OpenAI 03 is here

The OpenAI o3 reasoning model is the most capable thinking-first AI OpenAI has released to date, and it changes how I approach everything from code review to financial modeling in my consulting work. If you have been waiting for an AI that actually pauses, plans, and verifies before answering, o3 is the upgrade you have been waiting for.

Direct Answer: OpenAI o3 is a frontier reasoning model that uses a private chain-of-thought to break complex problems into sub-steps before responding. It scores 87.7% on GPQA Diamond, 96.7% on AIME 2024, and 71.7% on SWE-bench Verified, which means it now matches or exceeds expert human performance on graduate-level science, advanced math, and real-world software engineering benchmarks.

What Makes o3 Different From GPT-4o and o1

The earlier GPT-4 family answered fast. The o-series thinks first. o3 extends what o1 started by spending more compute at inference time, which OpenAI calls test-time scaling. In practice, that means the model generates internal reasoning tokens, evaluates multiple candidate paths, and then commits to an answer. For a Chartered Accountant like me running through a multi-step tax calculation or a deferred revenue schedule, that extra deliberation is the difference between a confident wrong answer and a correct one.

Three concrete improvements over o1:

• Math: AIME 2024 score jumps from o1's 83.3% to 96.7%, putting o3 in the top 0.1% of human competitors.
• Science: GPQA Diamond rises from 78% to 87.7% — past the average human PhD in the relevant field.
• Coding: SWE-bench Verified climbs from 48.9% to 71.7%, and Codeforces ELO hits 2727, which is grandmaster territory.

The ARC-AGI Breakthrough Everyone Is Talking About

o3 scored 75.7% on the ARC-AGI semi-private evaluation in low-compute mode and 87.5% in high-compute mode. For context, the previous best public model scored around 5%. ARC-AGI was designed by François Chollet specifically to resist memorization and reward genuine abstract reasoning. Crossing 85% is the threshold Chollet himself set for human-level performance on this benchmark. That does not mean o3 is AGI, but it does mean the model can solve novel visual reasoning puzzles it has never seen — a capability previous models genuinely could not.

o3 vs o3-mini: Which Should You Actually Use?

OpenAI shipped two variants. o3 is the full frontier model. o3-mini is a smaller, faster, cheaper sibling with three reasoning effort levels: low, medium, and high. Here is how I decide between them in client work:

• Use o3-mini (low): Quick code refactors, summarisation, drafting GoHighLevel email sequences, anything where latency matters more than depth.
• Use o3-mini (high): Mid-complexity coding, financial spreadsheet logic, structured data extraction. Often beats o1 at a fraction of the cost.
• Use full o3: Research-grade math, scientific analysis, debugging gnarly multi-file codebases, anything where being wrong is expensive.

For most of my 79,000+ students who are building AI workflows, o3-mini at medium effort is the sweet spot — capable enough for real work, fast enough to keep a chat conversational.

Deliberative Alignment: Why o3 Refuses Better

OpenAI introduced a new safety technique with this release called deliberative alignment. Instead of relying purely on RLHF guardrails, o3 reasons through OpenAI's safety policy in its chain-of-thought before answering. On internal jailbreak tests, this approach improved refusal accuracy on borderline prompts and reduced over-refusal on legitimate ones. For developers building consumer products, this matters because it means fewer false positives — the model is less likely to refuse a perfectly reasonable medical or legal question while still blocking actual abuse.

Pricing, Access, and the Compute Cost Reality

Access rolled out in stages. ChatGPT Pro and Team users got o3 and o3-mini first. API access followed for Tier 3+ developers. The pricing is where you need to pay attention: full o3 in high-compute mode on the ARC-AGI benchmark reportedly cost thousands of dollars per task. That is not a typo. Test-time compute scales linearly with the number of reasoning tokens, and o3 in high mode can burn through hundreds of thousands of internal tokens before answering.

Practical implications:

• Do not pipe o3 into a high-volume chatbot. You will go bankrupt.
• Reserve full o3 for batch jobs — overnight research runs, code audits, complex client deliverables.
• Use o3-mini for anything user-facing.
• Cache aggressively. If a question has been asked before, do not pay to reason through it again.

How I Am Using o3 in My Consulting Workflow

I run a Dubai-based AI practice and teach across 74+ courses, so I get to stress-test these models on real client problems. Three workflows where o3 has earned its place:

• Financial model audit: I feed o3 a client's Excel logic and ask it to find circular references, broken formulas, and assumption mismatches. It catches things I miss after eight hours of staring at the same sheet.
• GoHighLevel automation debugging: When a workflow has 40+ steps and is firing in the wrong order, o3 traces the dependency graph and identifies the broken trigger faster than I can.
• Course curriculum design: I ask o3 to find pedagogical gaps in my outlines — concepts I assume students know but have not actually taught yet. Its critique is sharper than any human reviewer I have used.

What o3 Still Cannot Do

It is not magic. o3 still hallucinates citations, still struggles with very long contexts, and still has no memory between sessions unless you build it. It is also slow — a hard problem can take 30 to 90 seconds to answer, which kills conversational flow. And critically, it does not have vision parity with GPT-4o yet, so multimodal workflows still need the older model.

The OpenAI o3 reasoning model marks the moment AI moved from fast pattern-matching to deliberate problem-solving, and the right way to capture that value is to match the model to the task. Your next step: pick one expensive, error-prone task in your workflow this week, run it through o3-mini at high effort, and measure the time saved against the API bill.

Keep Learning

If this was useful, these are worth reading next:

• The Future of Business: Turn Your SOPs into AI Agents (Automate Everything)
• Create 40 social media posts using ChatGPT and Canva in less than 2 minutes
• Or go further with the AI Mastery Course — used by 79,000+ students across 150+ countries.