AI Safety Tracker

Embeddable figures

8 figures

Eight figures are published as SVG images for use on other pages. Each image prints its title, sources and licences, date, and the address of its section here, and links to that section. An image placed from its published address is rebuilt with the data. Terms of reuse are on the terms page.

How far the open-weight frontier trails the overall frontier

Largest training run to date, all models against open-weight models. 8 months between the overall frontier reaching 3.8e25 FLOP and an open-weight model doing so. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/figures/.

Image
SVG, 948 by 502
<a href="https://aisafetytracker.org/capability/#open-weights"><img src="https://aisafetytracker.org/embed/open-weight-lag.svg" alt="Step chart of the largest training run to date for all AI models against open-weight models, with the lag between them in months, from AI Safety Tracker" width="948" height="502" loading="lazy" style="max-width:100%;height:auto"></a>

Training-compute thresholds in law, against the models above them

4 of 6 thresholds were written after a model had already crossed them. Each mark is the day an instrument took its final form. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/figures/. Compute Threshold Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data JSON CSV Training-compute thresholds written into law, executive action or published policy, with the text that sets each one. Whether a given model crosses a threshold here is arithmetic on Epoch's compute estimates, most of which are estimates rather than disclosures, and is not a legal finding about any model or developer. Cite AI Safety Tracker, "Compute Threshold Index". https://aisafetytracker.org/capability/#thresholds. Retrieved 23 October 2026.

Image
SVG, 948 by 485
<a href="https://aisafetytracker.org/capability/#thresholds"><img src="https://aisafetytracker.org/embed/compute-thresholds.svg" alt="Timeline of the training-compute thresholds written into law and policy against the AI models with a published compute figure, from AI Safety Tracker" width="948" height="485" loading="lazy" style="max-width:100%;height:auto"></a>

Frontier AI safety frameworks: what each developer commits to

7 of 8 frontier developers have published a framework. 4 commit to stopping, 3 to a decision process, and 4 to evaluation by someone else. Frontier Safety Framework Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data JSON CSV Our own reading of each developer's published safety framework and model cards. Every field is what the developer says about itself: these are self-assessments under self-written frameworks with no common scale and no external audit, and they are indexed here for exactly that reason. Not a ranking, and not a verification. Cite AI Safety Tracker, "Frontier Safety Framework Index". https://aisafetytracker.org/alignment/#frameworks. Retrieved 23 October 2026.

Image
SVG, 948 by 454
<a href="https://aisafetytracker.org/alignment/#frameworks"><img src="https://aisafetytracker.org/embed/safety-frameworks.svg" alt="Table of the frontier AI developers&#39; published safety frameworks, with the risk domains each covers and whether each commits to external evaluation and to stopping, from AI Safety Tracker" width="948" height="454" loading="lazy" style="max-width:100%;height:auto"></a>

Frontier AI releases against the system card published for each

Every release Epoch AI lists from the frontier developers since January 2025. 48 of 69 shipped with a system card on the day, 4 had one later and 17 have none. Release Safety Disclosure Index This site Licence MIT Updated monthly Retrieved , 10 days ago Data JSON CSV Our own reading of the system card, model card or safety report each developer published for each release Epoch AI lists from 2025, joined to Epoch's list. Every cell is what a developer says about its own model under its own framework: whether a card came out with the release, whether it reports dangerous-capability results, which government institutes it names as pre-release testers, and whether it says a risk threshold was reached. None of it is verified. A release Epoch adds after the last review is shown as awaiting review. Cite AI Safety Tracker, "Release Safety Disclosure Index". https://aisafetytracker.org/alignment/#releases. Retrieved 18 October 2026. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 10 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 18 October 2026 via AI Safety Tracker, https://aisafetytracker.org/figures/.

