The case that open-weight models have gotten too cheap to ignore

James O'Claire argues that open-weight models have gotten so cheap that the pricing of closed frontier models is starting to look like luxury positioning rather than a fair reflection of cost. His example: running DeepSeek V4 costs roughly $0.09 where a comparable Anthropic call runs around $5.00, close to a 50-fold gap. When the cheaper option is also good enough for most work, he says, the premium needs a better justification than brand.

His wider point is about where the floor is heading. Open-weight releases like DeepSeek V4, Xiaomi's Mimo, and Google's Gemma 4 keep pushing capable models out for free, and a smaller set of fully open projects, where the training data and pipeline are public too, go further. He singles out Allen AI's OLMo line and the NSF and Nvidia partnership behind it as the direction that matters, because reproducible data is what lets others build on the work rather than just download weights.

The piece is an opinion, and prices move week to week, so read the specific numbers as a snapshot. The argument underneath is harder to dismiss: if open models cover most tasks at a fraction of the price, closed labs may lean on warnings about safety and on policy restrictions to defend margins rather than on being clearly better. You can read the full post on his blog.

Why it matters

If you are choosing a model for production, the question is no longer only which is best but which is good enough at the price. For most workloads an open-weight model may now be the default, and a closed model the exception you justify case by case.