Sources
Every number and statement on AGI Meridian comes from the organisations below. Each name links to the original.
Open data
Free to reuse with credit. Our charts redraw their numbers.
| Organisation | What we use | Licence |
|---|---|---|
| Epoch AI | Capabilities Index, MirrorCode, FrontierMath, GPQA Diamond runs, SWE-bench Verified results, Furniture Assembly, EBR-bench | CC BY 4.0 |
| Hendrycks et al. (Center for AI Safety and co-authors) | "A Definition of AGI": ten-ability scores | CC BY 4.0 |
| Column Tax (TaxCalcBench) | Tax return accuracy (TaxCalcBench results) | MIT |
| PostTrainBench team | AI training other AI models (PostTrainBench results) | MIT |
| Thinking Machines (MMAU run) and MMAU authors | Hearing scores (speech, sounds and music) | CC BY 4.0 |
| Stack Overflow | Developer Survey 2025 and 2026 | ODbL 1.0 |
| International Energy Agency | Share of the world's electricity used by data centers, and their growth in 2025 | CC BY 4.0 |
Epoch AI data is licensed under CC BY 4.0, as are the Hendrycks et al. paper and MMAU. PostTrainBench is MIT. The Stack Overflow Developer Survey is under ODbL 1.0. We turn their numbers into charts and change nothing else.
OpenAI funded FrontierMath and has access to part of it (Epoch AI's disclosure).
Shown with credit
These publishers keep all rights. We show their numbers with credit and a link, and offer no download or embed.
| Organisation | What we use | Terms |
|---|---|---|
| Vals AI | AI research scores (RSI Index) and IOI programming olympiad scores. Vals AI owns the data. | © Vals AI. All rights reserved. |
| METR | Coding job length (time horizon) | © METR. All rights reserved. |
| Scale AI and Center for AI Safety | Remote Labor Index | © Scale AI. All rights reserved. |
| Anthropic and Google | Their own code-share and AI research work figures, linked on each page | Their own figures, linked to the original. |
| Stanford Digital Economy Lab | Employment of young workers in AI-exposed jobs | Used with credit. |
| ARC Prize | ARC-AGI-2 and ARC-AGI-3 scores | Used with credit. |
| Clay Mathematics Institute | Millennium Prize problems | Used with credit. |
| Hugging Face | AI Energy Score: energy used by AI that thinks before answering | Used with credit. |
| Central Statistics Office, Ireland | Data centers' share of Ireland's electricity | Used with credit. |
| PJM Independent Market Monitor | Cost of reserving power in the US's biggest grid that came from data centers | Used with credit. |
| ERCOT | Texas data center connection requests and the pause on new hook-ups | Used with credit. |
| Artificial Analysis | Humanity's Last Exam and GDPval-AA scores, on the US vs China page | Credited to Artificial Analysis, as its free data terms require. |
| Terminal-Bench | Terminal-Bench 4.0 scores, on the US vs China page | Used with credit. |
| Zapier | AutomationBench scores, on the US vs China page | Used with credit. |
| Datacurve | DeepSWE scores, on the US vs China page | Used with credit. |
| UC Berkeley RDI | Agents' Last Exam scores, on the US vs China page | Used with credit. |
| Mercor | FrontierCode scores (made by Cognition, hosted by Mercor) and the month-end accounting test | Used with credit. |
| SecureBio | Virology Capabilities Test: working out why a virus experiment failed | Used with credit. |
| Samaya AI | Forecasting company revenue from earnings reports | Used with credit. |
| Handshake AI | Building banking work ready for a client | Used with credit. |
| Nubank | Handling routine card and credit-limit questions, and helping customers sort out debt | Used with credit. |
| CRASH Lab | Naming the diagnosis from one scan (RADLE) | Used with credit. |
| Andon Labs | Drawing a floor plan from photos (Blueprint-Bench) | Used with credit. |
| Vectara | Hallucination leaderboard: how often AI adds claims a document does not contain | Used with credit. |
| WildHandBench | Reading messy handwriting | Used with credit. |
| RoboDojo | Doing chores with a robot | Used with credit. |
| Forecasting Research Institute | ForecastBench: predicting world events | Used with credit. |
| Waymo | Injury crashes, Waymo against human drivers | Used with credit. |
| Palisade Research | Shutdown tests: how often models got around being turned off | Used with credit. |
| Apollo Research | Tests reported in OpenAI's GPT-6 Astra report | Used with credit. |
| Frontier Security | Report on a Chinese model and the UK AI Security Institute's benchmark | Used with credit. |
| International Olympiad in Informatics | IOI 2026 results of the human contestants | Used with credit. |
| University and company studies on arXiv | Translating a novel, writing in an author's style (Stony Brook, Columbia and Michigan) and keeping shoppers happy in chat (Taobao), each linked to its paper | Used with credit. |
| Alek Safar | ClockBench | Used with credit. |
| AI Explained team | SimpleBench | Used with credit. |
Statements and reports
Organisations describing their own work, or testing other labs' models. We quote short passages, name the speaker and link the original.
- Anthropic
- OpenAI
- NSA, CISA and FBI
- Google and Alphabet
- xAI
- Alibaba
- DeepSeek
- Google DeepMind
- UK AI Security Institute
- International AI Safety Report
- ARC Prize Foundation
- Government and regulator releases, linked on each page that uses them
Staircase: redrawn after Tim Urban (Wait But Why, 2015), Eliezer Yudkowsky (2008) and Nick Bostrom (Superintelligence, 2014). Their art is theirs; ours is a new drawing.