When you put an AI model into a business process, you inherit everything you do not know about it: what data it was trained on, how it was tested, what happens to your prompts. Stanford's Foundation Model Transparency Index measures exactly that, and its latest edition shows that transparency is going backwards: the average score fell from 58 to 41 out of 100. We analysed the public score data for Anthropic, OpenAI, Mistral, Google and nine other developers.
What the transparency index measures
The Foundation Model Transparency Index (FMTI) is run by Stanford's Center for Research on Foundation Models. Each developer is scored on 100 yes/no indicators about its flagship model, split into three domains: upstream (34 indicators: training data, data labour, compute, energy), model (30: architecture, capabilities, risks, mitigations, release) and downstream (36: usage data, economic impact, monitoring, usage policies, accountability).
The third edition was published in December 2025 and scored 13 companies. Seven submitted their own transparency reports (AI21 Labs, Amazon, Google, IBM, Meta, OpenAI, Writer). For the other six (Alibaba, Anthropic, DeepSeek, Midjourney, Mistral, xAI) the team gathered public information itself, helped by an AI agent built on Claude, and every finding was checked by hand. Only 30% of contacted companies agreed to submit a report, compared with 74% in 2024. All data is public on GitHub, which is what we used for the charts below.
The 2025 ranking: enterprise vendors lead, frontier labs sit in the middle
The scores split into three groups. Enterprise-focused vendors IBM (95), Writer (72) and AI21 Labs (66) lead by a wide margin. In the middle are the big labs: Anthropic 46, Google 41, Amazon 39, OpenAI 35, DeepSeek 32, Meta 31 and Alibaba 26. At the bottom are Mistral (18), Midjourney and xAI (14 each).
Source: Stanford CRFM, 2025 Foundation Model Transparency Index (December 2025), 100 indicators: 34 upstream, 30 model, 36 downstream. Domain sums from the public score file on GitHub (stanford-crfm/fmti).
The pattern in the chart is the main story. The dark segment (upstream: training data and compute) is almost invisible for the frontier labs: Anthropic gets 3 of 34 points, OpenAI 1, Mistral 0. They earn most of their score in the downstream domain, through usage policies, documentation and monitoring. IBM is the only company that meets every upstream indicator.
Three editions: rankings flipped
The FMTI is one of the few metrics for AI companies tracked over time. In 2023 the average was 37; in 2024 it rose to 58 as companies engaged with the index; in 2025 it fell back to 41. The 2025 edition introduced new and stricter indicators, so part of the decline reflects a higher bar, but the authors also document real regressions: companies that used to disclose information and no longer do.
Source: FMTI October 2023, May 2024 and December 2025. Mistral and IBM were first scored in 2024. The 2025 edition changed the indicators, so the year-on-year comparison is indicative.
- Meta was first among the six companies scored in all three editions in 2023 (54) and is now last of them (31), a drop of 29 points in a year.
- OpenAI went from second (48) to second-to-last (35), down 14 points from 2024.
- Anthropic rose from 36 in 2023 to 46 and is now second among the six, behind AI21 Labs (66). It lost 5 points from 2024.
- Mistral dropped from 55 to 18, the largest fall in the index. It did not submit a report in 2025, so only its public documentation for Mistral Medium 3 counted.
- IBM rose from 64 to 95, by preparing a detailed report and disclosing six indicators no other company discloses.
Where the gaps are: training data, compute and real-world usage
The heatmap shows which topics each lab is open about. The pattern is consistent: companies are fairly open about release practices, usage policies and safety mitigations, and almost silent about how the model was built and how it is used in practice.
Source: our calculation from the FMTI 2025 score file. Each cell = indicators met / indicators in that subdomain. Selected subdomains and companies.
Anthropic stands out on usage data (3 of 5 indicators; Anthropic also publishes the Anthropic Economic Index on how Claude is used) and on usage policies (5 of 5). OpenAI is strongest on post-deployment monitoring (5 of 7) and mitigations. Google and Amazon disclose some compute and data-processing details but nothing on basic model properties. Mistral and xAI score little anywhere outside release and usage policies.
Source: FMTI 2025 score file, number of the 13 scored companies that meet each indicator.
Every company publishes terms of use and a change log. Only two disclose the energy or water used to train their flagship model (IBM and Writer), and no company discloses whether its benchmark test data overlaps with its training data. That last point matters for anyone comparing models by benchmark scores: without it, you cannot rule out that a model has seen the test. See also our analysis of AI energy use and of copyright and training data lawsuits, which the index names as one reason for training-data secrecy.
