The rumor: DeepSeek is A/B testing an unreleased model that reportedly beats V4 on coding, reasoning, and context handling, with a stated focus on visual coding workflows and backend efficiency. No benchmarks, no release date, no official word from DeepSeek.
Why it's worth talking about anyway
Every DeepSeek release so far has followed the same shape: ship an open-weight model that's good enough, price it at a fraction of what closed frontier labs charge, then let the market do the work of proving it out. V3 undercut GPT-4o on cost while beating it on several benchmarks. V4 did the same against Claude- and GPT-5-class competition. If a V5 keeps that pattern — competitive quality, open weights, aggressive pricing — the interesting part isn't the model. It's what it does to inference pricing across the board.
Every time an open-weight model closes the gap with closed frontier models, the cost floor for serving frontier-tier capability drops for everyone building on top of it, including the providers hosting it. That's the actual thing worth watching here, independent of whether this specific leak turns out to be real.
We'll know more once there's an actual release with actual benchmarks. Until then: rumor, not roadmap.