Physix Frontier · News Briefing Card (Hacker News · Oct 3, 2026)

Hacker News Debates the Million-Token-Per-Second AI Scenario

KEY FACTS

The article imagines an extreme scenario in which a single AI agent outputs one million tokens per second.
The author notes that one million tokens per second could mean four different things: context capacity, input processing, aggregate throughput, or single-agent output.
The piece uses Claude Opus 5.5 as an example, stating its context window is 1 million tokens while its output ceiling is smaller.
The article demonstrates with forty app drafts totaling 800,000 tokens: after adding a 60-second check, the end-to-end speedup is only 132.57x.
The author argues that once generation becomes cheap, taste, specifications, and judgment become the truly scarce resources.

KEY DATA

1M tokenClaude Opus 5.5 context window

800,000 tokenDemo draft output volume

132.57×End-to-end speedup

98.7%Check time share

PHYSIX OBSERVATION

This is a thought experiment rather than a measured test, but it punctures an old flaw in the industry narrative: equating generation speed with productivity. In real workflows, testing, verification, physical experiments, and human judgment do not speed up with tokens. While vendors compete on throughput, users should be asking how much time is saved end to end, not how pretty the peak benchmark looks.

Source: Hacker News report