OpenAI ships GPT‑5.4 mini/nano
What happened
OpenAI released GPT‑5.4 mini and nano, described as its most capable small models yet — cheaper and faster inference options published March 17. The new small models will lower inference costs for many use cases while leaving high‑throughput training demand intact for custom and agentic models. (9to5mac.com)
Why it matters
GPT‑5.4 mini is being served inside ChatGPT plus other product flows while GPT‑5.4 nano is positioned as an API‑only option for high‑throughput programmatic usage. (openai.com/index/introducing-gpt-5-4-mini-and-nano/, macobserver.com/news/openai-adds-gpt-5-4-mini-to-chatgpt-nano-goes-api-only/) OpenAI published live benchmark tables showing GPT‑5.4 scored 57.7% on SWE‑Bench Pro, GPT‑5.4 mini scored 54.4%, and GPT‑5.4 nano scored 52.4%, placing both small models close to flagship performance on coding tasks. (openai.com/index/introducing-gpt-5-4-mini-and-nano/) The company says mini is tuned for coding, tool use, multimodal reasoning and mid‑length contexts while nano targets short‑turn tasks like classification, extraction, ranking, and lightweight sub‑agent execution. (openai.com/index/introducing-gpt-5-4-mini-and-nano/, techcommunity.microsoft.com/blog/introducing-openai%E2%80%99s-gpt-5-4-mini-and-gpt-5-4-nano-for-low-latency-ai/4500569) API pricing tables position full GPT‑5.4 at roughly $2.50 per million input tokens and $15 per million output tokens while third‑party aggregators list GPT‑5.4 mini around $0.75/$4.50 per million and nano near $0.20/$1.25 per million, creating a multi‑tier cost ladder for routing simple work to cheap small models. (openai.com/api/pricing/, openrouter.ai/openai/gpt-5.4-mini/, llmbase.ai/news/openai-gpt-5-4-mini-and-nano-faster-models-for-coding-and-subagent-workloads/) OpenAI and industry writers frame mini/nano as deliberate “subagent” building blocks — systems where the flagship handles hard reasoning and the small models execute parallel, inexpensive subtasks — a split that preserves demand for high‑throughput training and larger models for custom or agentic workloads. (openai.com/index/introducing-gpt-5-4-mini-and-nano/, thenewstack.io/gpt-54-nano-mini/, cambrian-ai.com/wp-content/uploads/edd/2025/03/AI-Compute-Workloads-Shift.pdf)
Key numbers
- OpenAI released GPT‑5.4 mini and nano, described as its most capable small models yet — cheaper and faster inference options published March 17.
- (9to5mac.com) GPT‑5.4 mini is being served inside ChatGPT plus other product flows while GPT‑5.4 nano is positioned as an API‑only option for high‑throughput programmatic usage.
- (openai.com/index/introducing-gpt-5-4-mini-and-nano/, thenewstack.io/gpt-54-nano-mini/, cambrian-ai.com/wp-content/uploads/edd/2025/03/AI-Compute-Workloads-Shift.pdf)
What happens next
- The new small models will lower inference costs for many use cases while leaving high‑throughput training demand intact for custom and agentic models.
Sources
Quick answers
What happened in OpenAI ships GPT‑5.4 mini/nano?
OpenAI released GPT‑5.4 mini and nano, described as its most capable small models yet — cheaper and faster inference options published March 17. The new small models will lower inference costs for many use cases while leaving high‑throughput training demand intact for custom and agentic models. (9to5mac.com)
Why does OpenAI ships GPT‑5.4 mini/nano matter?
GPT‑5.4 mini is being served inside ChatGPT plus other product flows while GPT‑5.4 nano is positioned as an API‑only option for high‑throughput programmatic usage. (openai.com/index/introducing-gpt-5-4-mini-and-nano/, macobserver.com/news/openai-adds-gpt-5-4-mini-to-chatgpt-nano-goes-api-only/) OpenAI published live benchmark tables showing GPT‑5.4 scored 57.7% on SWE‑Bench Pro, GPT‑5.4 mini scored 54.4%, and GPT‑5.4 nano scored 52.4%, placing both small models close to flagship performance on coding tasks. (openai.com/index/introducing-gpt-5-4-mini-and-nano/) The company says mini is tuned for coding, tool use, multimodal reasoning and mid‑length contexts while nano targets short‑turn tasks like classification, extraction, ranking, and lightweight sub‑agent execution. (openai.com/index/introducing-gpt-5-4-mini-and-nano/, techcommunity.microsoft.com/blog/introducing-openai%E2%80%99s-gpt-5-4-mini-and-gpt-5-4-nano-for-low-latency-ai/4500569) API pricing tables position full GPT‑5.4 at roughly $2.50 per million input tokens and $15 per million output tokens while third‑party aggregators list GPT‑5.4 mini around $0.75/$4.50 per million and nano near $0.20/$1.25 per million, creating a multi‑tier cost ladder for routing simple work to cheap small models. (openai.com/api/pricing/, openrouter.ai/openai/gpt-5.4-mini/, llmbase.ai/news/openai-gpt-5-4-mini-and-nano-faster-models-for-coding-and-subagent-workloads/) OpenAI and industry writers frame mini/nano as deliberate “subagent” building blocks — systems where the flagship handles hard reasoning and the small models execute parallel, inexpensive subtasks — a split that preserves demand for high‑throughput training and larger models for custom or agentic workloads. (openai.com/index/introducing-gpt-5-4-mini-and-nano/, thenewstack.io/gpt-54-nano-mini/, cambrian-ai.com/wp-content/uploads/edd/2025/03/AI-Compute-Workloads-Shift.pdf)
