The most useful thing I’ve learned watching this latest round of open model releases is that “open” has stopped predicting anything. It used to be a reliable shorthand for cheap and second-tier. Now the top-ranked open coding model in the world costs roughly thirteen times more per million tokens than the cheapest credible open model sitting right next to it, and both hand you the weights. Openness, capability, and price have come apart into three separate questions, and I keep seeing teams answer one of them and assume they’ve answered all three. That’s how you end up with a procurement decision built on buzz and hype.
I’d run these four checks before deciding anything:
- Measure cost per completed task rather than per token, because a model that lists at half the price can land within ten percent once you count the calls it takes to finish the job.
- Find out what share of your input is genuinely cacheable, since the spread between cached and uncached is wide enough to design your prompts around: Caching is an architecture decision, not a billing footnote. Cached input costs one tenth of uncached input.
- Ask the provider what happens at peak, because a model that can’t serve you is the same as a model you don’t have. Open rarely means self-hosted. Most teams rent models or run a smaller derivative, so open weights move your pricing leverage, not your stack.
- Read the actual license instead of the word “open,” which grants nothing on its own. Open weights is not one license. Read the specific terms before you plan around them.
Subscribe to our weekly newsletter