DeepSeek ships V4-Flash official with 50% API price cuts and no V4-Pro
DeepSeek published the official V4-Flash on July 31, cutting API prices up to 50% and lifting the 304B-parameter model to 82.7 on Terminal Bench 2.1 against the V4-Pro preview's 72.1 — with architecture unchanged from the April preview, so the gain is entirely post-training. That makes agent performance a training-recipe problem rather than a scale problem, and DeepSeek is funding the race with roughly $7.4 billion raised from Tencent and NetEase at a 350 billion yuan valuation. The absent V4-Pro is the tell that the frontier tier slipped.
Source: caixinglobal.com ↗
the Flash model's architecture remains identical to its April preview, with the performance gains driven entirely by extensive post-training.
DeepSeek
Why this matters
- → Post-training alone lifted agent performance 14.6 points, reframing scaling as a training problem.
- → 50% API price cuts intensify margin pressure across frontier labs.
- → V4-Pro's absence suggests DeepSeek's frontier tier hit a wall.
Training trumps scale