OpenAI Releases o3 and o4-mini Reasoning Models Publicly

What Happened

OpenAI released its o3 and o4-mini reasoning models to paying subscribers and API developers, making the systems publicly available for the first time. The release represents OpenAI's most advanced reasoning-focused models to date, succeeding the earlier o1 and o3-mini series that the company had made available in limited or preview form.

Background

OpenAI has maintained a separate model line dedicated to extended chain-of-thought reasoning since the launch of its o1 model in late 2024. These models are designed to spend additional compute time working through problems before producing a response, a process the company describes as "thinking." The approach differs from the company's flagship GPT series, which prioritises conversational speed and general capability over deliberate, step-by-step problem solving.

The o3 model had been previously announced and demonstrated in research contexts, where OpenAI reported high scores on a range of established benchmarks including mathematics competitions, coding evaluations, and scientific reasoning tasks. The o4-mini was positioned as a faster and more cost-efficient alternative to o3, targeting developers and use cases where response latency and API cost are significant factors.

OpenAI is headquartered in San Francisco and was founded in 2015. The company operates as a capped-profit entity and counts Microsoft among its major investors. Microsoft has integrated OpenAI models into its Azure cloud platform and consumer products including Bing and Copilot.

Model Capabilities

According to OpenAI, o3 achieves state-of-the-art results on several third-party benchmarks at the time of release. The company reported the model performs strongly on tasks requiring multi-step mathematical reasoning, competitive programming, and PhD-level science questions as measured by the GPQA benchmark.

The o4-mini model, while smaller in scale, was reported by OpenAI to exceed the performance of the earlier o3-mini on most benchmarks while offering lower per-token pricing through the API. Both models support tool use, including web browsing, code execution, and image analysis within the same reasoning process, a capability OpenAI described as allowing the models to "think with tools."

OpenAI stated that o4-mini in particular showed strong results on mathematics benchmarks, including the AIME competition problems used to evaluate advanced high school and undergraduate mathematics ability.

Access and Pricing

Both models became available through OpenAI's ChatGPT interface for Plus, Team, and Pro subscribers. Access through the API was opened to developers simultaneously. OpenAI set o3 at a higher price point than o4-mini, consistent with the tiered pricing structure the company has applied across its model families.

Free-tier ChatGPT users did not receive immediate access to o3 at launch. OpenAI did not specify a timeline for broader free-tier availability at the time of release.

Market Context

The release came during a period of accelerated model launches across the industry. Google DeepMind has released multiple versions of its Gemini 2.5 Flash model with reasoning capabilities in recent weeks. Anthropic has updated its Claude 3.5 model line. Meta released its Llama 4 Scout and Maverick open-weight models. The pace of releases across major AI developers has increased throughout the first half of 2025.

OpenAI also faces competition in the reasoning model segment from Chinese AI developer DeepSeek, whose R1 model attracted significant attention earlier in 2025 after demonstrating competitive reasoning performance at lower reported training costs.

What Happens Next

OpenAI has indicated it plans to continue releasing models in the o-series line, with a full o3 pro version described as forthcoming for users on higher-tier subscription plans.