Anthropic's Opus 5 ships at half price, drops mandatory data retention

Anthropic made a move this week that should have enterprise buyers paying close attention. They quietly launched Opus 5, their latest flagship model, and it arrives with two surprises baked in: a sticker price that’s roughly 50% lower than the Fable-tier sibling it competes with, and the elimination of mandatory data retention. That second part is the real conversation starter. For regulated industries that’ve been sitting on the sidelines because they couldn’t stomach their prompts living on someone else’s servers, this removes a dealbreaker. No data retention means your queries don’t get logged for training or review unless you explicitly opt in.

I’ve watched enough model launches to know that pricing cuts usually come with fine print — lower rate limits, nerfed context windows, you name it. Anthropic isn’t saying much publicly about what corners they might have cut to hit this price point. But the half-price figure isn’t a promotional gimmick; it’s the baseline. If the benchmarks hold up and the model actually performs at the Opus tier, this puts real pressure on OpenAI’s enterprise pricing. Nobody wants to pay a premium for a model that retains their data by default when there’s a comparable option that doesn’t.

AMD, meanwhile, is taking a different swing at breaking the NVIDIA lock-in. Their ROCm.AI initiative is essentially vibe-coding a path around the CUDA moat — asking developers to let Claude optimize models directly for AMD silicon. It’s an admission wrapped in a product launch: they know their toolchain friction is the real barrier, not raw hardware capability. If AMD can make the optimization step invisible, they don’t need to beat CUDA on features. They just need to make the alternative painless enough that the cost savings actually materialize. That’s a big “if,” but it’s the most honest strategy they’ve tried yet.

The open-weight AI letter circulating among tech leaders is equally revealing for who didn’t sign it. The push to educate policymakers on the value of open models has broad industry support — until you notice certain names missing. I won’t speculate on who, but the absences tell you exactly where the fault lines are. Some companies have built moats they’d rather not see regulators accidentally fill in.