What Are Reasoning Models? The o-Series Explained
From the course OpenAI o-Series: Reasoning Models Masterclass
OpenAI's o-series represents a fundamental shift in how AI models solve problems. Instead of generating answers immediately, these models think first — producing an internal chain of thought before responding.
The Core Idea
Standard models (GPT-4o): Prompt → Answer Reasoning models (o1/o3): Prompt → Think → Think → Think → Answer
This extra thinking time dramatically improves accuracy on hard problems.
The o-Series Family
| Model | Released | Strength | Cost (Input/Output per M tokens) |
|---|---|---|---|
| o1-preview | Sep 2024 | First reasoning model | $15 / $60 |
| o1 | Dec 2024 | Improved reasoning | $15 / $60 |
| o1-mini | Sep 2024 | Fast reasoning | $3 / $12 |
| o3 | Early 2025 | Best reasoning | $10 / $40 |
| o3-mini | Early 2025 | Cost-effective reasoning | $1.10 / $4.40 |
| o4-mini | Mid 2025 | Latest efficient reasoner | $1.10 / $4.40 |
When Reasoning Models Shine
- Mathematics — Competition-level math problems
- Science — Physics, chemistry, biology reasoning
- Coding — Complex algorithms, debugging, architecture
- Logic — Puzzles, constraint satisfaction, planning
- Analysis — Multi-factor decisions, risk assessment
- Legal/Medical — Interpreting complex regulations or symptoms
Benchmark Performance
- AIME 2024 (Math Competition): o3 scores 96.7% vs GPT-4o at 63.6%
- Codeforces (Competitive Programming): o3 reaches 99th percentile
- GPQA Diamond (PhD-level Science): o3 scores 87.7% vs GPT-4o at 53.6%
- ARC-AGI (Reasoning): o3 scores 87.5% vs GPT-4o at 5%
When NOT to Use Reasoning Models
❌ Simple Q&A and factual lookups ❌ Creative writing and brainstorming ❌ Translation and summarization ❌ Casual conversation ❌ High-volume, low-complexity tasks
For these, GPT-4o or GPT-4o-mini is faster, cheaper, and often better.
Cost Considerations
Reasoning models generate "thinking tokens" that you pay for but don't see:
- A simple question might use 500 thinking tokens
- A hard math problem might use 10,000+ thinking tokens
- These hidden tokens significantly increase costs
- Monitor usage carefully in production applications
Key Takeaways
- •Reasoning models think before answering, producing internal chain-of-thought reasoning
- •They dramatically outperform standard models on math, coding, science, and logic
- •o3 scores 96.7% on AIME 2024 vs GPT-4o's 63.6% — a massive improvement
- •Don't use reasoning models for simple tasks — GPT-4o is faster and cheaper
- •Hidden thinking tokens significantly increase costs — monitor usage carefully
Progress is kept in this browser only.