OpenAI's o3 is a reasoning model that thinks before it answers — and on the hardest coding benchmark, SWE-bench Verified, it solves 69.1% of real GitHub issues, up from o1's 48.9%. Its smaller sibling o4-mini hits nearly the same scores for about a quarter of the cost. That's the short answer. The longer answer is more interesting.
Released April 16, 2025, o3 and o4-mini were the moment OpenAI's reasoning line stopped being a research demo and became the default choice for serious applications. The reason is not just higher benchmark numbers. It is that these were the first OpenAI models that could call tools — search the web, run Python, read an image — in the middle of their own chain of thought. That changes what you can build.
What Is the OpenAI o3 Reasoning Model?
The OpenAI o3 reasoning model is a large language model that generates an internal chain of thought before producing an answer, trading a few seconds of latency for substantially higher accuracy on hard problems. Released April 16, 2025, it scores 69.1% on SWE-bench Verified and over 2,700 Elo on Codeforces, and it can invoke tools like web search and Python during reasoning rather than only at the end.
The distinction from a standard model like GPT-4o matters. A chat model maps your prompt directly to an answer in one pass. A reasoning model spends "thinking tokens" first — exploring approaches, checking its own work, backtracking — and only then commits. For a factual lookup that is wasted effort. For a multi-step math proof or a gnarly debugging task, it is the difference between a confident wrong answer and a correct one.
o3 succeeded o1 and o3-mini, and shipped alongside o4-mini, a smaller and cheaper reasoning model. Both carry a 200K-token context window and can return up to 100K output tokens — enough to reason through and rewrite an entire codebase file in a single call.
OpenAI o3 vs o4-mini: Reasoning Model Benchmarks Compared
The surprising part of the April 2025 release was how close o4-mini runs to o3 despite costing a fraction as much. On pure math, o4-mini actually edges ahead. o3 keeps its lead on the broadest, hardest tasks — agentic coding, scientific reasoning, and visual problems — but for high-volume production work, the cheaper model is often the smarter default.
| Benchmark | o3 | o4-mini | o1 (prior gen) |
|---|---|---|---|
| SWE-bench Verified (coding) | 69.1% | 68.1% | 48.9% |
| Codeforces (Elo) | ~2,706 | ~2,719 | ~1,891 |
| AIME 2025 (math, w/ tools) | ~98.4% | ~99.5% | ~83% |
| GPQA Diamond (science) | ~83.3% | ~81.4% | ~78% |
| MMMU (visual reasoning) | ~82.9% | ~81.6% | ~77.6% |
| Context window | 200K tokens | 200K tokens | 200K tokens |
Figures are from OpenAI's April 2025 o3 and o4-mini launch benchmarks and prior o1 system-card results. Math and coding scores use the models' tool-enabled high-effort settings; real-world performance varies with reasoning-effort configuration and prompt design.
o3 Reasoning Model Pricing: What It Actually Costs to Ship
Pricing is where reasoning models went from experimental to deployable. In June 2025, OpenAI cut o3 API pricing by 80%, bringing it to roughly $2 per million input tokens and $8 per million output tokens. o4-mini sits below that at about $1.10 input and $4.40 output. The catch: reasoning models generate hidden thinking tokens you pay for as output, so a single "hard" query can burn 5,000–20,000 output tokens before you see a one-paragraph answer.
| Model | Input / 1M tokens | Output / 1M tokens | Best For |
|---|---|---|---|
| o3 | ~$2.00 | ~$8.00 | Hardest agentic + coding tasks |
| o3-pro | ~$20.00 | ~$80.00 | Max accuracy, latency-tolerant |
| o4-mini | ~$1.10 | ~$4.40 | High-volume math + coding default |
| GPT-4o | ~$2.50 | ~$10.00 | Chat, latency-sensitive tasks |
| o1 (prior) | ~$15.00 | ~$60.00 | Legacy — superseded by o3 |
| GPT-4o mini | ~$0.15 | ~$0.60 | Cheapest non-reasoning fallback |
Pricing reflects OpenAI's published API rates following the June 2025 o3 price reduction; figures are approximate and exclude cached-input discounts. Reasoning models bill hidden chain-of-thought tokens as output, so effective per-query cost exceeds the headline output rate.
How to Use the o3 Reasoning Model in AI Applications
The headline capability is tool use inside reasoning. o3 and o4-mini can decide, mid-thought, to run a web search, execute Python, analyze an uploaded file, or even generate an image — then fold the result back into the chain. OpenAI calls this "thinking with images" when applied to vision. For builders, it means a single API call can do what previously required a hand-wired agent loop.
- 01
Tune the reasoning_effort parameter
Low, medium, and high settings trade cost and latency for accuracy. Most production traffic should default to low or medium; reserve high effort for the queries that genuinely need it. The difference can be 3-5x in both token spend and response time.
- 02
Route, don't default
Use a cheap classifier or GPT-4o mini to decide whether a query needs reasoning at all. Sending every request to o3 is the most common way teams blow their budget — most prompts do not need a chain of thought.
- 03
Default to o4-mini, escalate to o3
For coding, math, and structured analysis, o4-mini delivers ~95% of o3's quality at ~25% of the cost. Run it as your baseline and fall through to o3 only when o4-mini fails a validation check or low-confidence signal.
- 04
Budget for hidden tokens
Reasoning tokens are billed but not always shown. Set max_completion_tokens caps and monitor actual output token counts in production — a complex query can quietly cost 10-20x a simple one even at the same model.
What the Reasoning Models Mean for AI Startups and Investors
From where I sit — having backed 65+ companies, many of them AI-native — the o3 generation reset two things. First, it raised the floor on what a thin application layer can do without custom agent infrastructure; a lot of last year's "agent orchestration" startups now compete with a single model call. Second, the 80% price cut made reasoning economically viable at consumer scale, which is why you saw a wave of products switch their default model in mid-2025.
The investing implication is blunt: if your AI startup's moat was "we wrap reasoning behind a nice UI," that moat is now thinner. The durable companies are the ones using these models to do something hard in a specific vertical — where proprietary data, workflow integration, and distribution matter more than the raw model. If you're tracking how AI model capability maps to startup valuations, our AI valuations dashboard follows where the private capital is actually going.
Reasoning models are no longer the expensive option.
At ~$2 per million input tokens and 69.1% on SWE-bench, o3 turned chain-of-thought from a demo into a default.
Route intelligently, default to o4-mini, and reserve o3 for the queries where a wrong answer is expensive.
Track how AI model capability is repricing private markets on the Value Add VC AI valuations dashboard. Originally published in the Trace Cohen newsletter.
Get VC data most people never see — free.
Weekly benchmarks, valuations, and fund data. No spam, unsubscribe anytime.