For longer prompts of more than 10,000 tokens, GPT-5.5 produces responses that are 19% to 34% shorter, which helps reduce the cost impact somewhat. For medium-length prompts between 2,000 and 10,000 tokens, however, responses are around 52% longer. With short prompts below 2,000 tokens, answer length remains roughly unchanged, meaning users face almost a full doubling of costs. The figures are based on OpenRouter’s own usage logs from April 2026.
|
Input length |
Avg. $/M tokens GPT-5.4 |
Avg. $/M tokens GPT-5.5 |
Change |
|---|---|---|---|
|
< 2K tokens |
$4.89 |
$9.37 |
+92% |
|
2K–10K |
$2.25 |
$3.81 |
+69% |
|
10K–25K |
$1.42 |
$2.15 |
+51% |
|
25K–50K |
$1.02 |
$1.65 |
+62% |
|
50K–128K |
$0.74 |
$1.10 |
+49% |
|
128K+ |
$0.71 |
$1.31 |
+85% |
A previous analysis by Artificial Analysis had estimated the price increase at around 20%, but that test relied on benchmarks rather than real-world workloads. Anthropic has also moved in a similar direction with Opus 4.7, where higher token usage pushed effective costs up by 30% to 40%. With both companies moving closer to potential IPOs, further price increases may remain a realistic scenario.
The effective cost increase of GPT-5.5 shows that nominal pricing alone no longer gives an accurate picture of model economics. For businesses and developers, real-world token behavior — especially output length across different task types — is becoming just as important as the official price per million tokens.
ES
EN