GPT-4o-mini
GPT-4o-mini is a lightweight large language model released by OpenAI, representing a compact version of the GPT-4o series. The model retains strong natural language understanding and generation capabilities while offering lower inference costs and faster response speeds, making it suitable for latency- and cost-sensitive applications such as text classification, summarization, dialogue systems, and multi-agent collaboration tasks.
Core Facts
OpenAI推出GPT-4o mini作为对轻量级高性能模型市场需求的回应
Model routing strategies can assign simple subtasks to smaller models (such as GPT-4o-mini) and delegate only complex reasoning to larger models in order to reduce costs.
GPT-4o is priced approximately 50% lower than GPT-4.
Timeline (last 90 days)
The GPT-4o-mini few-shot configuration had only one confidence value before and after calibration, and its correctness AUROC was only 0.500, indicating that the model has almost no ability to distinguish correct from incorrect predictions.
Introducing an auxiliary model (such as GPT-4o-mini) to summarize tool outputs brings 500ms to 2 seconds of external network latency as well as ongoing API call costs.
The routing layer can be divided into three tiers: the local/edge tier (such as Llama 3 8B, Mistral 7B, cost about $0.05–$0.15/1M tokens, latency <250ms), the mid-tier hosted tier (such as Claude 3.5 Haiku, GPT-4o-mini, cost about $0.25–$0.60/1M tokens, latency <800ms), and the frontier reasoning tier (such as Claude 3.5 Sonnet, o1, GPT-4o, cost about $3.00–$15.00/1M tokens, latency 2s–8s).
Rules generated by models with ≥3B parameters through GRPO training can serve as transferable knowledge to improve the downstream performance of GPT-4o-mini.
For a premium reasoning model costing $0.015 per query, a $0.00002 embedding call corresponds to an economic break-even hit rate below 1%; when routing to lightweight models (such as GPT-4o-mini or Claude 3.5 Haiku) with an inference call cost of $0.0002, an application needs a 15% to 20% cache hit rate to avoid losing money.
GPT-4o-mini scores about 18% accuracy on the same HS code benchmark; the best AI Agent with web search reaches 46.8%, but costs $1-10 per call and takes 10-30 seconds; human experts are about 95%.
The study evaluated fifteen pipeline configurations, ultimately producing 18,630 paired completion results generated by GPT-4o-mini
The expected cost of routing through the small model is Cs + Cl × f; the critical fallback rate is f = (Cl − Cs) / Cl; using GPT-4o-mini ($0.15/million tokens) and GPT-4o ($2.50/million tokens) as examples, the critical fallback rate is approximately 94%, but after incorporating the hidden cost of latency, the practical threshold is typically 30–40%.
The Terra and Sol models showed the most merciful behavior in HarvestBench tests (lowest kill rates), while GPT-4o-mini showed the most ruthless behavior (highest kill rate)
Model routing strategies can assign simple subtasks to smaller models (such as GPT-4o-mini) and delegate only complex reasoning to larger models in order to reduce costs.
3 more timeline events
All Facts (13)
OpenAI推出GPT-4o mini作为对轻量级高性能模型市场需求的回应
80%VerifiedModel routing strategies can assign simple subtasks to smaller models (such as GPT-4o-mini) and delegate only complex reasoning to larger models in order to reduce costs.
65%VerifiedGPT-4o is priced approximately 50% lower than GPT-4.
65%UnverifiedAI系统即使代码未变，仅更换底层模型版本（如从GPT-4o切换到GPT-4o-mini）或微调提示词，都可能导致输出行为发生显著变化
90%UnverifiedThe GPT-4o-mini few-shot configuration had only one confidence value before and after calibration, and its correctness AUROC was only 0.500, indicating that the model has almost no ability to distinguish correct from incorrect predictions.
50%UnverifiedIntroducing an auxiliary model (such as GPT-4o-mini) to summarize tool outputs brings 500ms to 2 seconds of external network latency as well as ongoing API call costs.
50%UnverifiedThe routing layer can be divided into three tiers: the local/edge tier (such as Llama 3 8B, Mistral 7B, cost about $0.05–$0.15/1M tokens, latency <250ms), the mid-tier hosted tier (such as Claude 3.5 Haiku, GPT-4o-mini, cost about $0.25–$0.60/1M tokens, latency <800ms), and the frontier reasoning tier (such as Claude 3.5 Sonnet, o1, GPT-4o, cost about $3.00–$15.00/1M tokens, latency 2s–8s).
50%UnverifiedRules generated by models with ≥3B parameters through GRPO training can serve as transferable knowledge to improve the downstream performance of GPT-4o-mini.
50%UnverifiedFor a premium reasoning model costing $0.015 per query, a $0.00002 embedding call corresponds to an economic break-even hit rate below 1%; when routing to lightweight models (such as GPT-4o-mini or Claude 3.5 Haiku) with an inference call cost of $0.0002, an application needs a 15% to 20% cache hit rate to avoid losing money.
50%UnverifiedGPT-4o-mini scores about 18% accuracy on the same HS code benchmark; the best AI Agent with web search reaches 46.8%, but costs $1-10 per call and takes 10-30 seconds; human experts are about 95%.
50%UnverifiedThe study evaluated fifteen pipeline configurations, ultimately producing 18,630 paired completion results generated by GPT-4o-mini
50%UnverifiedThe core idea of model routing is that not all tasks require the most powerful model; simple tasks can use lightweight models such as GPT-4o-mini, Claude 3 Haiku, or Llama 3.1 8B, potentially at just 1/50 or even 1/100 the cost of flagship models.
50%UnverifiedOpenAI的GPT-4o与GPT-4o-mini之间的每百万Token价格差距可达10倍以上
50%