DeepSeek's retrained V4-Flash beats its own Pro at a third the price
The 0731 rebuild's 13B-active MoE hits Terminal-Bench 82.7 and ships as ungated MIT weights you can self-host. Plus: OpenAI open-sources a Codex scanner.
Copy markdown
Flash overtook Pro on every agent benchmark
The July 31 rebuild jumps Terminal-Bench 2.1 from 61.8 to 82.7, DeepSWE from 7.3 to 54.4, and Cybergym from 38.7 to 76.7 — beating DeepSeek's own flagship V4-Pro on every agentic test. Same 284B total / 13B-active architecture; the gains are pure re-post-training, not a new design.
$0.14 in, $0.28 out — a third of Pro's cost
The API charges $0.14 per million input tokens ($0.0028 on a cache hit) and $0.28 output, against $0.87 output for V4-Pro. Simon Willison calls it 'the best value-per-intelligence model out there,' ranking it above the 428B MiniMax M3 while landing roughly 10x cheaper per task than Opus 5 or GPT-5.6 Sol.
Ungated MIT weights, 13B active — own it
Weights are live on Hugging Face as deepseek-ai/DeepSeek-V4-Flash-0731 under MIT with no gate, and only 13B params fire per token, so one strong GPU box (or a quant) self-hosts it with a 1M-token context on vLLM or SGLang. Build this weekend: a flagship-grade coding agent that never leaves your machine.
The catch: turn the reasoning effort up
Defaults underwhelm — the benchmark-level results only show at 'high' or 'max' effort (up to 384K output tokens), so budget for the extra thinking tokens. And the hosted API runs in China; reach for the open weights or a non-China provider when the code is sensitive.
Elsewhere: OpenAI open-sources a Codex security scanner
OpenAI released the Codex Security CLI under Apache-2.0 — 'npm i @openai/codex-security' scans repos, tracks findings across runs, verifies fixes, and drops security checks into CI. The catch: the scanning model itself stays a gated beta, so it's open plumbing bolted to a closed engine.