Evidence
- Sep 20Larger model Occamy was said to compete with on some benchmarks
- Sep 9Customers who tuned prompts on V4 Pro object to silent rerouting to a different model
- Aug 22DeepSeek V4 Pro (via Claude Code, max) scored 3,205 (12.3% gap closed) in 1.1 agent-days
- Aug 20Commenter says decompiling with LLMs works until closed models refuse, and open-weights models are catching up
- Aug 16Production build V4 Pro 0813 released to API and chat platforms with a major agent upgrade
Show all 10
- Aug 16On Artificial Analysis it sits below Kimi K3, ties GLM 5.2, and trails leading closed models by a few points
- Aug 16About 6 cents per task, far cheaper than Kimi K3 and GPT 5.6; only model in the high-intelligence, low-cost quadrant
- Aug 12Own chart puts it level with Fable 5 on Terminal Bench and ahead of Opus 4.8, at a fraction of the cost
- Aug 12Via Hermes agent fixed a broken FastAPI/Redis/SQLite report app end-to-end with no hints, using few tokens, though it took ~16 minutes
- Jul 20DeepSeek V4 reportedly about to leave preview for general availability