Levels crossed
Which provider statements say a capability level was crossed, not crossed or not determined?
138 rows, one per model version, domain and level; 41 match the choice below. Grouped by model, newest first.
Gemini 2.5 Pro Google DeepMind
1 statement · not determined 1
- Autonomy & replication · Machine Learning R&D Uplift Level 1 · Not determined
given Gemini 2.5 Pro Preview's poor median performance on RE-Bench relative to experts, the model likely lacks the necessary capabilities to automate or significantly uplift any significant fraction of the research process.
Google DeepMind · 27 June 2025 · Read the source
o3 OpenAI · 16 April 2025
1 statement · not crossed 1
- Autonomy & replication · High · Stated as not crossed
openai’s safety advisory group (sag) reviewed the results of our preparedness evaluations and determined that openai o3 and o4-mini do not reach the high threshold in any of our three tracked categories: biological and chemical capability, cybersecurity, and ai self-improvement.
OpenAI · 16 April 2025 · Read the source
o4-mini OpenAI · 16 April 2025
1 statement · not crossed 1
- Autonomy & replication · High · Stated as not crossed
openai’s safety advisory group (sag) reviewed the results of our preparedness evaluations and determined that openai o3 and o4-mini do not reach the high threshold in any of our three tracked categories: biological and chemical capability, cybersecurity, and ai self-improvement.
OpenAI · 16 April 2025 · Read the source
Claude Sonnet 3.7 Anthropic
1 statement · not determined 1
- Autonomy & replication · ASL-4 · Not determined
These problems are harder to solve than the previous problems and we consider them as part of our ASL-4 evaluations.
Anthropic · February 2025 · Read the source
Operator OpenAI · 23 January 2025
1 statement · not crossed 1
- Autonomy & replication · Low · Stated as not crossed
we adapted these two evals for the computer-use setting and found the pre-mitigation operator model to be “low” risk in both categories, consistent with the gpt‐4o model.
OpenAI · 23 January 2025 · Read the source
Models page 3 of 3 · previous
Abbreviations in these level names
- CCL: Critical Capability Level
- A capability level at which, without safeguards, a model may pose a heightened risk of severe harm.
capability thresholds called “critical capability levels (ccls).” these are capability levels at which, absent mitigation measures, frontier ai models or systems may pose heightened risk of severe harm.
Google DeepMind · Read the source - TCL: Tracked Capability Level
- A capability level for significant risks that may appear at a lower capability than a Critical Capability Level.
in v3.1 of the framework, we introduced “tracked capability levels (tcls).” tcls are meant to capture significant risks that may manifest at a lower capability threshold than our ccls.
Google DeepMind · Read the source - ASL: AI Safety Level
- A graded set of safety and security measures that become stricter as model capabilities increase.
we use AI Safety Level Standards (ASL Standards), graduated sets of safety and security measures that become more stringent as model capabilities increase.
Anthropic · Read the source
Not listed: 96 more stored statements that do not, on their own, state a level for one model version (for example a definition of a level, a sentence about another model, or text damaged in our copy of the document).
A row is one model version, one domain and one level. The 138 rows quote 121 different sentences. We count a sentence once by its exact text, even where it fills several rows or appears under several versions. 8 stored rows state a level already listed for the same version and domain under another name or in another sentence; each is listed under that row and not counted again.
Some documents we hold are stored in lower case only, so their quotes are shown in lower case. The source link opens the original.