Regulation is starting to require some of this
Part of what the index measures is now becoming law. Under the EU AI Act, obligations for general-purpose AI (GPAI) model providers have applied since 2 August 2025, and the European Commission's enforcement powers started on 2 August 2026. Providers must keep technical documentation, give downstream providers the information they need to integrate the model, have a copyright policy and publish a summary of the content used for training. Fines for GPAI providers can reach €15 million or 3% of worldwide turnover. In California, SB 53 has required large frontier developers to publish a frontier AI framework and transparency reports for new frontier models since 1 January 2026.
The index found that signatories of the EU GPAI Code of Practice score slightly higher than non-signatories, mostly thanks to downstream disclosures. Regulation will raise the floor on documentation, but it does not require the most useful information for buyers, such as data composition, train-test overlap or real usage statistics. For context on the wider rules, see our global AI regulation comparison and the frontier labs' safety frameworks.
What this means for companies buying AI
A low transparency score does not mean a model is bad or unsafe. Claude, GPT and Mistral models are among the most capable on the market. It means you have to ask for the information yourself and put it in the contract. Our checklist for vendor due diligence:
| Question for the vendor | Why it matters | Where to look |
|---|---|---|
| Are our prompts and outputs used for training, and how long are they kept? | GDPR, trade secrets | Commercial terms, data retention docs |
| Which evaluations were run, by whom, and are they reproducible? | Benchmark scores may not transfer to your task | System card, model card |
| What documentation do you provide to downstream providers under the AI Act? | You may need it for your own compliance | Art. 53 AI Act documentation |
| What is the training data cut-off and the published training data summary? | Copyright risk, knowledge gaps | Public training content summary |
| How are model versions retired and announced? | Stability of production systems | Deprecation policy, change log |
| Where is data processed, and by which sub-processors? | Data residency, GDPR transfers | DPA, sub-processor list |
Most importantly, build your own evaluation set from real tasks in your company. If vendors do not disclose train-test overlap, your own test data is the only benchmark you can fully trust. More in our guide to AI vendor selection and procurement.
Limits of the index
- It measures disclosure, not quality or safety. A company can be open about a weak model or quiet about a strong one.
- It scores one flagship model per company, and the models are from 2025 (Claude 4, o3, Gemini 2.5, Mistral Medium 3). Newer releases and system cards are not included.
- Effort matters. Companies that prepared their own reports scored considerably higher; Anthropic and Mistral did not, so their scores rely on public material only.
- The indicators changed in 2025, so the year-on-year decline is partly a stricter bar.
Frequently asked questions
What is the Foundation Model Transparency Index?
The FMTI is an annual index by Stanford CRFM that scores AI developers on 100 yes/no indicators about how their flagship model is built (data, compute), what it can do and risk (evaluations, mitigations) and how it is used (usage data, impact, policies). Editions: October 2023, May 2024 and December 2025.
How did Anthropic and OpenAI score?
In the December 2025 edition Anthropic (Claude 4) scored 46/100 and OpenAI (o3) 35/100. Both scored almost nothing on training data and compute (3 and 1 of 34 upstream points). The authors note that Anthropic discloses almost a strict superset of what OpenAI discloses; the exceptions are a bug bounty and a data retention and deletion policy.
Why did Mistral score so low?
Mistral fell from 55 in 2024 to 18 in 2025, the largest drop in the index. Mistral did not submit a transparency report in 2025, so the FMTI team scored its public documentation for Mistral Medium 3, an API-only model. It scored zero of the 34 upstream indicators. Mistral still meets some indicators others do not, such as publishing a model roadmap.
Is Manus included?
No. The 2025 index scores 13 developers of foundation models. Manus builds agents on top of models from other providers, so it is outside the scope of the index.
The bottom line
Transparency in AI fell in 2025, and the most capable labs disclose the least about training data and compute. Anthropic leads the frontier labs with 46/100, ahead of Google (41) and OpenAI (35), while Mistral dropped to 18. The EU AI Act now requires a minimum of documentation, but the information that matters most for choosing a model still has to be requested by the buyer and checked with your own tests.
Do you need to assess an AI vendor before you sign?
We help companies compare Claude, GPT and Mistral models on their own data, set up an evaluation set and prepare the questions and documentation the EU AI Act requires. As a Claude Certified Architect from Zagreb, we work with companies in Croatia and the DACH region.
Talk to an AI consultant