ML//model//GPT//o3

o3 is OpenAI's second-generation reasoning model (December 2024), which scales extended thinking further than o1. Its compute budget is configurable (low, medium or high thinking time), and more thinking buys better accuracy at a higher cost, which makes the trade between cost and accuracy explicit.

o3 is OpenAI's second-generation reasoning model (December 2024), which scales extended thinking further than o1. Its compute budget is configurable (low, medium or high thinking time), and more thinking buys better accuracy at a higher cost, which makes the trade between cost and accuracy explicit.

A common reading is that it explores several reasoning branches, scores them and keeps the best (tree search), which would explain both the accuracy leap and the cost; OpenAI has not published its mechanism, so o3 is MCTS is an analogy for search and evaluation at inference (Monte Carlo tree search, test-time compute), useful for thinking and unsupported as a description.

Branches can be evaluated by a PRM (scoring each step), by self-evaluation (RLAIF), by voting (NNN answers, the most frequent wins) or by external verification (math, code).

It achieved unprecedented scores on ARC-AGI, the benchmark designed to resist memorization and test genuine reasoning.

DeepSeek R1 matched o1-level with open weights and GRPO, but o3 pushed beyond, suggesting tree search or PRMs add value over group-relative scoring.