DeepSeek V4-Flash API live in public beta, agent scores top V4-Pro preview

Built on the same 13B-active architecture and pricing, the 0731 build beats the 49B-active V4-Pro preview on agentic benchmarks while using 12% fewer tokens, per DeepSeek. TerminalBench jumped 56.9→82.7 and DeepSWE hit 54.4%, near Opus 4.8. Scores are DeepSeek's claims, not yet independently verified.
How this story unfolded
3 weeks · 0 reports · 10 community posts · from Jul 18
DeepSeek by email
Get an email when DeepSeek has news
No news that day, no email.
More stories today
- CPUs and the rise of neurosymbolic AI
- Mollick asks for plan on cyber threats from open-weight Mythos/Astra models
- Dwarkesh Patel discusses the implications of continual learning for AI
- NoimosAI launches Social Agent, an AI research agent for social trends
- Stanford researchers deploy 37,000 AI agents as a virtual biotech