Astra 6 Ultrafast Mode: 300 Tokens/Sec Changes Agent Workflows, But Tools Are Now the Bottleneck

soumitrashukla9 · x · 2026-10-01

Researcher Sayash Kapoor shares takeaways from days with Astra 6's new ultrafast mode:

300 tokens/second output, 8x regular speed — fast enough to stay in flow on one task instead of context-switching, especially for work needing frequent human input. Sol-class models will be twice as fast again.
End-to-end speedup is only 2-4x: tool calling and code execution time are unchanged, so external factors may soon become the main bottleneck for agentic work. Arvind Narayanan had predicted both shifts: agent latency will be bottlenecked by shell commands, build pipelines, computer use, and web extraction — infrastructure that never faced the optimization pressure LLM inference has.
Computer use is one area where token speed really helps.

More from coding & agent