GPT-5.5 costs exactly twice what GPT-5.4 did at the API level - $5.00 per million input tokens (up from $2.50) and $30.00 per million output tokens (up from $15.00). But the actual hit to your bill depends heavily on what you're building.
OpenRouter, which routes API calls across dozens of AI models, tracked users who switched from GPT-5.4 to GPT-5.5 and measured the real cost difference. The result wasn't a flat 100% increase. Because GPT-5.5 produces shorter responses than its predecessor for the same prompts, real-world cost increases landed between 49% and 92%.
Where the Savings Partially Kick In
For prompts longer than 10,000 tokens (roughly 7,500 words of input), GPT-5.5 generates 19-34% fewer completion tokens. That efficiency partially offsets the doubled output price, bringing the net cost increase down to around 49-62% for those use cases rather than the full 100% you'd expect from the price card alone.
Short prompts don't get that benefit. Requests under 2,000 tokens saw cost increases of 92% - nearly the full price doubling - because completions didn't shrink proportionally. At the very long end, prompts over 128,000 tokens (about 90,000 words) saw costs jump 85%, suggesting the efficiency gains plateau or reverse for extremely large contexts.
Who Gets Hit Hardest
Anyone using ChatGPT through the API for short, conversational tasks - customer support responses, quick summaries, simple Q&A - faces the steepest increase. If you're processing long documents where GPT-5.5's tighter output actually saves tokens, the real-world increase is closer to 50%.
OpenRouter's full cost analysis breaks down the impact by prompt-length bucket, which is worth checking if you have actual usage logs to compare against.
A 50-92% price increase is significant on any API budget. If GPT-5.4 was handling your tasks well, running a direct quality comparison on your specific workloads before committing to the upgrade makes sense - the shorter completions GPT-5.5 produces might also affect output quality for tasks where thoroughness matters.