Last chance! 50% off DataCamp Premium
Sale ends in
OpenAI recently introduced o4-mini, a new reasoning model designed to be faster and cheaper than o3, while still supporting tools like Python, browsing, and image inputs. Unlike o3-mini, it’s also multimodal by default.
In this article, I take a closer look at what o4-mini can actually do. I ran a few practical tests across math, code, and visual input, and I’ve included OpenAI’s official benchmark numbers to show how it stacks up against models like o3 and o1.
You’ll also find details on pricing, availability, and when it makes sense to choose o4-mini over other models.
We keep our readers updated on the latest in AI by sending out The Median, our free Friday newsletter that breaks down the week’s key stories. Subscribe and stay sharp in just a few minutes a week:
What Is O4-Mini?
O4-mini is OpenAI’s most recent small-scale reasoning model in the o-series. It’s optimized for speed, affordability, and tool-augmented reasoning, with a 200,000-token context window and support for up to 100,000 output tokens—on par with o3 and o1.
Source: OpenAI
Where it differs most is in cost and deployment efficiency. o4-mini runs at $1.10 per million input tokens and $4.40 per million output tokens, compared to o3’s $10.00 input and $40.00 output pricing—a roughly 10x cost reduction for both input and output. As we’ll see in the benchmarks section, o4-mini offers comparable performance to o3.
o4-mini API pricing (left) vs. o3’s pricing (right)
Tool-wise, o4-mini is compatible with Python, browsing, and image inputs, and it integrates with standard OpenAI endpoints, including Chat Completions and Responses. Streaming, function calling, and structured outputs are supported. Fine-tuning and embeddings are not.
o4-mini vs. o4-mini-high
If you’ve opened the ChatGPT app, you might have noticed the option to choose between o4-mini and o4-mini-high. o4-mini-high is a configuration of the same underlying model with increased inference effort. That means:
- It spends more time internally processing each prompt,
- Generally produces higher-quality outputs, especially for multi-step tasks,
- But with slower response times and potentially higher token usage.
If you need faster completions, o4-mini may be the better choice. For more complex reasoning (especially if the reasoning needs to be done over coding or visual input), longer contexts, or precision-critical use cases, o4-mini-high could give you better outputs.
Testing O4-Mini
Let’s test o4-mini on a bit of math, coding, and multimodal input.
Math
First, I wanted to start with a simple calculation that typically confuses language models. I wasn’t trying to test basic arithmetic skills—instead, I wanted to see whether the model would choose the right method to solve it, either by reasoning step by step or by calling a tool like a calculator.
The output was wrong on the first attempt. I suggested using the calculator, and while the second answer was correct, there were still two issues. First, it described the result as “approximately,” even though no rounding or estimation was needed for this subtraction.
Second, its reasoning showed that it hadn’t actually used the calculator, despite the output claiming “The calculator shows” (which misrepresents how the result was obtained). Oddly, it also searched the web, which seems unnecessary for a basic math problem.
Next, I tested the model on a more challenging math problem, and the result was solid this time. It responded quickly and solved the problem using a short Python script, which I could see directly in its chain of thought. Being able to inspect the code as part of the reasoning process is a genuinely useful feature.
Generating a short P5js game
A few weeks ago, when I tested Gemini 2.5 Pro, I generated a quick p5.js game to see how it handled creative coding tasks. I wanted to run the same prompt with o4-mini, but this time I used o4-mini-high. The result wasn’t too far off from Gemini 2.5 Pro, which is great given that this is a smaller model.
The prompt was: Make me a captivating endless runner game. Key instructions on the screen. p5js scene, no HTML. I like pixelated dinosaurs and interesting backgrounds.
And the first result:
That dinosaur looks more like an old film camera to me, but the rest is great.
Multimodal reasoning with large input
Stanford recently released The 2025 AI Index Report, a 450-page (117,649 tokens) overview of the current state of the AI industry. I asked o4-mini-high to read it and predict 10 trends for 2026. It handled the task well and completed it in just 9 seconds.
Here are the trends it predicted for 2026:
- Near‑Human Performance on New Reasoning Benchmarks
- Ultra‑Low‑Cost Inference for Production LLMs
- Synthetic‑Plus‑Real Data Training Pipelines
- Widespread Adoption of AI‑Native Hardware
- Multimodal Foundation Models Standardized
- Responsible‑AI Certification and Auditing Regimes
- AI‑Driven Scientific Discovery Accelerates Nobel‑Winning Work
- Localized AI Governance Frameworks with Global Alignment
- Democratization of AI Education and Workforce‑Readiness
- AI Agents as Collaborative Knowledge Workers
Most of these made sense, though a few predictions lean toward optimism. Take number two—yes, inference costs are dropping, but “ultra-low” is a stretch, especially for larger models and less-optimized deployments.
Number four is similar: AI-native hardware is progressing, but widespread adoption depends on supply chains, software integration, and ecosystem support, which all take time.
As for number nine, we’re still quite far from anything that could be called full democratization of AI education, but to be fair, the model didn’t say “complete.”
O4-Mini Benchmarks
O4-mini was designed to be a fast, cost-efficient model, but it still performs well across a broad set of reasoning benchmarks. In some cases, it even surpasses larger models like o3.
