OpenAI Releases GPT-5.4 Mini and Nano, Which Might Be Extra Helpful Than the Large Mannequin – Decrypt

OpenAI is not slowing down. Lower than two weeks after launching GPT-5.4—itself launched simply two days after GPT-5.3—the corporate dropped two extra fashions on Tuesday: GPT-5.4 Mini and GPT-5.4 Nano.

These aren’t stripped-down variations of the flagship mannequin—they’re purpose-built machines designed for the type of work the place ready half a minute for a solution shouldn’t be an possibility.

OpenAI calls them its “most succesful small fashions but,” saying that GPT-5.4 Mini is greater than two instances quicker than GPT-5 Mini. If you happen to’ve ever watched a coding assistant suppose for 45 seconds earlier than enhancing three traces of code, you then perceive the attraction of a quick mannequin.

So why would anybody launch a much less correct mannequin on goal? The quick reply: as a result of accuracy is not all the time the bottleneck. If you happen to’re working a customer support chatbot that solutions the identical 200 questions all day, then you do not want the mannequin that scored greatest on PhD-level chemistry exams. You want the one which responds in below a second and prices a fraction of a cent per reply. That is the house these fashions are constructed for.

However it doesn’t imply these fashions are dumb or unreliable. On coding benchmarks, GPT-5.4 Mini scored 54.4% on SWE-Bench Professional—a check that measures a mannequin’s capability to repair actual GitHub points—in comparison with 45.7% for the outdated GPT-5 Mini and 57.7% for the complete GPT-5.4.

On OSWorld-Verified, which assessments how properly a mannequin can really function a desktop laptop by studying screenshots, Mini hit 72.1%, simply shy of the flagship’s 75.0%—and each clear the human baseline of 72.4%. GPT-5.4 Nano, in the meantime, scores 52.4% on SWE-Bench Professional and 39.0% on OSWorld—decrease than Mini, however nonetheless a significant leap over earlier Nano-class fashions.

“GPT-5.4 marks a step ahead for each Mini and Nano fashions in our inner evaluations,” Perplexity Deputy CTO Jerry Ma stated after testing each. “Mini delivers sturdy reasoning, whereas Nano is responsive and environment friendly for dwell conversational workflows.”

As an alternative of routing each single activity by way of an costly flagship mannequin, now you can construct programs the place the large mannequin plans and coordinates whereas smaller fashions deal with the precise grunt work in parallel—looking a codebase right here, studying a doc there, or processing a kind someplace else. As we noticed in our GPT-5.4 vs. Grok 4.20 comparability, the place the mannequin sits within the workflow issues as a lot as which mannequin you decide.

GPT-5.4 Mini runs at a price of $0.75 per million enter tokens and $4.50 per million output tokens by way of the API. GPT-5.4 Nano is even cheaper: $0.20 per million enter tokens and $1.25 per million output tokens—a worth level that makes working an enormous quantity of queries per day financially sensible for startups. For context, Nano is roughly 4 instances cheaper than Mini on inputs.

For normal ChatGPT customers, GPT-5.4 Mini is accessible in the present day to Free and Go customers by way of the “Considering” possibility within the plus menu. Paid subscribers who hit their GPT-5.4 price limits will routinely fall again to Mini. GPT-5.4 Nano, nonetheless, is API-only for now—OpenAI is clearly positioning it as a developer device, not a client one.