Image
SVG, 948 by 442
<a href="https://aisafetytracker.org/alignment/#releases"><img src="https://aisafetytracker.org/embed/release-cards.svg" alt="Bar chart of frontier AI model releases by developer, split by whether a system card was published on the day of release, later or not at all, from AI Safety Tracker" width="948" height="442" loading="lazy" style="max-width:100%;height:auto"></a>

Research on AI safety against research on frontier AI models

Works indexed by OpenAlex per year by research area, logarithmic scale. In 2025, 14 works on safety, interpretability, evaluation or catastrophic risk for every 100 on frontier models; 330 in 2020. OpenAlex OurResearch Licence CC0-1.0 Updated continuous Retrieved , today An open index of scholarly works. Counts here are matches on a stated search phrase in title or abstract, not a curated field classification: a paper that studies alignment without using the phrase is missed, and one that mentions it in passing is counted. The exact query behind every number is printed with it. Recent years undercount because indexing lags publication, and the current year is always partial. Cite OurResearch, "OpenAlex". https://openalex.org. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/figures/.

Image
SVG, 1088 by 518
<a href="https://aisafetytracker.org/alignment/#research"><img src="https://aisafetytracker.org/embed/research-volume.svg" alt="Line chart of research works per year on AI safety, interpretability, evaluation, catastrophic risk, governance and frontier models, from OpenAlex via AI Safety Tracker" width="1088" height="518" loading="lazy" style="max-width:100%;height:auto"></a>

Frontier AI training runs against published emissions figures

Every run since 2020 above 1.0e21 FLOP with a compute estimate, by developer, logarithmic scale. 1 of the 24 models at or above 1e25 FLOP has a published training-emissions figure. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/figures/. Environmental Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 7 datasets Our own reading of what frontier developers and the operators of their data centres have published about training energy, emissions, power sources and energy per prompt, with permits and press reporting marked separately from company statements. Most cells record an absence. Company figures are self-reported and mostly unaudited, and market-based emissions reflect contracts rather than the grid at the hour of use. Cite AI Safety Tracker, "Environmental Disclosure Index". https://aisafetytracker.org/environment/. Retrieved 23 October 2026.

Image
SVG, 1048 by 538
<a href="https://aisafetytracker.org/environment/#training-runs"><img src="https://aisafetytracker.org/embed/training-runs.svg" alt="Strip chart of frontier AI developers&#39; training runs by compute, with the runs that have a published training-emissions figure ringed, from AI Safety Tracker" width="1048" height="538" loading="lazy" style="max-width:100%;height:auto"></a>

Tokens processed per month, as each AI company states it

Every aggregate token count a company has stated, converted to tokens per month, logarithmic scale. The largest is ByteDance, Doubao models at 5.5 quadrillion tokens a month in June 2026. AI Usage Disclosure Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data 3 datasets Our own compilation of the token volumes, user counts and shares of use that companies and studies have put on the record, each linked to where it was stated. Scopes, units and token definitions differ between companies, so no two figures are comparable without the scope printed beside each. Figures known only through press reporting are marked as such. Not a market share estimate, and not a measure of any company's actual volume beyond what it chose to state. Cite AI Safety Tracker, "AI Usage Disclosure Index". https://aisafetytracker.org/usage/. Retrieved 23 October 2026.

Image
SVG, 1048 by 582
<a href="https://aisafetytracker.org/usage/#tokens"><img src="https://aisafetytracker.org/embed/token-volumes.svg" alt="Line chart of the monthly token volumes that Google, ByteDance, OpenAI and Microsoft have stated, by scope, from AI Safety Tracker" width="1048" height="582" loading="lazy" style="max-width:100%;height:auto"></a>

The law that applies to AI, by jurisdiction and how binding it is

86 instruments indexed in 11 jurisdictions. 48 are binding law, and 10 of those name AI in their own title. AI Law and Policy Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data JSON CSV Our own reading of primary instruments, with a link to each. Not legal advice, not a compliance checklist and not exhaustive. Instruments that never mention AI are included where they apply to it, which is most of the Australian list. Statuses and commencement dates go stale; every record carries its review date. Cite AI Safety Tracker, "AI Law and Policy Index". https://aisafetytracker.org/policy/. Retrieved 23 October 2026.

Image
SVG, 948 by 499
<a href="https://aisafetytracker.org/policy/#compare"><img src="https://aisafetytracker.org/embed/law-by-jurisdiction.svg" alt="Bar chart of the AI law and policy instruments indexed in each jurisdiction, split by how binding each is, with the binding law that names AI, from AI Safety Tracker" width="948" height="499" loading="lazy" style="max-width:100%;height:auto"></a>