GPT-5.4 Mini and Nano Pricing: Up to 4x Costlier Than GPT-5

What OpenAI shipped

OpenAI released two smaller models on March 17, 2026: GPT-5.4 mini and GPT-5.4 nano 1. Mini is pitched as a fast model that comes close to flagship performance on coding and computer-use tasks. It has a 400K-token context window and up to 128K tokens of output, and it is available across every ChatGPT tier, including Free and Go 1. Nano is the smallest option. It targets high-volume classification, extraction and routing work, and for now it is available only through the API 1.

Some launch coverage called the pair "near-flagship AI at budget prices." It said mini runs more than twice as fast as its predecessor and scores higher on benchmarks such as SWE-bench Pro and GPQA Diamond 5. That description is accurate in one sense. Compared with OpenAI's flagship models, these are cheaper. Compared with the models they replace, they are not.

The price jump

The published rates are consistent across outlets. GPT-5.4 mini costs $0.75 per million input tokens and $4.50 per million output tokens. Nano costs $0.20 input and $1.25 output 123.

The previous generation was much cheaper. GPT-5 mini listed at $0.25 input and $2.00 output, and GPT-5 nano at $0.05 input and $0.40 output 1. That means:

• Mini input: tripled, from $0.25 to $0.75
• Mini output: up about 2.25x, from $2.00 to $4.50
• Nano input: quadrupled, from $0.05 to $0.20
• Nano output: up roughly 3.1x, from $0.40 to $1.25

The widely repeated "3–4x" figure is correct for input tokens. Output increases are a little smaller, but they are still large. Developers on Reddit noticed quickly. One commenter warned others not to swap the new models into existing configurations without checking costs 3. Others pointed out that OpenAI's pricing pages briefly disagreed with each other, with one still showing the old mini rates before it was updated to the higher figures 3.

Reaction has been split. Developers are pleased with the capability gains and unhappy that the price of entry-level models went up 1.

Where the new pricing still competes

The higher prices do not make these models expensive relative to rivals. One API guide notes that nano's $0.20 input rate undercuts Google's Gemini 3.1 Flash-Lite, and its $1.25 output rate is slightly below Flash-Lite's $1.50. On paper, that makes nano the cheapest model from a major lab 2. The same guide notes the result depends on workload. Output-heavy tasks narrow the advantage 2.

Mini compares less favorably. One Reddit user's comparison listed Gemini 3 Flash at $0.50 input and $3.00 output, below mini on both. That user also noted Google still labels Gemini 3 models as previews 3.

Prompt caching changes the math for some users. Both models charge about 90% less for cached input: $0.075 per million tokens for mini and $0.02 for nano 12. Applications that reuse long system prompts or shared context can recover a meaningful share of the increase. Caching does not help with output tokens, which is where generative workloads spend most of their money.

The flagship context

The small models sit below a GPT-5.4 family that competes closely at the top. One community benchmark roundup, covering the full-size models rather than mini or nano, reported GPT-5.4 Pro leading a weighted SWE-bench score at 88.3% versus 79.3% for Claude Opus 4.6. On standard SWE-bench, Claude edged ahead at 80.8% to 80.1%, and Gemini 3.1 Pro led ARC-AGI-2 abstract reasoning at 77.1% 4. These are user-compiled figures and should be read cautiously. They do show that no single lab dominates, which limits how much pricing power any of them has.

What this means

The likely explanation is that OpenAI has moved its small models up a tier. GPT-5.4 mini is closer to a lighter flagship than to a commodity model, and OpenAI is pricing it for that capability. For agentic coding or computer-use tasks, where better results reduce retries and human review, faster and stronger output may justify the higher cost.

The picture is worse for nano's intended workloads. Classification and routing are where per-token price matters most and where the quality of a small model is often already sufficient. A fourfold input increase there is a real cost for anyone running large volumes, even if nano still beats competitors on headline rates.

The practical takeaway is straightforward:

1. Benchmark before migrating. Do not assume a model upgrade is free.
2. Measure your input-to-output ratio. It determines how much the increase actually costs you.
3. Use caching wherever possible.
4. Confirm the old GPT-5 mini and nano endpoints remain available for tasks that don't need the improvements.

The "budget" label applies only when these models are compared with flagships, not with their predecessors.