In March 2026, OpenAI operationalized1 the OpenAI Foundation and Anthropic launched the Anthropic Institute, followed by large-scale financial commitments from both sides:
The OpenAI Foundation, which holds a 26% equity stake in OpenAI Group, has already committed $530 million to causes like AI resilience, science, or economic transition with an expected investment of $1 billion this year and a $25 billion overarching commitment for the years ahead.
Anthropic has institutionalized its public benefit-oriented research through launching the Anthropic Institute and committing $654 million across initiatives such as a partnership with the Gates Foundation or Claude Corps, led by its Beneficial Deployments team.
We use the term “redistribution“ deliberately: the initiatives observed here transfer value after it has already been captured by the labs, rather than restructuring who owns or governs it at the point of creation. We use “philanthropic“ to describe financial commitments explicitly framed for public benefit and directed toward non-commercial beneficiaries, distinguishing them from AI labs’ standard commercial operations.
Recognizing that these programmes are newly established and still taking shape, this post serves two purposes: documenting a snapshot of what is currently visible in the public record of OpenAI Foundation and Anthropic, and surfacing the distributive choices already embedded in the design of these initiatives.
Mapping the redistribution
All data was collected between June 17, 2026 and July 9, 2026.
Inclusion criteria for the snapshot:
Funder: For OpenAI, we focus on commitments made by or through the OpenAI Foundation. For Anthropic, no equivalent legally separate entity exists, e.g. the Anthropic Institute and the Beneficial Deployments team operate within the for-profit entity. We therefore include Anthropic financial commitments that are explicitly framed as public benefit or charitable in purpose. Excluded: initiatives where OpenAI Foundation or Anthropic is named as a co-funder but is not the primary funder or initiating party, as the labs’ specific contribution and decision-making role could not be verified from public documentation.
Financial cash commitment to non-commercial beneficiaries: The initiative must involve a financial commitment where the primary stated purpose is public benefit or charitable impact. This criterion is satisfied if either: (a) the primary documentation uses language such as ‘public benefit,’ ‘charitable,’ ‘nonprofit,’ ‘social impact,’ or substantively equivalent terms; or (b) the primary documentation explicitly designates a non-commercial entity (e.g. such as a nonprofit organisation, academic institution, government body, or the general public) as the named primary recipient of the commitment.
Timing: The initiative must have been publicly announced between March 24, 2026 and July 9, 2026, the close of the data collection period. March 24th marks the OpenAI Foundation’s first public announcement as an operational grantmaking entity, that is, the moment it shared its programmatic priorities and overarching financial commitments. The Anthropic Institute was launched on March 11, 2026, but as an internal research arm rather than a disbursing entity. No Anthropic initiatives meeting the inclusion criteria were found that were announced between March 11 and March 24, making the two-week difference negligible for the data collection2. We therefore use March 24 as a single consistent snapshot boundary for both labs.
Public documentation: The commitment must be verifiable, requiring at least one publicly accessible primary source directly from OpenAI Foundation (www.openaifoundation.org) or Anthropic (www.anthropic.com) websites.
We note that this dataset captures a moment in time rather than a complete historical record; earlier or concurrent commitments not captured in public documentation may exist.
Findings
Because this dataset captures only a specific moment in time rather than a complete historical record, the following findings reflect the OpenAI Foundation’s and Anthropic’s current structural priorities rather than the full evolution of the labs’ distributive practices.
Figure 1: Stated total value by initiative. Data collected between June 17 – July 9, 2026. Figures represent announced commitments, not disbursements.
Sectors and themes
We identified three key clusters in the thematic map of the of the snapshot:
The first and largest is AI transition preparedness: this includes research to understand and measure, and initiatives to mitigate the societal disruptions AI will cause. For example, OpenAI’s $250M Economic Futures in the Age of AI programme funds labour market measurement and worker transition support; and Anthropic’s $200M Economic Futures Research Fund focuses on income support models and fiscal policy for a post-AI economy.
The second cluster is science and health R&D, including OpenAI’s $100M AI for Alzheimer’s initiative and Anthropic’s Gates Foundation partnership for neglected diseases and agricultural productivity.
The third, smaller cluster is cybersecurity and critical infrastructure, featuring Anthropic’s Project Glasswing, framed explicitly as defensive public-good infrastructure rather than commercial security provision.
Benefit types
Cash, featured in the inclusion criteria for the philanthropic initiatives in this dataset, is committed through different mechanisms, such as grants (e.g. AI Resilience Programme), donations (e.g. Project Glasswing), or salaries (e.g. Claude Corps fellows). Unrestricted cash grants, however, appear only in one initiative: OpenAI’s People-First AI Fund, which gives up to 10% of an organisation’s annual budget. It is also the smallest programme by total commitment ($50M) in the snapshot and the most geographically restricted (i.e. US 501(c)(3) only).
