Anthropic: frontier models with safety as the starting point
The prevailing view said safety and capability pulled in opposite directions. Moonshot bet that the most careful lab could also be among the most capable.
The Bet
Anthropic was founded in 2021 by a group of researchers who believed frontier AI systems should be built with safety as a starting point, not an afterthought. At the time, many people framed that as a trade-off: every bit of caution was assumed to cost capability.
Moonshot’s bet was that the trade-off was overstated. Understanding a model deeply, how it behaves, where it fails and how to steer it, is also how you make it more useful.
Understanding a model deeply is also how you make it more useful. Rigour and performance can compound together.
Moonshot’s View
We believed that as AI systems became more powerful, the customers with the most at stake would care most about reliability and trust. A lab that led on those qualities would not be a niche; it would be a default choice for serious work.
We also believed in the team’s research depth. In a field moving this fast, the ability to do careful science at the frontier is rare, and it is hard to copy.
Room for Disagreement
The strongest case against was that a safety-first lab would move too slowly in a race defined by speed, and that competing with larger, better-funded labs would be close to impossible for a newcomer.
Those were fair concerns. We disagreed with the idea that careful and competitive were opposites, and with the assumption that the frontier would stay closed to new entrants.
Notable
Anthropic, in brief.
- Sector
- AI safety
- Founded
- 2021
- Founders
- Dario Amodei, Daniela Amodei and others
- Moonshot backed
- Early
- The contrarian call
- Safety-first could also mean frontier-class
- Money raised
- $122.3B to date
- Company value
- $965B
- Moonshot 25
- No. 2 of 25 overall
Source: money raised and company value from Caplight, as of Nov 17, 2026. View source
Pitch Moonshot
Building something the crowd doesn’t believe in yet?
Tell us in about two minutes. No warm intro needed, and every pitch is read by the team.