The DeepSeek-V4-Pro model leaves its testing phase with better scores on agent tasks, but a significantly higher rate grid than the group's entry-level models.
DeepSeek has generally made available its V4-Pro model (version of August 13), released from preversion with optimizations for agentic uses and software engineering in real-life conditions. The company announces scores of 87.9 on Terminal Bench 2.1 and 62.7 on DeepSWE, two test benches specialized in the autonomy of IA agents.
Revers de la médain: pricing is increasing significantly compared to the cheapest models in the DeepSeek catalogue, with a dynamic price structure according to the high rush hours, which came into effect from 16 August. This increase illustrates a wider trend among suppliers of border models, which are now paying more for access to their most advanced capabilities.
Read the full article on the original source
Go to the source →