Small print.
Wide open.

Table of Contents

We  have  built  our  models,  platform,  and  applications  for  usefulness  in  production  environments  instead  of 
benchmark performance. This distinction matters, as many AI systems prove difficult to implement and fail to 
deliver  the  expected  value  in  complex  enterprise  environments,  where  governance  and  security  concerns 
remain among the top barriers to adoption. 

Our applications, such as Claude Code and Claude Cowork, multiply the impact of employees. Over time, our 
applications  have  become  more  capable  and  are  now  frequently  being  used  to  conduct  long-running 
autonomous work alongside employees. Organizations also integrate Claude as the intelligence that powers and 
transforms  internal  business  processes.  In  addition,  our  Claude  Platform  is  used  as  the  foundation  for 
enterprises building AI-native products.

As enterprise customers integrate Claude into increasingly complex workflows, they surface the use cases and 
requirements that shape our roadmap. They also provide feedback that allows us to continuously improve our 
models, platform, and applications as well as develop new solutions to advance their AI adoption. 

With each model and product advancement, Claude is able to address more valuable use cases, deepening its 
impact for existing customers while lowering the adoption threshold for new ones. When a model crosses the 
capability,  reliability,  and  trust  threshold  required  for  a  category  of  work,  adoption  in  that  category  generally 
moves quickly from experimentation into production. Coding was the first such category: within roughly a year, 
Claude progressed from suggesting individual lines of code to operating fleets of agents working autonomously 
across  production  codebases.  Cybersecurity  is  following  a  similar  curve,  with  Claude’s  performance  on 
ExploitBench,  a  benchmark  developed  by  Carnegie  Mellon  researchers  that  evaluates  AI  models’  ability  to 
exploit software vulnerabilities, increasing from 40% with Opus 4.8, to 78% with Claude Mythos 5, and to 91% 
with Opus 5.5 over a period of approximately four months. Cyberdefenders are now identifying and patching 
vulnerabilities  at  scale.  We  expect  these  improvements  in  model  capabilities  to  unlock  additional  use  cases. 
Claude  Security  and  our  recently  launched  OSS  Scanner  are  early  examples,  providing  the  ability  to  scan 
codebases on a continuous basis and propose fixes before attackers find vulnerabilities.

Claude as a Virtual Collaborator

We  believe  we  have  reached  an  inflection  point  where  AI  systems  can  function  as  genuine  collaborators  in 
professional work. They quickly complete substantive tasks that might otherwise occupy a skilled professional 
for several hours, a full day, or longer. AI systems are also becoming increasingly capable of performing tasks 
that were previously too difficult or costly to complete or surfacing original insights and analyses.

We refer to these AI systems that work alongside humans to perform complex, multi-step work at professional 
levels  as  Virtual  Collaborators.  Virtual  Collaborators  carry  the  access  rights,  system  connections,  and  context 
that  their  organizations  grant  them,  which  allows  them  to  function  more  accurately  and  reliably.  They  can 
navigate  and  use  internal  tools  and  execute  workflows  that  span  multiple  steps  that  require  sustained 
reasoning.  Virtual  Collaborators  are  AI  designed  to  be  embedded  where  work  gets  done,  and  our  aim  is  to 

189