DeepSeek's V4 Flash Beats Its Own Flagship on Agent Tasks

DeepSeek moved its V4 Flash model into general availability on July 31, and the notable part is not the launch itself but the scoreboard. After a new round of post-training, the smaller, cheaper Flash model now clears the company's own larger V4 Pro on agent benchmarks. The budget tier beat the premium one.

On Terminal-Bench, which measures how well a model handles multi-step work in a command line, V4 Flash scores 82.7. That is roughly ten points above the Pro version it sits below in the lineup. DeepSeek says the gain came entirely from a fresh training pass on the same underlying model, not a larger or redesigned architecture, which is a pointed reminder that raw scale is not the only lever left to pull.

The price is the argument

V4 Flash costs $0.14 per million input tokens and $0.28 per million output, with cache hits falling to $0.003. Set against the frontier tiers from OpenAI and Anthropic, which run into dollars per million tokens, that is a difference of one to two orders of magnitude. For developers building agents that make many calls in a loop, the running cost often matters more than a few points on a leaderboard, and DeepSeek is aiming squarely at that calculation.

The release adds to a stretch of strong showings from Chinese labs. It lands the same week Alibaba unveiled its 2.4-trillion-parameter Qwen 3.8 Max, and against the backdrop of Washington's debate over whether to restrict Chinese open-weight models that keep closing the gap on cost and capability. DeepSeek is one of the names cited most often in that conversation.

As ever, the benchmark figures are the company's own, and Terminal-Bench captures one slice of what an agent actually needs to do. Independent testing will settle how far the improvement carries. But the headline is hard to miss. A model most people would reach for to save money is now outscoring its pricier sibling on the tasks that matter for building software, and it costs almost nothing to run.