“Find published statistics on the share of usage or adoption of open-weight models versus proprietary models since ChatGPT launched. For each figure, give me the number claimed, exactly what was measured (downloads, tokens, deployments, survey responses), the time period, who published it, and the source URL.”
Fifteen published figures from seven primary publishers, spanning August 2023 through May 2026 fieldwork (ChatGPT launched November 30, 2022). Each figure is a percentage — but of different things: enterprise workload share, organization-level adoption where firms use both kinds, production deployment among adopters, weekly token volume on one routing platform, and stated intent. The lanes below keep them apart; unlike metrics are never joined into one trend.
| Period measured | Number claimed | Metric type | What was measured & denominator | Can categories overlap | Publisher | Sample / data scope | Source |
|---|---|---|---|---|---|---|---|
| Aug 2023 | nearly 77% | Intent / non-use | Respondents who don't use or don't plan to use commercial LLMs beyond prototypes in production | n/a (single question) | Predibase | 150 executives, data scientists, ML engineers across 29 countries | insidehpc.com |
| Nov 2023 | upwards of 85% closed | Market / usage share (exclusive) | Share of models in production that are proprietary | No — mutually exclusive | Menlo Ventures | 450+ enterprise executives, US & Europe | menlovc.com |
| Feb 2023 – Mar 2024 | 76% open | Adoption incidence (non-exclusive) | Companies using LLMs that chose at least one open-source model, often alongside proprietary models | Yes — orgs can use both | Databricks | Aggregated, anonymized platform data, 10,000+ global customers incl. 300+ Fortune 500 | databricks.com |
| 2024 | 81% closed / 19% open | Market / usage share (exclusive) | Enterprise usage market share of closed vs open-source solutions | No — mutually exclusive | Menlo Ventures | 600 US IT decision-makers, firms with 50+ employees | menlovc.com |
| Reported 2025 | 63% all orgs / 72% tech industry | Adoption incidence (non-exclusive) | Respondents whose organization used an open-source AI model | Yes — open use doesn't exclude closed | McKinsey, Mozilla Foundation, Patrick J. McGovern Foundation | 700+ technology leaders and senior developers, 41 countries | mckinsey.com |
| Jun 30 – Jul 10, 2025 | 13% open (from 19% six months earlier) | Market / usage share (exclusive) | Share of AI workloads using open-source models | No — workload share | Menlo Ventures | 150 technical decision-makers at enterprises and startups building AI apps | menlovc.com |
| 13 months ending Nov 2025 | Chinese OSS 13.0% avg (1.2% weekly low, nearly 30% in some weeks) | Token telemetry (exclusive shares) | Average weekly share of token volume on OpenRouter, 100 trillion tokens; 'OSS' = open-weight | No — token shares sum to 100% | OpenRouter with a16z | OpenRouter platform routing telemetry | openrouter.ai |
| 13 months ending Nov 2025 | Rest-of-world OSS 13.7% avg / rest-of-world proprietary 70% avg | Token telemetry (exclusive shares) | Average weekly share of token volume on OpenRouter | No — token shares sum to 100% | OpenRouter with a16z | OpenRouter platform routing telemetry | openrouter.ai |
| Late 2025 | 11% open (from 19% prior year) | Market / usage share (exclusive) | Overall enterprise open-source share | No — mutually exclusive | Menlo Ventures | ~500 US enterprise decision-makers | menlovc.com |
| May 19–29, 2026 | 79% open / 71% closed (50% both, 29% open only, 21% closed only) | Adoption incidence (non-exclusive) | Developers adding AI functionality who currently use each model type | Yes — 50% use both | Mozilla / SlashData | 1,494 qualified developers | stateofopensource.ai |
| May 19–29, 2026 | 51% open adopters / 63% closed adopters | Production deployment (per adopter group) | Adopters of each model type who reported reaching production | Denominators differ per group | Mozilla / SlashData | 1,494 qualified developers (adopter subsets) | blog.mozilla.org |
Method: figures compiled from primary publisher pages between November 30, 2022 (ChatGPT launch) and September 4, 2026; 15 plotted points from 11 distinct figure sets across 7 publishers. Publishers use “open source”, “OSS” and “open-weight” inconsistently — OpenRouter explicitly uses OSS as shorthand for open-weight; survey definitions vary. Percentages measure whatever each row's denominator states; non-exclusive adoption figures were never converted into complements, and proprietary share was never inferred from open-only statistics. Chinese and rest-of-world OSS token shares are kept separate as published. The search is broad but not guaranteed exhaustive; Menlo's enterprise estimates are the strongest comparable series, while platform telemetry and developer surveys answer different questions. Model-count shares, benchmark shares and raw download totals were excluded. The Predibase figure is stated intent, not measured usage.