OpenAI Releases GPT-5.4 Mini and Nano, Which Could Be More Useful Than the Big Model

OpenAI isn't slowing down. Less than 2 weeks aft launching GPT-5.4—itself released conscionable 2 days aft GPT-5.3—the institution dropped 2 much models connected Tuesday: GPT-5.4 Mini and GPT-5.4 Nano.

These aren't stripped-down versions of the flagship model—they're purpose-built machines designed for the benignant of enactment wherever waiting fractional a infinitesimal for an reply is not an option.

OpenAI calls them its “most susceptible tiny models yet,” saying that GPT-5.4 Mini is much than 2 times faster than GPT-5 Mini. If you've ever watched a coding adjunct deliberation for 45 seconds earlier editing 3 lines of code, past you recognize the entreaty of a accelerated model.

So wherefore would anyone merchandise a little close exemplary connected purpose? The abbreviated answer: due to the fact that accuracy isn't ever the bottleneck. If you're moving a lawsuit work chatbot that answers the aforesaid 200 questions each day, past you don't request the exemplary that scored champion connected PhD-level chemistry exams. You request the 1 that responds successful nether a 2nd and costs a fraction of a cent per reply. That's the abstraction these models are built for.

But it doesn’t mean these models are dumb oregon unreliable. On coding benchmarks, GPT-5.4 Mini scored 54.4% connected SWE-Bench Pro—a trial that measures a model's quality to hole existent GitHub issues—compared to 45.7% for the aged GPT-5 Mini and 57.7% for the afloat GPT-5.4.

On OSWorld-Verified, which tests however good a exemplary tin really run a desktop machine by speechmaking screenshots, Mini deed 72.1%, conscionable shy of the flagship's 75.0%—and some wide the quality baseline of 72.4%. GPT-5.4 Nano, meanwhile, scores 52.4% connected SWE-Bench Pro and 39.0% connected OSWorld—lower than Mini, but inactive a large leap implicit erstwhile Nano-class models.

"GPT-5.4 marks a measurement guardant for some Mini and Nano models successful our interior evaluations,” Perplexity Deputy CTO Jerry Ma said aft investigating both. “Mini delivers beardown reasoning, portion Nano is responsive and businesslike for unrecorded conversational workflows."

Instead of routing each azygous task done an costly flagship model, you tin present physique systems wherever the large exemplary plans and coordinates portion smaller models grip the existent grunt enactment successful parallel—searching a codebase here, speechmaking a papers there, oregon processing a signifier determination else. As we saw successful our GPT-5.4 vs. Grok 4.20 comparison, wherever the exemplary sits successful the workflow matters arsenic overmuch arsenic which exemplary you pick.

GPT-5.4 Mini runs astatine a complaint of $0.75 per cardinal input tokens and $4.50 per cardinal output tokens via the API. GPT-5.4 Nano is adjacent cheaper: $0.20 per cardinal input tokens and $1.25 per cardinal output tokens—a terms constituent that makes moving a immense magnitude of queries per time financially realistic for startups. For context, Nano is astir 4 times cheaper than Mini connected inputs.

For regular ChatGPT users, GPT-5.4 Mini is disposable contiguous to Free and Go users via the "Thinking" enactment successful the positive menu. Paid subscribers who deed their GPT-5.4 complaint limits volition automatically autumn backmost to Mini. GPT-5.4 Nano, however, is API-only for now—OpenAI is intelligibly positioning it arsenic a developer tool, not a user one.