OpenRouter measured real switcher-cohort usage and found GPT-5.5 costs 49-92% more than GPT-5.4 despite shorter completions on long prompts.
Key Takeaways
GPT-5.5 doubled list prices: input $2.50 to $5.00/M tokens, output $15 to $30/M tokens.
For prompts over 10K tokens, completions shrink 19-34%, partially offsetting the price hike.
For prompts under 10K tokens, completions are the same length or longer, so the full 2x price hits hardest.
Cost per million OpenRouter tokens rose 49% (50K-128K range) to 92% (under 2K range) across the switcher cohort.
Methodology used same-user before/after comparison on OpenRouter request logs; same tokenizer family means direct token counts are comparable.
Hacker News Comment Review
The single substantive comment flags a key methodology gap: multi-turn efficiency is not measured, and stronger models often close tasks in fewer turns, which could materially change the net cost picture for agentic workflows.
Notable Comments
@jsnell: argues turn-count reduction for agentic coding may be a larger efficiency lever than per-response verbosity, a variable the analysis does not control for.
