On April 16, 2025, OpenAI released o3 and o4-mini, two reasoning models in its o-series that think before answering and, for the first time in that family, decide on their own when to search the web, run Python, inspect images or call developer-defined functions while they reason. o3 became OpenAI's most capable model for hard coding, math and visual problems, while o4-mini offered much of that reasoning at a fraction of the price. For developers, the launch marked the point where "reasoning model" and "agent" started to mean the same thing.

Key Facts#

  • Release date: April 16, 2025. API model IDs are o3 and o4-mini, with snapshots o3-2025-04-16 and o4-mini-2025-04-16.
  • Tool use inside reasoning: OpenAI says these are its first reasoning models that can agentically use and combine every tool in ChatGPT, including web search, Python analysis, image understanding and image generation.
  • Thinking with images: the models can crop, zoom and rotate uploaded images as part of their chain of thought, rather than only describing them.
  • Vendor-reported results: on OpenAI's 477-task SWE-bench Verified subset, o3 scored 69.1% and o4-mini 68.1%, versus 48.9% for o1. With a Python interpreter, o4-mini reached 99.5% pass@1 on AIME 2025.
  • Launch pricing per million tokens: o3 at $10 input and $40 output, o4-mini at $1.10 and $4.40. OpenAI cut o3's price by 80% to $2 and $8 in June 2025, when it released o3-pro.
  • Limits: 200,000 input tokens and up to 100,000 output tokens, according to Microsoft's Azure OpenAI documentation.
  • Developer access: Chat Completions and the Responses API, with organization verification required for some developers.
  • Also announced: Codex CLI, an open-source coding agent for the terminal, plus a $1 million fund offering grants of $25,000 in API credits.

What Happened#

OpenAI described o3 as its most powerful reasoning model across coding, math, science and visual perception. The company said external experts found that o3 made 20% fewer major errors than o1 on difficult real-world tasks, with the biggest gains in programming, business consulting and creative ideation. o4-mini was pitched as the efficient option: a smaller model with strong math, coding and vision results and much higher usage limits, making it suitable for high-volume workloads that still benefit from reasoning.

The more important change was how the models use tools. OpenAI trained both with reinforcement learning not just to call tools, but to decide when a tool is worth calling. In ChatGPT, that meant a single question could trigger several web searches, a Python forecast and a generated chart, with the model changing course as results arrived. OpenAI said most such answers still arrive in under a minute.

In ChatGPT, Plus, Pro and Team users got o3, o4-mini and o4-mini-high in place of o1, o3-mini and o3-mini-high, with Enterprise and Edu following a week later. In the API, OpenAI highlighted Responses API features that matter for agents: reasoning summaries, and the ability to keep reasoning tokens around function calls so the model does not lose its train of thought between tool invocations. The company also said built-in tools such as web search, file search and code interpreter would follow inside the model's reasoning.

On safety, OpenAI said it rebuilt its refusal training data, added a reasoning-based monitor for biological risk prompts, and evaluated both models under its Preparedness Framework. It concluded that neither crossed the framework's "High" threshold in biological and chemical risk, cybersecurity or AI self-improvement.

Background#

OpenAI's o-series began with o1 in 2024, followed by o3-mini on January 31, 2025. Those models traded latency for accuracy by spending extra compute on internal reasoning, but they were awkward partners for tools. With o3, OpenAI reported that reinforcement learning scales the way pretraining does, with more compute producing better results, and it pushed training and inference-time compute up by another order of magnitude.

The release came just two days after GPT-4.1, which covered fast, non-reasoning workloads. Together, the two launches gave API customers a clear split: GPT-4.1 for speed and volume, o3 and o4-mini for problems that need deliberate, multi-step thinking. OpenAI's own "What's next" note already pointed to merging the two lines into one family, a promise it kept with GPT-5 in August 2025.

Why It Matters for Developers#

Reasoning models that choose their own tools change how you build agents. Instead of orchestrating every step in code, you describe the tools well and let the model plan. OpenAI's function-calling guide for these models recommends spelling out tool order for critical workflows, stating when not to call a tool, and remembering that system messages are treated as developer messages by o-series models. Those are design decisions, not afterthoughts, and our guide to AI agent architecture patterns covers when to let the model plan versus when to hard-code the flow.

In .NET, Microsoft.Extensions.AI exposes reasoning effort through ChatOptions.Reasoning, and the OpenAI adapter maps it to the API's reasoning effort setting. Combined with automatic function invocation, a reasoning agent needs little code:

C#
using Microsoft.Extensions.AI;
using OpenAI.Chat;

string apiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY")
    ?? throw new InvalidOperationException("Set OPENAI_API_KEY.");

IChatClient client = new ChatClient("o4-mini", apiKey)
    .AsIChatClient()
    .AsBuilder()
    .UseFunctionInvocation()
    .Build();

var options = new ChatOptions
{
    Reasoning = new ReasoningOptions { Effort = ReasoningEffort.High },
    Tools = [AIFunctionFactory.Create(GetOrderStatus)],
};

var response = await client.GetResponseAsync(
    "Is order 4711 eligible for a refund? Check its status first.", options);
Console.WriteLine(response.Text);

static string GetOrderStatus(string orderId) =>
    orderId == "4711" ? "Delivered on 2025-04-02" : "Unknown order";

Cost needs deliberate attention. Reasoning tokens are billed as output tokens, so a short visible answer can hide a long, expensive thought process. Track token usage per request and per feature, as described in our guide to observability and cost control for LLM apps, and default to o4-mini unless evaluations show that o3 is worth the premium.

The "thinking with images" capability also matters for business software. Invoices, whiteboard photos, architecture diagrams and low-quality scans are everyday inputs, and a model that can zoom and rotate before answering handles them more reliably. See multimodal AI in .NET for patterns that pass images safely through your pipeline. Finally, treat the benchmark figures above as OpenAI's own measurements, and confirm them against your own tasks, and your function calling schemas, before switching production traffic.

What's Next#

OpenAI said at launch that it would converge the o-series reasoning skills with the conversational and tool-use strengths of its GPT models. That happened on August 7, 2025, when GPT-5 folded fast answers and deeper reasoning into one system. o3-pro followed on June 10, 2025, alongside the o3 price cut, and Codex CLI grew into a larger coding-agent product.

As of September 2026, o3 and o4-mini remain in OpenAI's SDK model list and in Azure's model catalog, but they are no longer the frontier. Teams that built on them should keep their tool schemas and evaluation suites, because those assets carry over to newer reasoning models with little change.

Sources#