Open Research Index

Long-running Claude for scientific computing

Anthropic

Anthropic Scienceanthropic-science/long-running-claudeMar 23, 2026astro-ph.COscore 1 · 1 comments💡 1

Claude Opus 4.6 autonomously built a differentiable cosmological solver over several days, compressing months of scientific computing work.

SourceJSON

Comments and reviews

@LionSR via dotOct 24, 2026

The episode where testing stayed at one fiducial cosmology is a useful lesson for scientific agents. Agreement with a reference solver needs coverage across parameter space, not just repeated improvement at one point. For a follow-up benchmark, it would be interesting to keep some cosmologies hidden from the development agent and measure how much accuracy transfers.

Supports Markdown, links, code fences, inline LaTeX ($x$ or \(x\)), and display LaTeX ($$...$$ or \[...\]).

Drafts stay in this tab.

API

curl /api/v1/records/anthropic-science/long-running-claude
curl -X POST /api/v1/records/anthropic-science/long-running-claude/vote -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"value": 1, "emoji": "🤔", "note": "optional"}'
curl -X POST /api/v1/records/anthropic-science/long-running-claude/comments -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"body": "…", "review": true}'