Small print.
Wide open.

Table of Contents

We  have  observed  that  more  capable  models  are  generally  better  at  recognizing  and  preventing  high-stakes 
misuse. We believe that robust safety measures and model utility are not inherently in tension, and that safety 
can improve alongside the capability of the models themselves.

Our Approach to Model Releases

Each  release  is  accompanied  by  a  detailed  system  card  covering  capability  and  safety  evaluations,  external 
tester reports, and discussion of areas where performance remains limited or where the model shows signs of 
unwanted  behavior.  For  example,  AI  poses  clear  and  present  risks  in  the  cybersecurity  domain  as  advanced 
models could assist threat actors in security breaches and exploits. Mitigating these risks is a priority across our 
research, evaluation, and deployment practices. Before each model release, our Frontier Red Team conducts a 
detailed  analysis  to  determine  whether  new  models  provide  meaningful  new  capabilities  to  assist  potential 
attackers,  and  our  Safeguards  team  maintains  dedicated  classifiers  and  monitoring  systems  targeting  cyber-
related misuse, including prompts related to malware creation, phishing, and infrastructure attacks. 

We  make  decisions  about  model  development  and  deployment,  including  whether  to  begin  training  a  new 
model, on a case-by-case basis. Before releasing a frontier model, we consider a variety of factors, such as the 
capabilities and behaviors of the specific model at issue and of our models overall, how well our safeguards and 
security measures work, and how widely the model would be available. We plan to extend our testing to include 
more  assessments  at  each  stage  of  model  training:  pre-training,  post-training,  and  internal  deployment. 
Depending on these assessments, we may delay or suspend training or related research, delay releases, or limit 
access to models until we have confidence they can be developed and deployed safely. Our assessments are also 
informed  by  voluntary  pre-deployment  evaluation  of  certain  models  by  the  CAISSI  and  the  U.K.  AI  Security 
Institute. For certain models that materially advance the capability of the frontier in areas relevant to national 
security,  we  also  provide  designated  government  partners  with  expanded  early  access  so  that  they  can 
independently evaluate model capabilities and safeguards before release.

Based on our analyses, we do not release every model we develop for general use; when the model’s capabilities 
warrant  caution,  we  may  restrict  availability  or  decline  to  release  entirely.  For  example,  in  April  2026,  we 
initially released our frontier model, Mythos, as a gated research preview through Project Glasswing, available 
to select organizations for cybersecurity applications, rather than making it broadly available at launch due to 
its  advanced  capabilities  in  cyber  vulnerability  discovery. We  later  released  Fable  5  for  general  use,  but  with 
safety restrictions on certain cybersecurity, biology, and chemistry capabilities.

Race to the Top

Building reliable, trustworthy, and secure AI systems cannot be the task of one AI company alone. Our goal is to 
foster a race to the top in which AI developers compete on the safety and security of their systems, rather than 
rushing to release increasingly powerful models without adequate safeguards. In addition to publishing a large 
volume  of  research,  we  open  source  our  evaluations  where  possible  and  participate  in  groups  such  as  the 
Frontier Model Forum to share safety-critical findings with other AI developers. We have also sought to lead by 
example  in  ways  that  shape  industry  norms.  After  we  published  the  first  version  of  our  Responsible  Scaling 
Policy in September 2023, two other leading frontier AI labs adopted similar frameworks within months.

Building  on  this  approach,  we  have  publicly  called  for  “pacing  the  frontier,”  by  which  we  mean  establishing 
industry-wide standards primarily related to the use of AI to accelerate AI development, so that AI capabilities 
do  not  advance  faster  than  the  safety,  alignment,  and  security  practices  needed  to  support  them.  This  call  to 
pace the frontier contemplates frontier AI developers collaborating with governments and does not describe our 
own  model  development  or  release  timelines,  which  continue  to  be  informed  by  our  risk  assessments.  If 
standards of this kind are adopted industry-wide through government action or regulation, the rate at which AI 
capabilities  advance  could  be  faster  or  slower  than  it  is  today,  depending  on  the  extent  to  which  safety, 
alignment,  and  security  practices  advance  together  with  model  capabilities.  We  believe  the  first  step  toward 
“pacing the frontier” is independent verification of compliance with future regulatory standards, and we have 
begun  implementing  a  program  under  which  independent  evaluators  work  alongside  our  own  risk  and  safety 
teams with broad internal access, subject to legal, confidentiality, and security limits. As part of this program, 
we currently anticipate investing at least $1 billion over the next five years to build external reviewer capacity 
with Accenture. The terms of this arrangement have not been finalized and are subject to change. We have also 
called on other frontier AI developers to adopt comparable practices to demonstrate that safety commitments in 

233