Beyond cash, there are two key accompanying benefit groups:
Proprietary model access and credits: predominantly in Anthropic’s initiatives, e.g. model access in Project Glasswing or $2,500 in Claude licenses and API credits per fellow in Claude Corps.
Infrastructure and ecosystem support: e.g. Claude Corps builds human capital infrastructure by upskilling the fellows and their respective host organisations; AI for Alzheimer’s supports the creation of a shared data infrastructure to enable the wider scientific community to chart disease progression and predict drug activity.
Relatedly, cases that were excluded from the dataset, yet signal another form infrastructure support, included OpenAI and Anthropic acting as one of the founding members of new civil society vehicles for societal AI preparedness: Anthropic was a founding member of RAISE US (theme: American workforce transition) and OpenAI of the Appia Foundation (theme: shared standards for AI). Anthropic and OpenAI Foundation also co-funded Intercept which aims to eradicate respiratory diseases, advancing a non-AI public good.
Beneficiaries and geographic focus
Of the eight initiatives in the dataset:
two have explicit US-only eligibility requirements that include 501(c)(3) status, US-based operations, and US work authorisation (People-First AI Fund, Claude Corps)
two have confirmed grantees before the public launch, offering no open application process (AI for Alzheimer’s: six named US research institutions; Project Glasswing named 12 US-headquartered corporations as launch partners).
the Anthropic-Gates partnership is the most LMICs-facing, including beneficiaries in sub-Saharan Africa, and India alongside the US; the initiative doesn’t offer a public application pathway, emphasising instead the Gates Foundation’s long-standing experience
the specific beneficiaries are still to be announced for the remaining three programmes (Economic Futures Research Fund, AI Resilience, Economic Futures in the Age of AI).
Observations
Where eligibility is defined, it is frequently restricted to the United States
Where an initiative specifies who is eligible to apply, the criteria often limit participation to the United States: the People-First AI Fund is open only to US-based 501(c)(3) organisations that operate primarily within the 50 states or the District of Columbia; Claude Corps requires fellows to be authorised to work in the US and host organisations to be US-based nonprofits; and AI for Alzheimer’s names six US research institutions as its grantees.
However, there are a few caveats we want to highlight. Several of the largest commitments, including OpenAI Foundation’s Economic Futures in the Age of AI or Anthropic’s Economic Futures Research Fund, had not named beneficiaries by the close of data collection, and Anthropic describes its fund as intended to be globally applicable. Moreover, two initiatives reach beyond the US by design. First is the Gates Foundation partnership described below and second is Project Glasswing, whose US-headquartered launch partners were joined in early June 2026 by roughly 150 further organisations across more than fifteen countries. The pattern we observe, therefore, is that US-restricted eligibility is a common characteristic in the pilot grant and fellowship programmes, yet the initiatives without stated eligibility criteria, which include some of the largest commitments, leave the geographic question open.
Support aimed at lower- and middle-income countries runs through an intermediary global philanthropy
One initiative in the snapshot is designed primarily for lower- and middle-income countries: Anthropic’s four-year, $200M partnership with the Gates Foundation. Its largest component addresses health in these countries, including work on polio, HPV, and preeclampsia, and forecasting for malaria and tuberculosis through the Foundation’s Institute for Disease Modeling, with education work extending to sub-Saharan Africa and India, and economic-mobility work aimed at smallholder farming.
The partnership reaches recipients through the Gates Foundation’s existing programmes and delivery networks rather than through an open application process, so who benefits in these countries appears to be determined by the Foundation rather than by an open call. At the same time, the partners state that the datasets, benchmarks, and tools the work produces will be released as public goods and made freely available, and the Gates Foundation frames the effort around closing the resource gap with equity as its stated goal. Whether routing support through an established global philanthropy best serves recipients’ own priorities, such as local capability-building and data ownership, is an important question for these programmes to answer as they mature, and one this snapshot cannot settle.
Both labs are funding macro level societal scaffolding for AI transition
At the broadest level, several commitments fund research and policy work on how societies should adapt to AI – work usually associated with governments and public institutions, here supported by the companies building the technology. OpenAI Foundation and Anthropic both fund study of labour-market disruption and income-support models. Anthropic pairs its fund with an Economic Policy Framework that opens by arguing that the “policymaking process was built for a slower world” and recommends measures such as stronger labour-market statistics and modernised unemployment insurance that can scale quickly. Claude Corps, on the other hand, directly pilots what an AI-adapted society may look like amid vast, AI-driven economic change.
What the snapshot shows is that AI developers are not necessarily standing in for governments but are actively shaping the terms on which governments may eventually act. They do this by generating evidence and authoring some of the proposals that may shape public policy responses. The agenda for societal adaptation to AI is therefore being partly set by the companies whose technology is driving the need for adaptation.
