DeepSeek V4 Flash Enters Public Beta, Agent Benchmarks Surpass V4 Pro

DeepSeek's V4 Flash model hits public beta with agent scores far exceeding the V4 Pro preview — Terminal Bench 82.7, DeepSWE 54.4 — at just $0.14 per million input tokens.

DeepSeek has officially launched the V4 Flash API in public beta, and the benchmark numbers tell a striking story: the new model dramatically outperforms the V4 Pro preview on every agentic benchmark that matters.

The headline figures: Terminal Bench 2.1 at 82.7, DeepSWE at 54.4, Cybergym at 76.7, and NL2Repo at 54.2. These are code-agent and full-stack development metrics, and they put V4 Flash in a different league from its Pro sibling. The model architecture remains unchanged from Flash Preview — 284 billion total parameters with 13 billion active via mixture-of-experts — meaning all gains came from re-post-training alone.

Independent analysis by Artificial Analysis places V4 Flash at an Intelligence Index of 49.9, ranking it ahead of DeepSeek V4 Pro (44.3) and just behind Gemini 3.6 Flash (50.1). Context: Claude Opus 5 leads the index at 60.7, while GPT-5.6 Sol scores 58.9. But price tells the sharper story — at $0.14 per million input tokens and $0.28 per million output, V4 Flash is the cheapest model per intelligence task across the entire leaderboard, costing just $0.027 per evaluation. Cache hits drop to $0.003 per million tokens, a 98% discount.

DeepSeek is shipping native Responses API support and specific adaptation for Codex. The V4 Flash release keeps the same API endpoint — just swap the model name to deepseek-v4-flash. The V4 Pro official release is coming soon.

An open-weight model with MIT license and 1M token context window, beating the preview of its own bigger sibling on agentic code tasks, at the lowest cost-per-intelligence on the market. The gap between "flash" and "pro" may need a new name.

Sources: DeepSeek API Change Log, Artificial Analysis