DeepSeek7.0 out of 10 took its V4-Flash API out of preview on July 31, publishing DeepSeek-V4-Flash-0731 as the official release. The company said on X that the update keeps the exact same model architecture and size as the April preview — a sparse mixture-of-experts model with 284 billion total parameters and roughly 13 billion active per token — and applies only to the Flash API; the V4-Pro API and DeepSeek’s app and web models are unchanged for now, with an official V4-Pro release still to come.
What changed
DeepSeek said the gains come entirely from re-post-training the existing checkpoint with a pipeline focused on coding, agentic tool use and reasoning, not from a new architecture. The clearest jump is on Terminal-Bench 2.1, a benchmark for complex command-line agent work, where 0731 scores 82.7 versus 61.8 for the Flash preview and 72.1 for V4-Pro-Preview — meaning the smaller, cheaper Flash model now outperforms DeepSeek’s own flagship preview on agentic execution despite activating a fraction of its parameters per token. DeepSeek and third-party trackers reported similar jumps on internal agent benchmarks DeepSWE and DSBench-FullStack. The model retains a 1-million-token context window and three selectable reasoning-effort levels, and ships with a speculative-decoding draft module carried over from the preview checkpoint, so existing serving configurations continue to work.
Pricing unchanged, for now
API pricing holds at the preview’s rates — $0.14 per million input tokens (cache-miss) and $0.28 per million output tokens, among the cheapest access to a top-tier open-weight coding and agent model. DeepSeek has separately flagged a planned peak-pricing window that would double billing for roughly seven hours a day, though no effective date has been set.
Why it matters
The release underscores how quickly open-weight labs are closing the gap with proprietary frontier models on agentic coding tasks, and it does so via cheaper post-training rather than costlier pretraining runs — a pattern that keeps pressure on rival labs’ pricing for agent-oriented API traffic. With V4-Pro’s official release still pending, DeepSeek’s roadmap suggests further agent-benchmark gains are still to come across its full model line.