GPT-5.4 Mini and Nano Pricing: Small Models Get Pricier

What OpenAI shipped, and what it costs

OpenAI's new small models, GPT-5.4 mini and GPT-5.4 nano, are being marketed in a familiar way. They are described as fast, efficient, and close to the flagship model's performance without the flagship's price. The mini model is aimed at harder work such as coding and reasoning, while nano is pitched at simpler jobs like sorting information and quick data extraction 5. OpenAI says the mini runs more than twice as fast as its predecessor and scores higher on benchmarks including SWE-bench Pro and GPQA Diamond 5.

The list prices are $0.75 per million input tokens and $4.50 per million output tokens for mini, and $0.20 input and $1.25 output for nano 12. Cached input drops to $0.075 for mini and $0.020 for nano 2. Taken alone, those numbers look modest, and some coverage has called the pricing "friendly" 5.

The comparison with the previous generation is less flattering. GPT-5 mini cost $0.25 per million input tokens and $2.00 per million output tokens. GPT-5 nano cost $0.05 input and $0.40 output 2. That makes the new mini three times more expensive per input token and the new nano four times more expensive 12. Output prices rose by a similar margin 1.

A broken pattern

The price increase matters because it reverses a long-running expectation. From 2023 through 2025, each new generation of small models arrived dramatically cheaper than the one before 2. Developers came to treat that decline as something they could plan around. The usual habit was to build on today's small model and assume the next one would be both better and cheaper.

That assumption no longer holds. Developers who sized their pipelines around GPT-5 mini prices are not getting a free upgrade. They are facing a repricing 1. Hacker News commenters were quick to set the old and new rates side by side 2, and the price gap has drawn attention across developer communities 1.

The case that it's still a bargain

Not everyone sees the change as a step backward. Supporters judge the models on what they can do for the money, not only on generation-over-generation cost.

Simon Willison, a widely read commentator on AI releases, calculated that nano can describe 76,000 photos for $52 1. For vision-heavy workloads handling millions of images, that kind of cost per task may matter more than where a model ranks on a leaderboard 1.

A DEV Community breakdown compares the new models with Anthropic's Claude Haiku 4.5, priced at $1.00 input and $5.00 output. By that measure, mini is somewhat cheaper than Haiku, and nano is roughly four times cheaper on input and three times cheaper on output 4. The same piece offers a practical rule of thumb [4]:

• Nano for repetitive, high-volume work such as classification, log parsing, and bulk extraction
• Mini for tasks that involve interacting with user interfaces or computer use

There is also a consumer angle. Free ChatGPT users now have access to the mini model, which approaches the flagship on several benchmarks, including coding and computer use 2. Before this, free users had no direct access to GPT-5.4-level capability 2.

Pricing keeps moving underneath

Any given price snapshot seems unlikely to last long. A CloudZero overview of OpenAI's lineup describes a later pricing ladder in which GPT-5.4 mini and nano are labeled legacy tiers 3. In that ladder, a GPT-5.6-family model called Luna costs $0.20 input and $1.20 output after a July 30 price cut, which is roughly four times cheaper than GPT-5.4 mini 3. A "balanced" tier called Terra, at $2 input and $12 output, undercuts the standard GPT-5.4 at $2.50 and $15 3.

The full spread runs from $0.20 to $30 per million input tokens 3. The same overview notes that GPT-5.4 doubles its input rate for long-context use 3. Its first cost-saving recommendation is to move workloads to new tiers whenever prices change 3.

The takeaway for product teams

One reading reconciles these sources. OpenAI appears to be pricing small models by the value of their capabilities rather than lowering costs automatically each generation. Mini and nano cost more because they do more, and OpenAI seems to be betting that customers will pay for that.

If that is the strategy, the effect on planning is significant. The steady price decline that made "wait for the next model" a dependable cost plan has become uneven. In this case, cheaper options arrived through a separate product line and a later price cut, not through the natural successor to the old small models 3.

For product managers, the practical lesson is to stop treating model costs as a line that only goes down. Teams should budget per task instead of per token, compare prices across vendors, and plan to re-check which tier each workload uses every time OpenAI changes its prices. Based on the past year, that will happen often.