What OpenAI is hiding behind the mega sale! 🤑

OpenAI dropped a simple but “wait, what?” kind of a move on Tuesday, and I've been scratching my head over it.

Sam Altman’s crew just torpedoed the prices for their o3 model API, down a whopping 80%.

At $8 per million tokens output, it is now priced the same as the company’s primary coding model the GPT-4.1 and cheaper than the flagship 4o model, which is at $10 right now.

At the same time, OpenAI introduced the o3 Pro for developers, also priced drastically lower than its prohibitively-expensive predecessor in the premium segment, the o1 Pro.

The new model is priced at $20/$80 input-output versus the $150/$600 we were asked to pay for the o1 Pro.

This fire sale, especially for o3, took me by surprise.

GPT-4o handles most things just fine, so why practically give away such a premium model, especially when AI companies are usually crying poor over compute costs?

Part of it is obvious: competition. Google’s Gemini-2.5-Pro has emerged as a developer darling, powering AI agents like Cursor and Windsurf with its combo of smarts and cheap rates. OpenAI had to hit back on price.

But I think there's a bigger game afoot. This isn't just about price wars. It's a signal that foundational AI companies are gearing up to build and deploy AI Agents themselves, aiming to grab a slice of that lucrative secondary market.

LLMs aren't just becoming smarter; they're morphing into full-blown agents, ready to shake things up.

Is It Already the End of AI Agents?

I have frequently argued that pricing is going to play a major role in adoption of LLMs. Incredible LLMs like Claude 4 Opus and GPT-o1-pro haven’t found their due, because of that.

At $8 per Mtok, o3 now outprices Gemini-2.5-Pro, which was ranked by us as at Artificially Boosted as the top LLM overall, all things taken.

But more importantly — o3 is built for agents.

This thing is optimized for tool use, function calls, and tight instruction following — the holy trinity of agentic workflows.

Think about what companies like Perplexity, ManusAI, Replit, and Lovable have done.

They’ve built impressive products and found real markets by constructing intricate systems around base LLMs. This often involves:

• Complex orchestration to manage multi-step tasks.

• Custom "glue code" for calling external tools and APIs.

• External vector databases for memory and RAG (Retrieval Augmented Generation).

• Logic for breaking down user requests into actionable steps.

Now, imagine if the foundational LLM itself comes with much of this "agent-like" intelligence baked in.

o3 can natively handle more complex instruction sequences, manage tool interactions more intuitively, and perhaps even perform rudimentary planning out-of-the-box — thus, the game changes.

It means the heavy lifting traditionally done by these "wrapper" companies could be significantly reduced. Instead of needing to build complex external systems, developers might find that the LLM itself, with clever prompting, can manage more of the agentic workflow. This doesn't mean all external infrastructure disappears, but it does mean:

• Building powerful agents gets easier and cheaper.

• OpenAI (and other foundational model providers) can offer more "complete" solutions directly.

This is how LLMs morph into full-blown agents, potentially disrupting the very secondary market of specialized agent builders that grew up around them. OpenAI isn't just dropping prices; they're signaling an ambition to own more of the agent stack.

Now, I am not saying that this would be the end of agentic apps, but that the goal-post would move, and the infrastructure could again move to be inside of an app store in ChatGPT.

With everything being disrupted this fast, it isn’t alway easy to predict.

Was this email forwarded to you? Subscribe by clicking here.