O4-mini: Compact Reasoning Model Overview

Updated 6 July 2026

O4-mini is a compact, reasoning-oriented model deployed via API-based prompted inference in empirical evaluations.
It demonstrates strong zero-shot performance for travel mode prediction and competitive speed in medical QA, with mixed few-shot improvements.
Evaluations emphasize operational behavior over internal architecture, highlighting task-dependent strengths and interference from domain-specific prompts.

Searching arXiv for papers mentioning o4-mini and related evaluations. o4-mini is represented in recent arXiv literature primarily through empirical evaluations rather than through a standalone architectural report. The available record portrays it as a reasoning-oriented, compact, closed-source model that is deployed through API-based prompted inference rather than parameter updating. In the current literature, it appears in two main roles: as a classifier-like inference engine for travel mode choice prediction from natural-language-encoded survey records, and as a commercial online baseline for medical multiple-choice question answering with Web Search enabled. Across these settings, the most consistently documented strengths are strong zero-shot performance and high response speed, while the most prominent limitations are mixed responsiveness to few-shot or domain-enhanced prompting, incomplete replication metadata, and task-specific evaluation scope (Zhang et al., 20 Jan 2026, Vrettos et al., 24 Jun 2025).

1. Identity and positioning in the literature

The travel-behavior study "TransMode-LLM: Feature-Informed Natural Language Modeling with Domain-Enhanced Prompting for Travel Behavior Modeling" characterizes o4-mini as a “reasoning-specialized” or “reasoning-oriented” model and contrasts it with GPT-4o and GPT-4o-mini as “general-purpose” models. In its model-selection framing, the O-series models “represent third and fourth generation reasoning frameworks explicitly optimized for multi-step logical inference.” Within that study, o4-mini is treated not as an explanation generator for human inspection but as a classifier-like predictor that maps a prompted textual trip description to a transportation mode label. The implementation is in Python using the OpenAI API, and the paper states that “the o-series models (o3-mini and o4-mini) do not support temperature adjustment as their reasoning architecture uses fixed sampling parameters” (Zhang et al., 20 Jan 2026).

The medical QA study "Accurate and Energy Efficient: Local Retrieval-Augmented Generation Models Outperform Commercial LLMs in Medical Tasks" positions OpenAI’s o4-mini as a strong closed-source reference model and describes it as a cost-efficient reasoning model. In that paper, it appears as o4-mini-Search, with Web Search activated, and serves as one of two commercial online comparators against locally deployed medical RAG systems. The comparison is therefore not merely model-centric; it is framed around privacy, controllability, auditability, and environmental efficiency in hospital-deployable settings (Vrettos et al., 24 Jun 2025).

A recurrent feature across both studies is the absence of direct architectural disclosure. Neither paper reports layer counts, tokenizer design, context length, training corpus, optimization schedule, or full prompt dumps for o4-mini. The literature therefore supports an operational description of o4-mini more strongly than a mechanistic one.

2. Reframing structured tasks as text inference

In TransMode-LLM, o4-mini is evaluated on single-record travel mode prediction from structured household travel survey data that have been converted into standardized natural-language narratives. The framework is explicitly a three-phase pipeline: statistical feature analysis, natural-language encoding, and LLM adaptation under zero-shot, one-shot, few-shot, and domain-enhanced prompting. Formally, the paper defines a structured trip record as $X = \{x_1, x_2, \ldots, x_n\}$ , maps it into text through

$D: X \rightarrow \mathcal{T},$

and then models mode prediction as

$P(m|D(X)) = \text{LLM}_{\alpha}(\text{Prompt}(D(X)), \theta),$

where $\alpha$ includes GPT-4o, GPT-4o-mini, o3-mini, and o4-mini, and $\theta$ denotes the learning paradigm. The final decision is printed in the paper as

$\hat{m} = \arg\max_{m \in \mathcal{M} P(m|D(X)),$

with an apparent typesetting omission near the closing brace (Zhang et al., 20 Jan 2026).

The empirical setting is the 2022 NextGen National Household Travel Survey. After a two-stage cleaning process, the final dataset contains 22,868 trips and 85 attributes. The prediction problem is restricted to six modes—Car, Van, SUV/Crossover, Pickup truck, School bus, and Walk—which together cover over 90% of observed trips. The class distribution is strongly imbalanced: Car 37.9%, SUV/Crossover 33.5%, Pickup truck 10.7%, Van 6.0%, Walk 6.8%, School bus 2.1%. This imbalance motivates the use of F1-macro and F1-weighted in addition to accuracy.

The selected input is a 15-variable subset obtained through literature-based candidate identification and feature importance aggregation across nine methods. The selected features are travel distance, travel duration, trip purpose, age, gender, driving license status, employment status, household size, number of vehicles, homeownership, household income, urban/rural designation, population size, rail availability, and gasoline price. The travel paper also provides a fixed narrative style for presenting these features, for example: a traveler’s age, gender, driver status, household composition, income, urban context, rail access, trip purpose, distance, time, and gasoline price are rendered into a single contextual description ending with the question of the “most likely transportation mode” (Zhang et al., 20 Jan 2026).

