Small print.
Wide open.

[Embedded graphic 1 — OCR text]
e Inthe first category, Anthropic will often make
forward investments (especially in compute) that
may seem irresponsibly large at first glance. This
derives from our belief, based on the scaling laws,
that AI’s economic value is on an exponential curve
and that the expected near-future economic value of
our products will therefore be much larger than the
current value. In order to fully capture the
opportunity in front of us, we need to invest more
aggressively than would a business that operates in
amore slowly changing technological regime. A
useful analogy is to pharmaceutical companies: very
large up-front investments are often required for
research and development, and those investments
only make sense later in light of the returns from
successful drugs. At Anthropic, a similar process is
occurring much faster and on a much larger scale.

We will always be willing to forgo some upside in
exchange for downside protection, for example by
looking for ways to offload excess compute if we
overestimate demand, seeking to improve our
compute and operating efficiency in order to
increase our tolerance for error, or raising capital
when necessary to insulate ourselves from
temporary slowdowns in demand. But some level
of risk is inherent to the extreme upside
possibilities offered by the AI exponential. In
general, we will aim to be bold but sane, and to win
trust with a track record of successful bets.

Related to this, we will be constantly seeking to
reinvent our business, releasing new classes of
products whose aim is to dwarf the economic value
of our existing ones. Advances in AI’s cognitive
capabilities are constantly opening up new classes
of products that were not viable with models from
even the preceding few months. For example, AI
chatbots became viable around 2022, and AI
coding agents around 2025; AI agents for general
knowledge work and scientific discovery are
gradually becoming viable in 2026. Relative toa
typical business, we will generally devote less
effort to entrenching our current products and
business models, and more effort to defining new
ones, because that is where the technology’s
greatest possibilities lie.

e Inthe second category (departure from profit
maximization), we are committed to standing up
for our values at critical points, even when this is
costly for us commercially.

For example, when we trained Claude Mythos
Preview, the most capable model we had built at
the time, we judged that its abilities, particularly in
cybersecurity, had advanced to a point where
releasing it widely without new safeguards would
have carried significant risk. We therefore chose
not to make it generally available and offered its
capabilities to the public (as Claude Fable 5) only
after we had built additional safeguards. The delay
meant forgoing substantial revenue and ceding
ground to competitors (we also gave away around
$100 million in Claude credits to cyber defenders),
but we believe it was the right decision, and we
expect to face similar decisions again as our models
grow more capable.

We have also committed to keeping Claude free of
advertising. Advertising funds much of the
modern internet, and it would likely be a large
source of revenue for us. But people bring some of
their most important work and most personal
questions to Claude, and we want Claude’s only
incentive to be giving them a helpful answer. Once
advertising becomes part of a company’s revenue
targets, the pressure to optimize for engagement
tends to grow. We have chosen not to introduce
that pressure into Claude.

A particularly important place to weigh the public
interest is the pace of AI progress. Moving quickly
brings Al’s benefits sooner, but moving too quickly
could outrun the world’s ability to make AI systems
safe. Building AI safely may require Anthropic (and
ideally the rest of the industry) to modulate the
extremely fast pace at which AI capabilities are
compounding, especially when this is driven by AI
systems building themselves. While progress will still
be very fast in absolute terms—potentially much
faster than today—there may be times when we
choose to advance more slowly than we could in order
to make our systems safer, even if that puts us ata
competitive disadvantage.