Dev burns 5B tokens a day on open models as multi-agent swarms emerge as a new scaling axis
xeophon · x · 2026-09-17
Developer Florian Brand reports his daily token usage exploded from under 100M to over 5B tokens (open models only), driven by more reliable long-running agents, cheap GPU access, and multi-agent swarms. Key points: single agents are hitting wall-clock-time limits, making parallel swarms an underexplored scaling axis; swarms excel at data work and broad research via agent-to-agent communication; SOTA open models can maintain persistent fleets of sub-subagents that decompose, delegate, and relay findings. He also teases a section on GPT-6's training. Even with visible inefficiencies at every step, "point a swarm at a problem and waste a bazillion tokens" is working better than expected.
More from coding & agent
- Typesafe AI Launches Jev: A Decision Engine at 70-500ms and $0.042 Per Million Tokens — brandon_galang · 2026-09-17
- Open-source Helicon app brings Meta's Muse Code CLI to Windows desktop — alexandr_wang · 2026-09-17
- Netlify to livestream a Grok Bot autonomously building and deploying a site — thisiskp_ · 2026-09-17
- OpenJev open-sources Jev-style semantic decisions on a single RTX 3090 with a frozen 4B model — alexcovo_eth · 2026-09-17
- Browser Use CEO on agentic engineering: humans are the bottleneck, he stopped reading code — David Ondrej · 2026-09-17
- Jev: an open-source action-picker that splits agent thinking from clicking — alexcovo_eth · 2026-09-17