OpenAI drops o3-pro, its most powerful reasoning model yet

If you throw a PhD-level physics problem or a gnarly architecture bug at a standard AI, it usually hallucinates an answer in half a second. A reasoning model, on the other hand, sits back and actually thinks.

OpenAI just launched o3-pro, billing it as its most capable AI model to date. It officially replaces o1-pro and is already live for ChatGPT Pro and Team subscribers as of Tuesday. Enterprise and Edu tier users will see it rollout next week. If you build with code, the model also landed in OpenAI's developer API on Tuesday afternoon.

Running it through the API won't come cheap. OpenAI set pricing at $20 per million input tokens and $80 per million output tokens. For context, a million tokens is roughly 750,000 words, which is longer than Tolstoy's entire "War and Peace."

Smarter reasoning with a longer wait

According to OpenAI's changelog, expert reviewers consistently preferred o3-pro over the vanilla o3 across every tested category, with the biggest jumps coming in science, education, coding, business, and writing assistance. Reviewers rated it noticeably higher for accuracy, clarity, and following instructions.

OpenAI also shared some internal benchmark bragging rights. On the AIME 2024 math benchmark, o3-pro scored higher than Google's Gemini 2.5 Pro. On GPQA Diamond, a tough benchmark designed around doctorate-level science questions, it edged out Anthropic's newly released Claude 4 Opus.

The new model can also wield external tools. It can search the live web, dig through uploaded files, analyze visual inputs, run Python code, and tap into ChatGPT memory for personalized context.

The catch? It takes even longer to spit out answers than o1-pro did. On top of the slow response times, o3-pro currently cannot generate images and doesn't work with OpenAI's Canvas workspace. Temporary chats are also temporarily shut off due to a technical bug.

Why it matters

For researchers, software engineers, and data specialists tackling high-stakes problems, o3-pro looks like an immediate upgrade worth leaning into. It is not designed for casual email drafting or everyday chat; it is essentially an expensive digital specialist brought in when getting the wrong answer carries a real cost.

Across the industry, the launch signals that the frontier intelligence race is far from slowing down. Anthropic just dropped Claude 4 Opus, and Google has been touting Gemini 2.5 Pro. By putting o3-pro on the board, OpenAI is fighting hard to keep the benchmark crown in its court.

We think pairing a deep-thinking reasoning engine with native tool access like Python and web browsing is the right move, because reasoning in a vacuum only gets you so far. The lingering problem, however, is practical friction. Between the $80 per million output token price tag and noticeably slower speeds, o3-pro will remain a luxury model that only high-margin business workflows can comfortably justify.