Math and logic
These tests focus on structured reasoning, often requiring multiple steps to solve math problems correctly. The AIME benchmarks, for instance, are drawn from competitive high school math contests.
- AIME 2024: o4-mini (no tools) scores 93.4%, outperforming o3 (91.6%), o3-mini (87.3%), and o1 (74.3%).
- AIME 2025: o4-mini (no tools) scores 92.7%, again beating o3 (88.9%), o3-mini (86.5%), and o1 (79.2%)
Source: OpenAI
Coding
These benchmarks test the model’s ability to write, debug, and understand code. Codeforces is a competitive programming dataset, while SWE-Bench and Aider Polyglot simulate real-world engineering tasks.
- Codeforces: o4-mini (with terminal) scores an ELO of 2719, slightly above o3 (2706). o3-mini scores 2073, and o1 lags far behind at 1891.
- SWE-Bench Verified: o4-mini scores 68.1%, just behind o3 at 69.1%, and ahead of o1 (48.9%) and o3-mini (49.3%).
- Aider Polyglot (code editing): o4-mini-high scores 68.9% (whole) and 58.2% (diff). For comparison, o3-high scores 81.3% (whole) and 79.6% (diff).
Source: OpenAI
Multimodal reasoning
These benchmarks measure how well the model processes visual inputs like diagrams, charts, and figures. The tests combine image understanding with logical reasoning.
- MMMU: o4-mini scores 81.6%, close to o3 (82.9%) and better than o1 (77.6%).
- MathVista: o4-mini scores 84.3%, trailing o3 (86.8%) but significantly above o1 (71.8%).
- CharXiv (scientific figures): o4-mini reaches 72.0%, behind o3 (78.6%) but above o1 (55.1%).
Source: OpenAI
General QA and reasoning
These datasets test broad question-answering ability, including PhD-level science questions and open-ended reasoning across subjects. Tool use is often required to score well.
- GPQA Diamond: o4-mini scores 81.4%, behind o3 (83.3%) but above o1 (78.0%) and o3-mini (77.0%).
- Humanity’s Last Exam:
- o4-mini (no tools): 14.28%
- o4-mini (with Python + browsing): 17.70%
- For context, o3 (with tools) scored 24.90%, and Deep Research scored 26.60%.
Source: OpenAI
O4-Mini vs O3: Which One Should You Use?
O4-mini and o3 are based on the same underlying design philosophy: models that can reason with tools, think step-by-step, and work with text, code, and images. But they’re built for different trade-offs. In short:
- Pick o3 if precision is your top priority and cost is less of a concern.
- Pick o4-mini if speed, volume, or cost-efficiency matter more than squeezing out a few extra benchmark points.
O3 is stronger overall
O3 scores higher than o4-mini on benchmarks like CharXiv, MathVista, and Humanity’s Last Exam. It also shows better consistency on complex tasks that require longer reasoning chains or precise code generation. If you’re working on research problems, scientific workflows, or anything where accuracy matters more than speed or cost, o3 is likely the safer pick.
O4-mini is much cheaper and faster
The difference in price is significant: o4-mini costs $1.10 for input and $4.40 for output, compared to o3 at $10 and $40 respectively. That’s nearly a 10x difference across the board. You also get faster response times and higher throughput, which makes it a better fit for production environments, real-time tools, or anything running at scale.
Multimodality is not a deciding factor
Both models are multimodal and tool-capable. Both support Python, browsing, and vision. If you’re just looking for that feature set, either one will work.
How to Access o4-mini in ChatGPT and API
You can use o4-mini in both ChatGPT and the OpenAI API, depending on your plan.
In ChatGPT
If you’re using ChatGPT with a Plus, Pro, or Team subscription, o4-mini is already available. You’ll see it listed as “o4-mini” and “o4-mini-high” in the model selector.
These replace the older o3-mini and o1 options. Free users can also try o4-mini in a limited form by selecting “Think” mode in the composer before sending a prompt. Note that:
- o4-mini is the default fast version.
- o4-mini-high runs with increased inference effort for better quality at the cost of speed.
In the API
o4-mini is also available through the OpenAI API under the Chat Completions and Responses endpoints. Both versions (standard and high) support tool use, including browsing, Python, and image inputs, depending on how you’ve configured your toolset.
The model ID for the current snapshot is o4-mini-2025-04-16 . Pricing is set at $1.10 per million input tokens and $4.40 per million output tokens, with caching and batching supported. If you’re working with the Responses API, you also get reasoning token visibility and more control over tool execution within multi-step chains.
Conclusion
o4-mini offers solid performance across math, code, and multimodal tasks, while cutting inference costs by an order of magnitude compared to o3. The benchmarks show that it holds up well, even in tasks where models typically struggle, like multimodal reasoning or coding.
It’s also the first time a mini model from OpenAI includes full tool support and native multimodality. That alone puts it ahead of o3-mini and o1, both of which it replaces in most ChatGPT tiers. And with the addition of o4-mini-high, users get a bit more control over the balance between speed and output quality.
Whether o4-mini is the right model depends on the task. It won’t beat o3 in terms of raw capability, but for anything that needs to run quickly, scale easily, or stay within a tighter budget, it’s a strong option—and in many cases, good enough.
