rule of thumb is that one token is about three-quarters of a word.

Meta reports that Llama 3 was pretrained on about 15 trillion multilingual tokens.

Meta reports that Llama 4 Scout was trained on 40 trillion tokens.

AllenAI reports a 5.93T-token mix for the released OLMo 3 7B recipe.

The dataset card describes The Stack v2 as a roughly 900B-token code corpus.

DeepSeek reports that DeepSeek-V3 was pretrained on 14.8 trillion high-quality and diverse tokens.

Uses one workweek of annual oversight per participating physician as the default assumption.

An unsealed February 28, 2023 Meta email says Books3 plus Gutenberg is about 30B tokens.

The paper describes Common Pile as an 8TB openly licensed corpus.

The dataset card describes FineWeb2 as a 20TB corpus assembled from 96 Common Crawl snapshots.

The dataset card describes The Stack v2 as 67.53TB in total before deduplication.

Starts at 1 percent to reflect a deliberately macro-scale oversight regime rather than a small benchmark panel.

Uses one workday per physician reviewer as a default assumption for drafting examples, refining rubrics, and calibration.

The released Tulu 3 70B preference mixture contains 337,186 generation pairs.

The released Tulu 3 SFT mixture contains 939,344 examples.

The dataset card lists 463.6B raw tokens in the main stage and 176.2B raw tokens in the cooldown stage; this input sums both stages.

An unsealed appendix filed in February 2025 says 35.7TB of the downloads came from Z-Library and LibGen.

Uses 0.01 percent of physicians as a deliberately small specialty-network comparison against the larger default oversight regime.

Internal target for a hypothetical public-knowledge corpus, not an observed empirical statistic.

Judge Alsup wrote that Anthropic downloaded at least five million books from LibGen.

Judge Alsup wrote that Anthropic downloaded at least two million books from Pirate Library Mirror.

Judge Alsup wrote that Anthropic assembled a central library of more than seven million pirated books.

The dataset card describes FineWeb2 as containing more than 5 billion documents.

The June 2025 order identifies the Books3 dataset as containing 196,640 books.

Meta reports that Llama 4 Maverick was trained on 22 trillion tokens.

AllenAI reports 5 trillion training tokens for OLMo 2.

Alibaba says Qwen3-Coder was pretrained on 7.5 trillion tokens.

Alibaba says Qwen3 models were pretrained on about 36 trillion tokens across 119 languages and dialects.

Scale's April 3, 2025 update says HLE was finalized to 2,500 questions.

The GPQA paper and dataset card describe the main benchmark as 448 expert-written multiple-choice questions in biology, physics, and chemistry.

An unsealed appendix filed in February 2025 says Meta downloaded 81.7TB from shadow libraries.

The dataset card describes The Stack v2 as containing 3.28 billion unique files.

The dataset card describes FineWeb2 as containing more than 3 trillion words.

This is a rough historical benchmark for WSJ/Dow Jones newsroom size, not a freshly disclosed audited count.

WHO reports a global density of 17.2 doctors per 10,000 people in 2022. This input turns that density into a rough headcount by multiplying it by the repo's January 1, 2025 world-population benchmark.

News Corp reported about 23,900 employees as of June 30, 2024.

Uses the Census Bureau projection for the U.S. population on January 1, 2025.

Uses the Census Bureau projection for world population on January 1, 2025.

The paper describes Common Pile as drawing from 30 sources.

Uses the article count publicly shown on the Taylor & Francis platform homepage.

HealthBench gives the exact count of 262 physicians, and OpenAI's January 8, 2026 healthcare announcement separately describes a global network of more than 260 licensed physicians across 60 countries.

Uses Reddit's Q4 2025 daily active uniques as the latest official public scale benchmark.

Hypothetical productivity assumption for Stackipedia.

Derived from the English deduplicated counts reported by the project: 20.5T tokens over 14.5B documents.

This is an industry rule-of-thumb for a typical full-length book, not a census of Books3 itself.

Hypothetical default used for Stackipedia scenarios; no external public source is claimed.

Hypothetical engagement assumption for Stackipedia.

Aya contributes 100,000 multilingual instruction examples in the released mixture.

FLAN v2 contributes 89,982 examples in the released mixture.

Derived from the published PeS2o paper count in Dolma v1.6.

Derived from the Project Gutenberg token count in Dolma v1.6.

Derived from the published The Stack token count in Dolma v1.6.

Alibaba says 70 percent of Qwen3-Coder pretraining data was code.

Derived from the published Reddit token count in Dolma v1.6.

Derived from the Common Crawl and C4 token counts in Dolma v1.6.

Longpre et al. estimate that about 45% of C4 URLs are restricted for AI use by terms or robots exclusions.

This combines the five Persona subsets in the released Tulu 3 SFT mixture.

WildChat GPT-4 contributes 100,000 examples in the released mixture.

Meta reports 5.0 million H100-80GB GPU hours for Llama 4 Scout training.

The statute sets ordinary statutory damages at not less than $750 and not more than $30,000 per work, before willfulness or innocent-infringer adjustments.

DeepSeek reports 2.788 million H800 GPU hours for the full DeepSeek-V3 training run.

The October 2025 order says the parties identified 482,460 works for settlement administration.

DeepSeek reports 2.664 million H800 GPU hours for DeepSeek-V3 pretraining.

The October 17, 2025 filing describes an expected per-work award of about $3,000 across 482,460 identified works.

Meta reports 2.38 million H100-80GB GPU hours for Llama 4 Maverick training.