Anthropic's Opus 5.5 System Card Discloses Multi-Agent Scaling Results
Anthropic's Claude Opus 5.5 system card (section 8.12) discloses rare multi-agent scaling law experiments, reportedly scaling up to 100 parallel agents — the first time a major model vendor has published multi-agent scaling data at this scale in a system card.
Confirmed
- Multiple posters (@maksymandr, gleech (Joseph Van Pelt), xeophon, jyangballin) confirmed the section covers how capability changes as agents scale; jyangballin shared links and screenshots.
- First report of multi-agent runs scaling up to 100 parallel agents.
- ProgramBench experiments show multi-agent systems reach the same outcomes as single agents, but faster.
Contested details
- jyangballin noted the ProgramBench experiment used only 166 of 200 tasks.
- A reply to xeophon's post claims similar experiments have been run before, with intervals between them apparently growing.
Why it matters
This is a rare first-party disclosure on multi-agent scaling, offering vendor data on whether massive parallel agent decomposition is worthwhile, and may inform cost-benefit decisions in multi-agent system design.
2026-09-23 ~ 2026-09-23 · 6 related posts
Primary sources
- [source] Opus 5.5 system card reveals scaling law results for multi-agent systems — maksym_andr · 2026-09-23
- Opus 5.5 system card reveals multi-agent scaling law results — xeophon · 2026-09-23
- Opus 5.5 system card first to report multi-agent scaling up to 100 parallel agents — zephyr_z9 · 2026-09-23
- Opus 5.5 System Card Section 8.12 Reveals Multi-Agent Scaling Results — gleech · 2026-09-23
- [source] Anthropic's Opus 5.5 system card: multi-agent runs match single-agent outcomes faster — jyangballin · 2026-09-23
- [source] Opus 5.5 system card Section 8.12 has interesting multi-agent scaling law results — jyangballin · 2026-09-23