OpenAI o-Series

The OpenAI o-Series is a family of reasoning models developed by OpenAI that generate extended internal deliberation before producing an answer, trading speed for accuracy on complex problems in mathematics, coding, science and visual reasoning. Announced in preview in September 2024 and expanded through 2025, the line was later folded into the unified GPT-5 system and repositioned as a thinking mode rather than a separate model family.^[1][2]

History

OpenAI released o1-preview and o1-mini on 12 September 2024, describing models trained with large-scale reinforcement learning to think for longer before responding. The company reported that o1 scored 83 per cent on the qualifying examination for the International Mathematical Olympiad, against 13 per cent for GPT-4o, making the family the reference point for an emerging reasoning-model race.^[1][2] The full o1 model followed in December 2024, and o3 was announced on 20 December 2024 alongside a safety access programme for external researchers. o3-mini arrived on 31 January 2025 with three adjustable reasoning-effort levels and availability to free-tier ChatGPT users; in February 2025 OpenAI began surfacing summaries of its reasoning process, widely read as a response to the open reasoning traces shown by DeepSeek's R1.^[3][6][7] The full o3 and the smaller o4-mini were released on 16 April 2025, with o3-pro following on 10 June 2025.^[4][5] As the GPT-5 generation matured, the family was progressively retired from ChatGPT, with o3 discontinued on 26 August 2026.^[6]

Key Concepts and Technology

The o-Series popularised test-time compute: spending more computation during inference, not only during training, so a model can explore approaches, check its own work and correct errors before answering. Practical controls included reasoning effort settings that let users trade latency for depth, and native tool use — web search, code execution and image analysis — invoked mid-reasoning. The family's commercial success accelerated a competitive wave: Google added thinking modes to Gemini, Anthropic introduced extended thinking in Claude, and DeepSeek's R1 showed that reasoning performance could be approached with open weights at far lower cost.^[2][6]

Applications and Impact

The family moved thinking models from research demonstrations into everyday products, powering agentic coding tools, data analysis, document review and STEM tutoring. It also reset expectations for benchmark performance and pricing: o-models were significantly more expensive per token than their GPT-4 counterparts, establishing a tiered structure that later generations made automatic. There was no o5 — GPT-5 merged fast and reasoning behaviour behind an internal router, making deliberation a default capability rather than a model choice.^[4][6]

>See Also

• GPT-5
• Reasoning Models
• Test-Time Compute
• DeepSeek
• OpenAI