China's DeepSeek has released the finished version of its V4 Pro model, and the company's benchmark numbers suggest it has closed most of the gap with Anthropic's flagship Claude Fable—while costing a fraction of the price.
From Preview to Production
DeepSeek's April preview of V4 Pro drew a wave of scrutiny after independent testers put it through their paces, and the results were unflattering. In those early benchmarks, the model landed roughly 18 points behind Anthropic's top-tier offering, raising questions about whether the Chinese lab could keep pace with the leading American AI developers.
The completed model tells a markedly different story, at least according to DeepSeek's own figures. The company says the gap has narrowed dramatically, with Claude Fable now edging out V4 Pro by only about 5% on the metrics DeepSeek highlights. That kind of leap between a preview and a shipped model is unusual and points to significant refinement during the final development stretch.
A five-percent performance edge is a tough sell when the alternative costs 45 times more.
The Price Argument
The headline pitch isn't raw capability—it's value. DeepSeek argues that Claude Fable's slim performance advantage comes attached to a price tag that is roughly 4,500% higher than V4 Pro's. For businesses running large volumes of AI queries, that math reframes the conversation away from which model tops the leaderboard and toward which one makes economic sense at scale.
DeepSeek has built its reputation on undercutting rivals while staying competitive on quality, a strategy that rattled markets when its earlier models demonstrated that frontier-level results didn't require frontier-level spending. The V4 Pro release leans hard into that same playbook.
Key points from DeepSeek's framing:
- The finished V4 Pro trails Claude Fable by only about 5% on benchmarks.
- The April preview had lagged by roughly 18 points before final tuning.
- Anthropic's model is presented as costing around 45 times more.
As always, vendor-supplied benchmarks deserve caution, and independent evaluations of the production model will be the real test of whether the numbers hold up outside DeepSeek's lab.