DeepSeek's super-agent is 1.9x faster but burns 3.8x the tokens
AIFlow_ML · x · 2026-11-03
The author digs into DeepSeek's own comparison between its super-agent and a neutral single session: the super-agent finishes tasks 1.9x faster, but consumes 23.2M tokens vs 6.1M — about 3.8x more. DeepSeek itself admits trading more tokens for shorter, steadier runtimes.
The cost argument: on rented tokens that's a direct bill; on your own GPUs it's power and time. He also notes the coding task in the test was just a tool that burns time, not a realistic coding workload.
Related event: DeepSeek Super-Agent Is 1.9x Faster but Burns 3.8x Tokens(4 posts)→
More from coding & agent
- Neuralift launches MCP server to bring customer-data intelligence into Claude and agents — jonathanmendez · 2026-11-03
- Containment failure vs authority failure: two distinct agent security failures — rseroter · 2026-11-03
- Benchmark: Claude Code calls your MCP tool only 34.5% of the time when it overlaps built-ins — Future_AGI · 2026-11-03
- Search2o: an agent server bundling framework, runtime, semantic router and state — venkat11 · 2026-11-03
- Anthropic's daily merged code changes jumped from ~2,500 to 30,000+; 7-person team rebuilt dev stack around agents — msharmas · 2026-11-03
- 259 indie maker income posts analyzed: median $400, top 10 take 52% of all revenue — heyshrutimishra · 2026-11-03