Small print.
Wide open.

Table of Contents

clarity  of  mission  reduces  friction,  accelerates  decision-making,  and  fosters  a  culture  where  safety  and 
capability are pursued in tandem.

Talent and Culture

Talent is one of our most important advantages, because technology, capital, product, and distribution all flow 
from superior talent. Talent density matters more than talent mass. Brilliant people are inspired and motivated 
when  surrounded  by  equally  brilliant  colleagues,  creating  an  environment  that  attracts  more  exceptional 
people. This talent can leverage Claude to maximize each individual’s impact. 

We believe we have assembled the deepest concentration of AI research talent in the world, including many of 
the  researchers  who  pioneered  the  techniques  that  underlie  modern  LLMs.  Our  founding  team  originated 
several  foundational  advances  in  AI,  including  Scaling  Laws,  GPT-2  and  GPT-3,  RLHF,  and  mechanistic 
interpretability.  Our  conviction  that  safety  and  capability  must  advance  together  helps  us  attract  and  retain 
world-class  talent,  including  researchers  and  engineers.  For  example,  we  have  experienced  approximately  1% 
regrettable attrition of our employee base overall in the TTM period ended September 30, 2026.

We maintain a commitment to open internal collaboration and truth seeking, which is underpinned by our low-
ego,  high-impact,  and  mission-aligned  culture.  Researchers  share  results  openly,  and  colleagues  critique  and 
redirect in real time. When multiple teams pursue parallel approaches, the norm is to rigorously evaluate which 
approach is demonstrating stronger and more aligned results. This collaborative dynamic is difficult to replicate 
because it requires both the foundation of trust we have built from the earliest days of our organization and a 
relentless commitment to preserving such trust. That trust manifests in ways that are uncommon at our scale: 
teams that consistently pull in the same direction rather than at cross-purposes.

We believe our culture has driven our talent density as we have scaled.

Frontier and Safety Research

Our  frontier  model  leadership  is  a  direct  product  of  our  talent  density.  We  believe  our  researchers  have 
consistently pushed the boundaries of intelligence, capability, and reliability further than bigger teams or those 
with  larger  budgets.  Claude  is  widely  recognized  for  its  performance  on  real-world  tasks  across  industries, 
including coding, medicine, law, science, and mathematics. 

Our approach to research provides key advantages to our business:

•

•

•

Our  Model  Research  and  Compute  Performance  Are  Co-designed  from  the  Start.  Architectural 
decisions  are  made  with  compute  efficiency  in  mind  from  day  one.  Our  model  architecture  and 
inference  infrastructure  are  designed  in  a  single  unified  process.  AI  researchers  and  inference 
infrastructure engineers share planning cycles, reporting lines, and performance targets. This approach 
enables us to develop model versions that are both more capable and more efficient compared to prior 
versions.

Our  Pre-training  Process  Is  Engineered  for  Capital  Efficiency.  We  focus  on  a  clear  pipeline  for 
developing model improvements to architecture, training methodology, and data. These are tested at a 
smaller scale before they consume significant compute. The bar for inclusion in a frontier training run 
is  high  and  strictly  evidence-based:  we  require  demonstration  that  a  given  improvement  holds  as 
models  increase  in  size,  not  merely  that  it  performs  well  at  reduced  scale.  Frontier  training  runs  are 
planned, scheduled capital deployments with known inputs and rigorous inclusion criteria. The result 
is  a  repeatable  process  for  converting  research  investment  into  capability  gains,  while  managing  the 
risks inherent to frontier training runs. 

Our Post-training Investment Is Purpose-Built for Agentic Capabilities. Post-training has grown from 
a finishing step into a major research investment, reflecting the market shift from assistants to agents. 
This  is  where  our  models  hone  their  ability  to  write  code,  use  tools,  follow  instructions,  and  operate 
software  on  a  user’s  behalf.  We  train  against  complete,  realistic  tasks  whose  complexity  reflects  real 
professional work. 

187