Conclusion
This snapshot documents what the institutionalisation of AI-driven redistribution currently looks like at the two leading frontier AI labs. Three limitations bound what can be concluded from the data: we are looking at a highly concentrated and narrow sample of only two American labs; we capture a brief moment in time rather than a complete historical record; and we base this initial analysis entirely on public announcements, which are inherently curated and frequently lack full operational details.
However, many of these initiatives are explicitly described as pilots, with stated ambitions to grow in funding, scope, and reach. As they expand, the question of who benefits, on what terms, and who is excluded become increasingly consequential. This is why, we end this post with further research directions, emerging directly from the limitations and open questions of this snapshot:
Further research directions
Deepening and extending this dataset
Longitudinal tracking: Collecting data on AI benefits sharing by OpenAI and Anthropic prior to March 2026 would allow observation of how their practices have evolved over time. Monitoring future announcements will also continue adding to the dataset, allowing researchers to map the trajectory of these pilot programs as they mature and scale.
Beneficiary tracing: The current dataset captures announced commitments, not disbursements or outcomes. Analysing how benefits diffuse beyond the primary named beneficiaries would add an impact layer this snapshot cannot provide.
IP and data terms: Across all initiatives in this snapshot, IP and data ownership terms are either not stated or ambiguous in publicly available documentation. Understanding who owns the outputs enabled by these commitments would shed direct light on the degree of capacity-building versus dependency these initiatives produce.
Other labs and geopolitical dynamics: This snapshot covers only OpenAI Foundation and Anthropic. Extending the analysis to other frontier labs and examining the geopolitical dynamics shaping their respective distributive approaches would situate the current findings in a broader competitive context.
Expanding the scope of analysis
Geographic scaling: Given the explicit plans for some of these pilots to grow in reach, a productive next step would be examining which programme models are transferable to other regions and which are not, e.g. what conditions made Claude Corps or the People-First AI Fund viable in the United States, and what structural adaptations would be required to run equivalent programmes in sub-Saharan Africa, South Asia, or Latin America.
Managing the public benefit and commercial incentive tension: Several initiatives in this dataset simultaneously advance the labs’ product ecosystems and stated public benefit goals. For example, OpenAI Foundation’s Alzheimer’s initiative explicitly frames an “active learning cycle” using patient data to improve AI models. The distinction between philanthropy and market development may therefore not be fully separable in some initiatives, raising questions about optionality and recipient autonomy. Future research could examine what governance mechanisms could make these dual-interest arrangements more transparently and equitably structured.
Dependency and structural lock-in: Related to the above, future iterations of these programmes could address the risk of structural dependency more directly. Programmes that distribute proprietary API credits or embed proprietary platforms as the primary benefit mechanism create reliance on continued lab provision. This is not inherently problematic in the short term, but it limits recipient optionality over time. Further research on future reiterations could explore what a more capacity-building-oriented version of the programmes would look like, e.g. one that builds skills, infrastructure, or outputs that retain value independently of any single platform.
Individual-level redistribution: The current analysis focuses on institutional commitments. A separate and complementary direction would examine redistribution from individuals whose wealth is AI-driven (e.g. employees and founders of frontier labs).
Theoretical positioning
The initiatives documented here sit at the intersection of at least three existing literatures: AI benefit-sharing, corporate social responsibility, and technology diffusion, each of which carries different analytical implications. Future work that explicitly tests these framings against empirical data, including longitudinal data as these programmes mature, would sharpen the analytical tools available to researchers, policymakers, and the labs themselves.
Acknowledgements: We thank Jacob Shaal for comments on an early draft of this post.
Appendix
Data collection categories:
Table 1: Data collection categories for the snapshot of OpenAI’s and Anthropic’s AI-driven redistribution.
Missing-data codes
Not stated - Information not provided in any publicly available primary source at time of data collection.
TBC (To be confirmed) - Information not yet available at time of data collection but publicly confirmed as forthcoming by the funder (e.g. Economic Futures Research Fund: “We’ll have more to share soon”).
Borderline inclusion case
Project Glasswing (Anthropic) - The $100M in usage credits to critical infrastructure organisations (including large commercial entities such as AWS, Apple, and Microsoft) is the borderline component. This initiative is included on the grounds that (1) Anthropic frames it explicitly as public-good-oriented societal resilience infrastructure and (2) the other $4M in direct donations to Alpha-Omega/OpenSSF ($2.5M) and the Apache Foundation ($1.5M) straightforwardly meet all four inclusion criteria.
Full dataset
Table 2: AI-driven redistribution: a snapshot of current commitments by OpenAI Foundation and Anthropic (data collected June 17, 2026 - July 9, 2026).
Note: OpenAI Foundation was announced after recapitalization in October 2025, but the first public announcement “Update on the OpenAI Foundation” on the “OpenAI Foundation” website was released in March 2026.
On March 12, 2026, Anthropic announced Claude Partner Network, a commercial partner investment, which does not meet the “financial commitment to non-commercial beneficiaries” criterion.