A leaked image showed five upcoming OpenAI models. The flagship would get the attention, but the mini and nano prices were the part that changed what I might actually build.
Looking back OpenAI announced GPT-4.1, mini, and nano later on April 14. The leaked prices were accurate. My descriptions of the models below were guesses made before that announcement, not specifications from OpenAI.
Five Models, One Leak
A leaked image hints at five launches:
- GPT-4.1: The new flagship, building on GPT-4o with sharper reasoning.
- GPT-4.1 mini: Mid-sized, balancing cost and performance for most business use cases.
- GPT-4.1 nano: The smallest, targeting edge devices or budget-conscious deployments.
- o3: A separate line from GPT, possibly focused on planning or specialized reasoning.
- o4-mini: A mini version of a model (o4) that hasn’t been fully revealed yet — suggesting either technical hurdles or a staggered rollout.
Pricing and Latency
The leaked image also showed tiered pricing:
| Model | Input (per 1M tokens) | Cached Input | Output (per 1M tokens) | Latency vs GPT-4o |
|---|---|---|---|---|
| GPT-4.1 | $2.00 | $0.50 | $8.00 | Similar |
| GPT-4.1 mini | $0.40 | $0.10 | $1.60 | 40% faster |
| GPT-4.1 nano | $0.10 | $0.025 | $0.40 | 50% faster |
GPT-4.1 nano at $0.10 per million input tokens is 20x cheaper than the flagship. For large-scale AI workloads, that gap changes what’s worth building.
Anthropic and Mistral Got Here First
Anthropic and Mistral had already made model size and price part of the product choice. The flagship 4.1 would probably be impressive, but the mini and nano models mattered more to me: they made high-volume experiments easier to justify.
The o4-mini’s early appearance is odd. My guess: OpenAI wants to test a smaller version before rolling out the full o4, or they’re still fixing the main model.
My read before the announcement
- GPT-4.1: Better multimodal support, stronger reasoning, fewer hallucinations.
- Mini/nano: Speed and affordability, not bleeding-edge capability.
- o3: OpenAI’s answer to planning, logic, and agentic workflows.
- o4-mini: A preview of the next-gen architecture with training wheels on.
This was still a leak, so the model descriptions were my read rather than confirmed specifications. The pricing table was enough to make the direction clear: OpenAI wanted one model family to cover workloads that previously needed very different cost calculations.
One quick signal
Did this earn your time?
Thanks. That gives me something concrete to check.
About the author
KahWee Teng
I’m a staff software engineer in San Francisco. I write from systems I’ve built and maintained: frontend migrations, AI-assisted workflows, and hardware running at home.
