GPT-5.5 Pricing: What It Actually Costs
TH5 09, 2026 · 2 phút đọc
OpenAI launched GPT-5.5 with a 2x price increase over GPT-5.4. Input went from $2.50/M to $5.00/M, output from $15/M to $30/M. OpenAI says the new model is less verbose, so the real cost impact should be smaller than the headline numbers suggest.
We pulled real usage data to find out.
The result: actual costs went up between 49% and 92%. The price hike is partially offset by shorter completions, but only on longer prompts.
How we measured this
We took users whose most-used model was GPT-5.4 before launch, then switched to GPT-5.5 after. Same users, same workflows, different model version.
Both models share the same tokenizer family, so no normalization needed. The comparison is clean.
Completions got shorter, but only on long prompts
| Prompt Size | Median Completion (5.4) | Median Completion (5.5) | Change |
|---|---|---|---|
| Under 2K | 121 | 129 | +7% |
| 2K to 10K | 140 | 213 | +52% |
| 10K to 25K | 211 | 143 | -32% |
| 25K to 50K | 185 | 150 | -19% |
| 50K to 128K | 188 | 136 | -28% |
| 128K+ | 215 | 143 | -34% |
Above 10K tokens, GPT-5.5 produces 19% to 34% fewer output tokens. Below 10K, completions are the same length or longer. The 2K to 10K bucket actually got 52% more verbose.
What it actually costs
Using billed costs across the same cohort, here’s the average cost per million tokens:
| Prompt Size | Avg $/M (5.4) | Avg $/M (5.5) | Change |
|---|---|---|---|
| Under 2K | $4.89 | $9.37 | +92% |
| 2K to 10K | $2.25 | $3.81 | +69% |
| 10K to 25K | $1.42 | $2. | +51% |
| 25K to 50K | $1.02 | $1.65 | +62% |
| 50K to 128K | $0.74 | $1.10 | +49% |
| 128K+ | $0.71 | $1.31 | +85% |
Long prompts get the biggest break because shorter completions absorb part of the price hike. Short prompts take the full hit, since completion length didn’t drop.
Takeaway
If your workload runs on short prompts, GPT-5.5 will roughly double your bill. If you’re running long context tasks, expect closer to a 50% increase. Worth benchmarking your actual usage before assuming the new model is a drop-in upgrade.
Methodology
- Source: MixRoute request logs
- Cohort: Users whose top model was GPT-5.4 before launch, then GPT-5.5 after
- Sample: Text-only, non-cancelled requests
- Windows: GPT-5.4 from April 21 to 23, 2026. GPT-5.5 from April 25 to 28, 2026 (launch day excluded)
- Normalization: Cost per million MixRoute tokens, bucketed by prompt size
- Excluded: Media requests, cancelled requests, zero-token requests