This formulation is significant because it turns a conventional tabular discrete-choice-style problem into prompt-based text classification. A plausible implication is that o4-mini is being tested less for generic language generation than for its ability to recover structured behavioral regularities from linguistically encoded context.

3. Zero-shot and few-shot performance in travel mode choice

Within the reported travel-mode experiments, o4-mini is the best-performing zero-shot LLM in every sample-size condition. The paper evaluates sample sizes of 100, 200, 500, and 1,000, using identical test sets within each sample-size configuration for fair comparison. Zero-shot accuracy for o4-mini is 0.4500, 0.5000, 0.5600, and 0.5750, respectively. In the same four settings, GPT-4o records 0.4000, 0.4750, 0.4400, and 0.4500; GPT-4o-mini records 0.3500, 0.4750, 0.4000, and 0.4350; and o3-mini records 0.3500, 0.4500, 0.4400, and 0.4500. Relative to classical baselines, o4-mini loses to LogitBoost at sample size 100 but surpasses both LogitBoost and Gradient Boosting at 200, 500, and 1,000 (Zhang et al., 20 Jan 2026).

Sample size	Zero-shot accuracy	F1-macro / F1-weighted
100	0.4500	0.2887 / 0.4058
200	0.5000	0.3308 / 0.4745
500	0.5600	0.4030 / 0.5261
1000	0.5750	0.4479 / 0.5416

These F1 values matter because they indicate that o4-mini’s zero-shot accuracy is not explained merely by majority-class prediction. At sample size 1,000, for example, GPT-4o zero-shot records F1-macro 0.2825 and F1-weighted 0.2991, GPT-4o-mini records 0.2724 and 0.2890, and o3-mini records 0.3228 and 0.3029, all substantially below o4-mini’s 0.4479 and 0.5416. The paper therefore presents o4-mini as the strongest zero-shot model not only in raw accuracy but also in weighted and, increasingly with larger samples, macro-balanced performance.

Few-shot learning produces a more nuanced pattern. The best few-shot configuration without domain enhancement improves on zero-shot at sample sizes 100, 200, and 500, but not at 1,000. The exact best no-domain few-shot results are: 0.5000 with 5 examples at 100; 0.5500 with 3 examples at 200; 0.5800 with 5 examples at 500; and 0.5400 with 1 example at 1,000. The corresponding changes relative to zero-shot are $+11.1\%$ , $+10.0\%$ , $+3.6\%$ , and $-6.1\%$ . The F1 behavior is similarly mixed: at 1,000, few-shot slightly improves F1-macro from 0.4479 to 0.4647 while reducing accuracy from 0.5750 to 0.5400 and F1-weighted from 0.5416 to 0.4913. The study uses this pattern as a counterexample to the assumption that more in-context examples always help (Zhang et al., 20 Jan 2026).

4. Prompt engineering, domain scaffolds, and interference effects

The prompting conditions tested for o4-mini span three broad families. First, zero-shot prompting supplies the task framing and the natural-language trip description without demonstrations. Second, one-shot and few-shot prompting supplies 1, 2, 3, 5, or 10 demonstration examples drawn from training data via stratified sampling that prioritizes mode diversity. Third, domain-enhanced prompting augments the prompt with standardized mode definitions inspired by the National Household Travel Survey and a structured three-step decision process consisting of Feasibility Check, Contextual Analysis, and Mode Selection (Zhang et al., 20 Jan 2026).

The domain scaffold is transportation-specific. In Step 1, Feasibility Check, the model reasons about required speed from distance and time and rules out physically implausible modes. In Step 2, Contextual Analysis, it considers factors in order of empirical importance derived from the feature-selection stage. In Step 3, Mode Selection, it integrates the evidence and selects the most practical mode while respecting constraints such as “school bus only for school-related trips.” The paper frames this as a domain-enhanced chain-of-thought style scaffold, but it does not print the exact prompt template, system message, JSON schema, or output-parsing protocol.

For o4-mini, domain-enhanced prompting is not a consistent benefit. The best few-shot accuracies without and with domain enhancement are as follows:

Sample size	Best few-shot without domain	Best few-shot with domain
100	0.5000 (5 examples)	0.4500 (1 example)
200	0.5500 (3 examples)	0.5750 (3 examples)
500	0.5800 (5 examples)	0.5500 (2 examples)
1000	0.5400 (1 example)	0.5250 (10 examples)

Thus domain enhancement helps o4-mini only at sample size 200 and lowers accuracy in the other three conditions. The F1 evidence is similarly mixed and often negative. At 100, domain enhancement lowers accuracy and F1-weighted but slightly increases macro F1; at 200, accuracy and macro F1 improve slightly while weighted F1 declines slightly; at 500 and 1,000, all reported metrics worsen, with especially sharp degradation in macro F1 at 1,000. The paper contrasts this with GPT-4o, which improves in accuracy in all four sample sizes under domain enhancement.

