GPT-5.4 Mini and Nano Pricing: Faster Models, Higher Bills

What OpenAI shipped

OpenAI has released GPT-5.4 mini and GPT-5.4 nano, two smaller versions of its GPT-5.4 flagship aimed at high-volume, latency-sensitive work. The company says the models answer up to twice as fast as the earlier GPT-5 mini variants while keeping much of the larger model's reasoning ability 3. OpenAI positions mini for harder jobs such as coding and reasoning. Nano is meant for lighter tasks like classification and quick data extraction 5.

The speed claims hold up in community testing. Reported throughput is roughly 180–190 tokens per second for mini and about 200 for nano. The older GPT-5 mini managed around 55–60 tokens per second at normal priority 3. Mini also posts higher scores on benchmarks including SWE-bench Pro and GPQA Diamond 5.

The pricing problem

The cost is where opinions split. Some coverage calls the new rates friendly: $0.75 per million input tokens for mini and $0.20 for nano 5. That reading only works if you compare against the flagship. GPT-5.4 itself costs $2.50 input and $15.00 output per million tokens 1.

Against the models they replace, the new tiers cost more [2]:

Model	Input ($/M)	Output ($/M)
GPT-5 mini (prev.)	$0.25	$2.00
GPT-5.4 mini	$0.75	$4.50
GPT-5 nano (prev.)	$0.05	$0.40
GPT-5.4 nano	$0.20	$1.25

The increases work out as follows:

• Mini: input tripled and output rose 2.25 times.
• Nano: input quadrupled and output roughly tripled.

Cached input helps a little, at $0.075 for mini and $0.02 for nano 2. That discount only matters for workloads that reuse long prompts. This gap explains the developer pushback 2. Small models are usually chosen because they are cheap. When the budget tier triples in price, it starts to look like a mid-tier product.

How it stacks up against rivals

Against competing small models, the picture is more mixed than the generational comparison suggests.

Claude Haiku 4.5. Mini looks strong here [2]:

• It is cheaper, at $0.75 versus $1.00 input and $4.50 versus $5.00 output.
• It has twice the context window, 400K versus 200K tokens.
• It scores far higher on computer use, 72.1% versus 50.7%.

Gemini 3 Flash. Google's model costs less at $0.50 input and $3.00 output. It offers a 1M-token context window and edges mini on GPQA Diamond, 90.4% to 88.0%. Mini is faster, though, at roughly 180–190 tokens per second against Flash's ~130 2.

Bargain-tier options. Further down the price scale, Gemini 3.5 Flash-Lite lists at $0.30 input and $2.50 output. DeepSeek's V4-Flash charges just $0.14 input and $0.28 output 1.

In practice, GPT-5.4 mini competes as a capable mid-range model, not as a rock-bottom one. Teams running millions of simple calls have cheaper alternatives.

The security angle

The speed gain matters most for security tooling. At these throughput levels, small models become practical for inline jobs such as pull-request scanning and log triage without slowing pipelines down 3. In agentic setups, a Codex-style system can let GPT-5.4 handle planning and final judgment while handing narrower subtasks to cheaper, faster subagents 3.

The same speed carries a risk. A compromised prompt or poisoned repository could spread insecure code patterns quickly across an automated pipeline 3. Faster review loops are only as trustworthy as the inputs they process.

The small-model launch also fits a wider push by OpenAI into cybersecurity. On April 14, the company announced GPT-5.4-Cyber, a variant described as "cyber-permissive" and tuned for defensive security work. It lowers refusal thresholds for legitimate tasks 4. Access runs through an expanded Trusted Access for Cyber program, and its top tiers are reserved for defenders who verify their identity with OpenAI 4. For now, the model is limited to vetted vendors, organizations and researchers 4.

Our reading

In our view, OpenAI is repricing its small-model tier around capability rather than cost. Mini and nano are clearly better and much faster than their predecessors, and mini outperforms Haiku on several fronts. But calling them budget models glosses over a 2–4x rise in per-token cost.

For developers, the practical question is whether the speed and benchmark gains offset the higher bill:

• Latency-critical work such as agent loops and inline security scanning may justify the upgrade.
• Bulk, low-complexity tasks will likely move to Gemini Flash-Lite, DeepSeek, or remaining stock of the older GPT-5 tiers.