OpenAI just dropped GPT-5.4 mini and GPT-5.4 nano. The benchmarks look great. The pricing? Not so much.
The price jump nobody's talking about
GPT-5.4 mini costs $0.75 per million input tokens and $4.50 per million output tokens. That's roughly 2-3x more expensive than GPT-5 mini, which most developers were using as their "cheap" model for high-volume tasks.
GPT-5.4 nano is even worse comparatively: $0.20/$1.25 per million tokens, which is about 3x the cost of GPT-5 nano.
Here's the pricing side by side:
| Model | Input (per M tokens) | Output (per M tokens) | vs Previous |
|---|---|---|---|
| GPT-5.4 mini | $0.75 | $4.50 | ~2.5x more |
| GPT-5 mini | $0.30 | $1.80 | baseline |
| GPT-5.4 nano | $0.20 | $1.25 | ~3x more |
| GPT-5 nano | $0.075 | $0.40 | baseline |
Why this matters more than you think
If you're running agents, the cost multiplication is compounding. Agents make multiple calls per task. A 10-call agent chain that cost $0.05 on GPT-5 mini now costs $0.12 on GPT-5.4 mini. Run that 100 times a day and you've gone from $5/day to $12/day.
Over a month, that's the difference between a $150 bill and a $360 bill. For the same workload.
The auto-upgrade trap
Here's the part that'll catch people off guard: many tools and frameworks default to the latest model version. If your code points to gpt-5-mini and OpenAI routes that to 5.4, your costs just tripled without you changing a single line of code.
Check your model strings. Pin your versions. Don't let auto-upgrades surprise you.
What you can actually do
1. Pin your model versions explicitly. Don't use gpt-5-mini. Use gpt-5.0-mini-2025-xx-xx or whatever the frozen version string is.
2. Track your spending in real time. Billing dashboards update after the fact. By then the damage is done. I use a live token counter in my Mac menu bar that shows costs as they accumulate. $5 one time purchase. When you can see the meter running, you naturally pace yourself differently.
3. Benchmark whether 5.4 is actually worth it for YOUR use case. The SWE-Bench Pro score went from ~48% to ~54%. That's meaningful for coding tasks. For simple classification or extraction? You're probably paying 3x more for improvements you don't need.
4. Consider mixing models. Use 5.4 mini for tasks where the quality jump matters (complex reasoning, coding). Keep 5.0 mini or nano for high-volume, simpler tasks where the cheaper model is good enough.
The bottom line
Better model, higher price. That's fair. What's not fair is the silent cost increase for anyone using "latest" model routing. Check your configs today before the end-of-month bill teaches you the hard way.
How are you handling the pricing change? Sticking with 5.0 mini or upgrading?
Top comments (0)