The authors interpret this architecture-dependently. They argue that reasoning-specialized models “might be exploiting specialized computational pathways that get disrupted by any external domain information,” and they suggest that such models develop “highly structured internal representations and inference mechanisms.” On that view, standardized definitions and explicit stepwise scaffolds can introduce “conflicting signals” or distractors rather than useful constraints. This suggests that, for o4-mini in this task family, simple zero-shot or lightly few-shot prompting may be preferable to elaborate domain injection (Zhang et al., 20 Jan 2026).

5. Medical question answering, speed, and efficiency trade-offs

A second empirical portrait of o4-mini comes from medical multiple-choice question answering. The evaluated benchmark contains 1,000 questions with answers: 500 MedQA questions and 500 PubMedQA questions. The setup is zero-shot inference only; the paper does not describe a train/validation/test split or any supervised adaptation. For both o4-mini and DeepSeekV3-R1, Web Search was activated. The local comparison systems are textbook-grounded RAG pipelines built over the Cecil Textbook of Medicine, with FAISS indexing and several open-access generator models (Vrettos et al., 24 Jun 2025).

The paper reports the following values for o4-mini-Search: Accuracy 57.0% $D: X \rightarrow \mathcal{T},$ 0, Precision 52%, Recall 62%, F1 56.8%, Latency per question 0.5 sec/Q, Throughput 2 Q/sec, PPW 0.19, and CO2 footprint 1140 g. Although the table formatting is corrupted in the energy fields, the discussion reconstructs total energy at approximately 3.0 kWh over the 1,000-question benchmark, based on an external assumption of 3 Wh per prompt and Germany carbon intensity 0.38.

Metric	o4-mini-Search	llama3.1:8B-RAG
Accuracy	57.0%	58.5%
F1	56.8%	57%
Latency per question	0.5 sec/Q	2.99 sec/Q
Throughput	2 Q/sec	0.33 Q/sec
PPW	0.19	0.52
CO2 footprint	1140 g	473 g

Relative to the best local system, llama3.1:8B-RAG, o4-mini is slightly worse in accuracy and markedly worse in reported energy efficiency, but substantially faster. The paper states that llama3.1-RAG significantly outperforms o4-mini in accuracy under a Wilcoxon signed-rank test with $D: X \rightarrow \mathcal{T},$ 1. It also reports that llama3.1-RAG achieved “2.7x more accuracy points per kWh and 172% less electricity usage than o4-mini while achieving higher accuracy,” although the percentage phrasing is mathematically nonstandard and the paper’s own reconstructed arithmetic is more consistent with about 63.3% lower electricity use and about 58.5% lower CO2 emissions.

This medical benchmark establishes a different facet of o4-mini than the travel paper. In travel behavior modeling, o4-mini is the strongest model in the reported zero-shot conditions. In medical QA, it remains competitive and fast, but it is outperformed by a domain-grounded local RAG on both accuracy and efficiency. The strongest supported interpretation is therefore task-conditional rather than universal.

6. Evidence boundaries and comparative context

The literature on o4-mini remains operationally informative but mechanically incomplete. In the travel study, the exact train/test split proportions are not reported; there are no appendices with o4-mini prompt dumps, outputs, or case studies; and the paper does not evaluate calibration, uncertainty, cost, latency, or robustness to prompt order. It also restricts the task to six mode classes in a U.S. car-oriented context and does not report subgroup performance across demographics. In the medical study, energy for commercial models is not directly measured but estimated from external assumptions; the exact prompts, temperature, top- $D: X \rightarrow \mathcal{T},$ 2, max tokens, system prompts, and answer-format normalization procedures are not reported; and there are no MedQA-versus-PubMedQA or Step 1-versus-Step 2 breakdowns for o4-mini (Zhang et al., 20 Jan 2026, Vrettos et al., 24 Jun 2025).

These gaps matter because they delimit what can be claimed. The current arXiv evidence supports statements about observed behavior in prompted inference settings, not about internal architecture, training data provenance, calibration properties, or deployment robustness under distribution shift. A plausible implication is that reproducibility for o4-mini currently depends more on high-level protocol reconstruction than on exact prompt-level replication.

A useful comparative backdrop is provided by the broader compact-model literature. The "Phi-4-Mini Technical Report" explicitly states that Phi-4-Mini is not “o4-mini,” but presents it as a relevant comparison point for the broader category of small but high-capability reasoning models. Phi-4-Mini is described there as a 3.8-billion-parameter decoder-only Transformer with 128K context, 200,064-token vocabulary, and Group Query Attention, while a reasoning-trained variant reaches AIME 50.0, MATH-500 90.4, and GPQA Diamond 49.0. That paper is therefore best read not as documentation of o4-mini itself, but as evidence that the research community increasingly studies compact models through the lens of disproportionately strong reasoning and task efficiency (Microsoft et al., 3 Mar 2025).

Taken together, the available literature presents o4-mini as a model whose distinctive empirical profile lies in strong zero-shot reasoning-oriented inference, high speed in online QA settings, and a nontrivial susceptibility to prompt-induced interference when explicit domain scaffolding is imposed. The strongest presently supported conclusion is therefore specific: o4-mini is best understood as a high-performing compact reasoning model in API-mediated inference workflows, with strengths and failure modes that are highly dependent on task framing and prompting regime.