Small print.
Wide open.

Table of Contents

run to attempt to achieve the intended model improvements. We have in the past experienced, and may in the 
future experience, such failed or underperforming training runs, resulting in a loss of invested resources with 
limited or no corresponding benefit to our business. Given the scale of compute and other resources required 
for frontier model training and the extended timelines involved, a failed or underperforming training run could 
also  delay  our  product  roadmap,  affect  our  competitive  position,  and  adversely  affect  our  business,  financial 
condition, and results of operations.

In  addition,  new  models,  features,  or  capabilities  may  not  be  introduced  on  schedule,  may  not  perform  as 
intended,  may  require  significant  iteration  after  release,  may  fail  to  gain  meaningful  adoption,  or  may  take 
longer than expected to gain meaningful adoption, any of which could reduce customer demand or place us at a 
competitive disadvantage. In addition, considerations relating to model safety, interpretability, alignment, and 
welfare may lead us to limit or delay the deprecation of older models or to continue supporting legacy versions 
alongside  newer  releases,  creating  trade-offs  in  product  development  and  compute  distribution  that  could 
adversely  affect  innovation  velocity,  infrastructure  utilization,  and  revenue.  In  addition,  customers  who  have 
integrated  older  models  into  their  workflows  may  resist  migration  to  newer  versions,  and  any  accelerated  or 
forced deprecation could result in customer attrition, contractual disputes, or reputational harm. We have also 
changed, and may in the future change, usage limits, rate limits, or model and product capabilities, which could 
adversely affect our reputation and the perceived value of our offerings. Some of these changes have resulted, 
and  could  in  the  future  result,  in  claims  that  our  disclosures  regarding  model  access  or  capabilities  were 
inadequate or misleading.

Further,  the  industry  benchmarks  commonly  used  to  test  AI  models  may  not  accurately  reflect  real-world 
performance  or  utility,  and  even  models  that  perform  well  on  such  benchmarks  may  not  meet  customer 
expectations when deployed in production environments, which could lead to customer dissatisfaction, reduced 
usage,  or  reputational  harm.  If  we  are  unable  to  effectively  innovate  and  improve  our  models,  platform,  and 
applications  while  addressing  these  challenges,  or  if  competitors  are  able  to  deliver  AI  solutions  that  are 
perceived  as more  capable,  safer,  or  more  reliable,  or  that  offer  adequate  performance  at  a  lower  cost,  our 
business, financial condition, and results of operations could be adversely affected.

In addition, innovation in AI involves addressing complex and evolving challenges related to data quality, data 
integrity, training methodologies, bias, safety, accuracy, and the behavior of model outputs. AI algorithms may 
be  flawed,  and  AI  systems  may  contain  bugs,  errors,  or  defects  or  fail  to  follow  instructions  or  produce 
inaccurate, misleading, offensive, inappropriate, or otherwise harmful outputs; reflect or amplify unwanted or 
unknown  biases  present  in  training  data;  or  raise  concerns  relating  to  intellectual  property,  privacy,  or  other 
legal  and  ethical  considerations.  We  must  continually  invest  resources  to  identify,  mitigate,  remediate,  and 
reduce the occurrence of such issues, including through managing and improving datasets, training techniques, 
evaluation methods, safeguards, and human oversight. However, we may not be able to obtain the datasets that 
we need, including through licenses or strategic data partnerships. In addition, we or our data providers may 
use synthetic or AI-generated data in the training process, which could introduce quality issues, amplify biases 
or  errors,  contribute  to  model  degradation,  or  raise  novel  legal  questions  regarding  ownership,  licensing,  or 
intellectual property rights that may not be timely or clearly resolved by courts or regulators. Such unintended 
consequences, failures, or misuse of our models could impair customer trust, slow adoption, expose us to legal 
or regulatory scrutiny, or adversely affect our brand and reputation. 

If  we  are  unable  to  secure  adequate  compute  or  if  we  fail  to  direct  our  compute  effectively 
across our workloads, we may be unable to train models that remain competitive with those 
of our competitors or adequately serve our customers.

Our  ability  to  develop,  train,  improve,  and  deploy  frontier  models  depends  on  our  access  to  enormous  and 
increasing  quantities  of  specialized  compute.  The  training  of  Claude  models  has  required,  and  we  expect  will 
continue to require, substantial compute. Industry participants, academic researchers, and our own experience 
suggest  that,  in  certain  contexts,  applying  greater  amounts  of  compute  during  model  training  has  been 
associated with improvements in model capabilities. These observations are consistent with aspects of broader 
empirical  relationships  among  model  performance,  model  size,  training  data,  and  compute  scale  that  are 
commonly referred to as “Scaling Laws,” although the extent, efficiency, and durability of such improvements 
are  uncertain,  and  these  relationships  may  plateau,  diminish,  or  fail  to  materialize  at  larger  scales  or  under 
different  conditions.  Accordingly,  our  ability  to  remain  at  the  frontier  of  AI  capabilities  is  substantially 
dependent on our ability to secure compute at the scale required to sustain all our workloads, including to train 

29