GPT-5.5 Costs 49-92% More Than GPT-5.4 in Practice
GPT-5.5 doubled OpenAI's list price over GPT-5.4: $5 per million input tokens and $30 per million output, up from $2.50 and $15. OpenRouter's cost analysis, published May 4, measured what that did to real bills. The answer: costs rose between 49% and 92%, depending almost entirely on prompt length.
That range is the useful part. GPT-5.5 is more token-efficient on long prompts, which claws back part of the price increase. On short prompts it isn't, and you pay nearly the full doubling. If your traffic is mostly short requests, the upgrade costs about twice as much. If it is mostly long-context work, it costs about half again as much.
How OpenRouter measured it
OpenRouter took a "switcher cohort": users whose main model was GPT-5.4 before launch and who then moved to GPT-5.5. Both models use the same tokenizer family, so token counts are directly comparable. The comparison is the same users running similar workloads, not a synthetic benchmark.
The numbers by prompt length
| Prompt size | Completion tokens vs GPT-5.4 | Effective cost vs GPT-5.4 |
|---|---|---|
| Under 2K | +7% | +92% |
| 2K to 10K | +52% | +69% |
| 10K to 25K | -32% | +51% |
| 25K to 50K | -19% | +62% |
| 50K to 128K | -28% | +49% |
| 128K+ | -34% | +85% |
Two patterns stand out:
- Short prompts get the worst deal. Under 2K tokens, completions are about the same length, so the doubled rate passes straight through. This is where chat, classification, and most tool calls live.
- The 2K to 10K band is the surprise. Completions got 52% longer, not shorter. Typical agent turns and RAG answers fall here.
From 10K up, GPT-5.5 writes 19% to 34% less than GPT-5.4, which is why the increase settles near 50% to 60%. The 128K+ bucket jumps back to +85%, which is consistent with input tokens dominating the bill at that size: fewer output tokens can't offset a doubled input rate.
Is the upgrade worth it?
On LMRank, GPT-5.5 scores 9.4 against GPT-5.4's 8.9. That is a real gap, and for hard reasoning and long agentic runs it can justify the cost. For everything else, compare the alternatives at May 2026 list prices:
- GPT-5.4 (8.9, $2.50/$15): still available. For short-prompt traffic it costs roughly half as much as GPT-5.5 for the same work.
- GPT-5.4 Mini (8.3, $0.75/$4.50): the obvious home for classification, extraction, and short chat.
- Claude Opus 4.7 (9.6, $5/$25): same input price as GPT-5.5 and cheaper output. If you are paying flagship rates anyway, test both.
- DeepSeek V4 Pro (9.0): launched at $1.74/$3.48 and, at the time of writing, discounted 75% to $0.435/$0.87. The discount matters more than the list price for budgeting.
What to do
Don't upgrade by default. Split your traffic by prompt length and look at each bucket:
- Under 10K tokens: keep GPT-5.4 or move down to Mini unless an eval shows GPT-5.5 winning on your task.
- 10K to 128K: GPT-5.5's shorter completions make the premium roughly 50%. Route here if quality moves your metric.
- Measure cost per completed task, not cost per token. Token efficiency is now a model property you have to benchmark, just like accuracy.
Compare pricing and scores across models on the LMRank leaderboard.