OpenAI o3

OpenAI o3 was announced on 20 December 2024 and released on 16 April 2025 alongside o4-mini. Its importance is easy to state: it was the first widely available model that could call tools in the middle of its own reasoning. Where o1 thought and then answered, o3 could search the web, run Python, read a file or crop and zoom into an image, feed the result back into its chain of thought, and carry on thinking. OpenAI described the pattern as thinking with images, and it is the direct ancestor of everything now sold as an agent. The benchmark record was strong for its moment: 83.3% on GPQA Diamond, 91.6% on AIME 2024, 69.1% on SWE-bench Verified and a Codeforces rating around 2,700 Elo. The preview version had already made headlines in December 2024 by scoring 87.5% on the ARC-AGI semi-private set in its high-compute configuration — at a cost per task that ran into thousands of dollars, a detail that mattered as much as the score. Price was the model's second story. o3 launched at 10 and 40 dollars per million input and output tokens, and on 10 June 2025 OpenAI cut that by 80%, to 2 and 8 dollars, on the same day it shipped the higher-compute o3-pro. A frontier reasoning model at commodity prices did more to spread agentic patterns through the industry than any single capability claim. The line is now being wound down. o3 leaves ChatGPT on 26 August 2026 at the end of a 90-day sunset, and in the API the o3-2025-04-16 and o3-pro-2025-06-10 snapshots shut down on 11 December 2026, alongside older GPT-5 snapshots. Its function has not disappeared, only moved: in the GPT-5 family, reasoning effort and tool use inside the chain of thought are settings rather than a separate model.

▸Videos

No videos yet.

▸Reviews