Model components / Fine-tuned / chat models

o3 / o4-mini

Previous reasoning tier (2025): full-scale o3 plus cost-efficient o4-mini, both using extended chain-of-thought. Powered Codex agent and ChatGPT's coding mode through 2025. Folded into the GPT-5 line in late 2025; OpenAI consolidated the o-series and GPT branches and is no longer maintaining o-series as a separately branded reasoning family. Current reasoning capability lives in GPT-5.x with thinking enabled and in GPT-5.3-Codex.

OpenAI o3 and o4-mini reasoning models, both released Apr 16 2025 (o3-mini preceded Jan 31 2025). Post-trained reasoning/instruct models, API-only + ChatGPT, no downloadable weights. Succeeded by the GPT-5.x reasoning line but still selectable. The o4-mini AIME figure in the capability record is not on either cited source and is flagged. Verified 2026-08-13 via the OpenAI o3/o4-mini system card and the Wikipedia article.

Openness

1 high confidence

1.0

weights: closed
data: closed
code: closed
license: Proprietary(API+ChatGPT only)

Proprietary OpenAI reasoning models; no weights, data, or code released.

https://en.wikipedia.org/wiki/OpenAI_o3 recorded 2026-08-13
o3-mini released 2025-01-31, o3 and o4-mini 2025-04-16, o3-pro 2025-06-10; served through ChatGPT tiers and the API, with no weights, corpus or code published
https://cdn.openai.com/pdf/2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf recorded 2026-08-13
"OpenAI o3 and o4-mini System Card, OpenAI April 16, 2025" - still live and still the primary release document; a safety card rather than a weights or data release

Adoption

5 medium confidence

5.0

o3/o4-mini were rolled out as selectable models inside ChatGPT (Plus/Pro and some free) plus the API; ChatGPT reached ~900M WAU (Feb 2026) rising toward ~1B (May 2026). Reach attributed to the ChatGPT surface these models power, not a standalone o3 count. Now superseded as default by GPT-5.x but still used at scale.

https://en.wikipedia.org/wiki/OpenAI_o3 recorded 2026-08-13
"OpenAI released o3-mini to all ChatGPT users (including free-tier) and some API users"; o3 and o3-pro on the paid tiers; no standalone per-model user count published

Capability

4 high confidence

4.0

Frontier-class reasoning at its Apr 2025 launch, state of the art on Codeforces, SWE-bench and MMMU at the time. It has since been surpassed by GPT-5.x, Opus 4.8 (88.6 SWE-bench) and Gemini 3.5, so on a frontier-relative scale it is a 4 rather than a 5. The o4-mini figure of 99.5 does not appear on the Wikipedia source and rests on OpenAI's system card alone; either way the score is set by the o3 figures.

https://en.wikipedia.org/wiki/OpenAI_o3 recorded 2026-08-13
"o3 achieved a score of 87.7% on the GPQA Diamond benchmark"; "On SWE-bench Verified ... o3 scored 71.7%, compared to 48.9% for o1. On Codeforces, o3 reached an Elo score of 2727, whereas o1 scored 1891". No o4-mini AIME figure on the page.
https://cdn.openai.com/pdf/2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf recorded 2026-08-13
primary system card, still resolving; models described as combining "state-of-the-art reasoning with full tool capabilities"

Verified 2026-08-13