So OpenAI made o3-mini free. Not free-with-an-asterisk, not "free for the first three tries," actually available to anyone with a ChatGPT account, no credit card, starting January 31st. If you've been paying twenty bucks a month for Plus just to poke at their reasoning models, congratulations, that subscription just got a little less special.
I don't think this is a coincidence and I don't think OpenAI wants you to think too hard about the timing either.
Two weeks ago DeepSeek dropped R1 and the AI world basically had a small stroke. Nvidia lost something like 17% of its value in a single day (January 27th, if you want the exact date to drop into an argument later), which is close to $600 billion in market cap gone before lunch. A Chinese lab had apparently trained a model that competes with OpenAI's o1 for a fraction of the compute cost, and it was open-weight, and it was cheap to run through the API, like, absurdly cheap compared to what OpenAI charges. Italy banned the DeepSeek app outright. The US Navy told its people not to touch it. Half of Twitter (I'm still calling it that, fight me) spent a weekend either panicking or gloating depending on which side of the AI-bubble argument they'd already picked.
And now, right in the middle of all that noise, o3-mini shows up free for everyone. Funny how that works.
I've been messing around with it since Saturday and here's my actual take, not the press-release version: it's good, it's fast, and it's clearly built to answer the DeepSeek question directly rather than to be some huge leap forward on its own. Ask it a multi-step math problem or a gnarly logic puzzle and you can watch it visibly "think" before answering, which is the whole reasoning-model bit everyone's doing now. It's noticeably snappier than full o1 was when that launched back in December. For code it's genuinely useful — I had it refactor a chunk of a Python script I've been neglecting for this very blog's backend (yes, techpad still runs on a pile of duct-taped Python scripts I wrote in 2013, don't judge me) and it caught an off-by-one error I'd walked past twice.
But "free" is doing a lot of work in that sentence. Free users get it with limits — you have to pick the "reason" option in the model picker, and there's a daily cap that Plus and Team subscribers don't have. It's not that OpenAI suddenly got generous. It's that DeepSeek proved you could give away frontier-adjacent reasoning ability without bankrupting yourself, and OpenAI apparently decided the PR cost of looking stingy was higher than the compute cost of not being stingy. That's a rational move! I just wish companies would say that instead of the usual "we're so excited to make this available to more people" copy that shows up in every single one of these announcements. Say the quiet part. We're all adults.
The thing that actually bugs me isn't o3-mini itself, it's the framing everywhere I've read about it this week, which treats this as some triumphant democratization story. It's a competitive response. That's fine! Competition is good, we get free stuff out of it, everybody wins a little. But calling it anything other than "OpenAI got spooked and reacted" feels like it's giving them credit for a decision that DeepSeek basically forced. If R1 hadn't happened on January 20th (the same day as the inauguration, incidentally, which is a wild news cycle to have shared), I don't think we're getting a free o3-mini in early February. Maybe eventually. Not now.
Anyway. I ran it against a handful of the same prompts I threw at DeepSeek R1 last weekend and honestly, for most everyday stuff, the gap is smaller than the discourse would have you believe. R1 is a little more willing to show its full chain of thought without you asking. o3-mini feels more polished on the output side, less rambly. Neither one is going to write your novel for you, and I remain extremely unconvinced by anyone who tells you it can.
I'll probably keep both installed for a while just to compare. Somebody has to.