Strawberry Is Not the $2,000 Thing

People keep trying to make Strawberry the villain of some apocalyptic pricing tier. It's not.

The $2,000 option — if it materializes — is apparently something more like q* plus Strawberry plus whatever new tool scaffolding they've been quietly assembling, plus GPT-4.5 doing cleanup in the back. An agentic combo. A stack pretending to be a model.

Strawberry on its own adds 10 to 20 seconds to each calculation. That's the tradeoff. You wait longer and you get a model that thought harder about the problem — which is a perfectly reasonable thing to sell, and also a perfectly reasonable thing to mostly not want, depending on the problem.

The latency question is genuinely open right now. There are tasks where 15 extra seconds of inference is irrelevant, because the alternative is a junior engineer taking three hours and still being wrong. There are other tasks where you're in a loop, calling the model 40 times to build something, and now you've added 10 minutes to a process that used to take 90 seconds.

Nobody knows which category most real workloads fall into. OpenAI probably doesn't either.

What's actually interesting is the bundling instinct — the idea that the powerful thing isn't Strawberry reasoning, it's Strawberry reasoning plus agents plus new tools, wrapped in one offering at a price that filters for enterprise buyers who don't read their own invoices. That's not a model release. That's a land grab dressed as a product.

Whether it works depends almost entirely on whether the reasoning steps are actually worth waiting for. Which is the kind of thing you can only find out by shipping it and watching what breaks.