DeepSeek has released an upgraded V4 Pro model, and the company's own testing shows the finished product performs far better than an April preview that trailed Anthropic's flagship Claude by 18 points in benchmarks. The new numbers also highlight a stark price difference: DeepSeek says Claude Fable is only about 5% better than V4 Pro but costs roughly 4,500% more.
The Preview That Missed the Mark
Back in April, DeepSeek rolled out a preview of V4 Pro. That version scored 18 points behind Anthropic's Claude in benchmark tests, a gap that raised questions about whether DeepSeek could keep pace with the top-tier models. But the company's own numbers for the finished V4 Pro tell a different story. According to DeepSeek, the final model is much stronger than the preview suggested, closing most of that gap.
It's not unusual for preview builds to underperform final releases. But the size of the improvement here is notable. DeepSeek hasn't published the exact benchmark scores, only that the finished V4 Pro is significantly better than the April version. That leaves room for independent verification, but it also suggests the preview wasn't representative of what the company could deliver.
A 4,500% Price Gap
The more striking claim comes down to cost. DeepSeek says Claude Fable, Anthropic's model, is only about 5% better than V4 Pro in performance. Yet it costs 4,500% more. That's a massive price-performance gap, and if it holds up, it would make V4 Pro a far more economical choice for developers and businesses that need serious AI horsepower without the premium price tag.
DeepSeek didn't provide specific dollar figures for either model, so the 4,500% figure is relative. But the implication is clear: for most use cases, the extra 5% of performance from Claude Fable would be hard to justify at that kind of markup. The company is positioning V4 Pro as the value option in a market where top-tier models often come with eye-watering costs.
What the Numbers Don't Say
DeepSeek's claims are based on its own benchmarks, not third-party tests. That's a key caveat. The company has a track record of releasing strong models, but the gap between the April preview and the final version is large enough that outside validation will matter. Independent evaluators will need to run V4 Pro through the same tests to see if the performance really matches up.
There's also the question of what "5% better" means in practice. Benchmarks measure specific tasks, and a 5% edge on a test suite doesn't always translate to a 5% improvement in real-world applications. For some workloads, the difference might be negligible. For others, it could be meaningful. DeepSeek's numbers are a starting point, not a final verdict.
The release of V4 Pro is a concrete step, and the company's pricing strategy is aggressive. But the real test will come when independent reviewers get their hands on the finished model. Until then, the 4,500% price gap is a headline, not a proven fact.
