OpenAI o3
OpenAI's full o3 release in April 2025, after a December 2024 teaser. Step-change on math, code, and frontier reasoning benchmarks; the first model to convincingly beat the ARC-AGI visual-reasoning challenge (long held up as a "true general intelligence" benchmark). Native tool use during reasoning — the model decides mid-thought to run code, search the web, or call external APIs.
How are Intelligence, Speed & Cost bucketed?
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
o3 set the high-water mark for frontier capability in 2025 and became the reference benchmark for "what's the best AI can do?" debates in policy, education, and labor-market analysis. The ARC-AGI result specifically reignited the AGI-timeline discussion across AI safety circles.
Core Capabilities
Context Window
Availability
Pricing Model
Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
What it feels like
- Step-change on ARC-AGI-1 — 75.7% at the public $10K compute limit, 87.5% at high compute (172x)
- But ARC-AGI-2 (March 2025) cratered the narrative: <3% on the harder benchmark vs 60% for the average human
- The released April-2025 o3 is NOT the December preview — preview had ARC-AGI-1 in its training, public version is more honest
- TechCrunch reported the released o3 scored lower than initial OpenAI claims implied
- Reasoning trace is much faster than o1 and the honesty/calibration of the chain improved noticeably
- Still expensive at high effort; still slow on the hardest tasks
Best use cases
- Frontier reasoning research and benchmark exploration
- Hard math (AIME, IMO-style) and graduate-level science (GPQA)
- Code that needs novel algorithm design rather than scaffolded refactors
- Tool-using agents that benefit from longer reasoning windows
Tools to try
Not ideal for
- Latency-sensitive chat — o3-mini or GPT-4o are far faster
- Tasks framed slightly outside training distribution — Apple's variable-rename study still applies
- Bulk inference — high-effort o3 is one of the most expensive options on the market
Model Evolution
o is OpenAI's language model family.