DeepSeek V4 Flash 0731 beats its own Pro flagship at $0.14/$0.28
A retrained Flash jumps 10 points to 50 on the Intelligence Index, adds native Responses API + Codex support, and undercuts GPT-5.6 Luna ~60% per task.
Copy markdown
The jump: 50 on the Index, +10 in three months
DeepSeek's retrained V4 Flash scores 50 on Artificial Analysis' Intelligence Index — up 10 points from April's release and, notably, 6 points ahead of its own V4 Pro flagship. Same 1M-token context, 284B total / 13B active params.
The price: $0.14 in, $0.28 out, per million
First-party API pricing is $0.14/$0.28 per 1M tokens with a 98% cache-hit discount ($0.0028/1M). That's roughly 60% cheaper per task than GPT-5.6 Luna — even after OpenAI's fresh 80% Luna cut — while landing just one point behind it on the Index.
Build this weekend: native Responses API + Codex
V4 Flash 0731 ships native OpenAI Responses-API format and documented Codex integration, making it a near drop-in for Responses-native agent harnesses at ~93 tok/s. Point your coding agent at it and eval against Luna on your own tasks.
The gains are in agent work, not trivia
GDPval-AA v2 Elo climbed to 1559 (from 1189) and Terminal-Bench 2.1 hit 79% — the improvement is concentrated in real-world, multi-step tool use, exactly where cheap models usually collapse. DeepSeek says it beats its Pro flagship on nine agent benchmarks.
The catch: text-only, weights still weeks out
No image or non-text input, so vision-dependent coding is out for now, and full open weights are only “coming weeks” — today it's API-only. As always with fast-moving Chinese providers, run your own eval set and re-confirm pricing before shifting production traffic.
Elsewhere: OpenAI's content-provenance checks reach the API
OpenAI's July 31 SDK bump (Python 2.52, Node 7.3, Go 3.49) exposes content-provenance verification through the API, so you can confirm whether content carries OpenAI's C2PA/SynthID signals programmatically instead of via the web-only tool.