DeepSeek's V4-Flash Now Beats Its Bigger Sibling on Agent Benchmarks
DeepSeek-V4-Flash exits preview with massively upgraded agent skills, native Responses API support, and Codex integration — same price, same architecture, much better performance.
- DeepSeek-V4-Flash-0731 is now in public beta with massively upgraded agent capabilities that DeepSeek claims surpass V4-Pro-Preview on agentic benchmarks.
- Same architecture, new behavior: Model size (284B total, 13B active) and pricing ($0.14/$0.28 per 1M tokens) are unchanged from the preview.
- Native Responses API + Codex support: V4-Flash now works directly as a backend for OpenAI Codex CLI without custom adapters.
- Strong coding benchmark scores: V4-Flash hits 79% on SWE-bench Verified at $0.28/M output — roughly 1/90th the price of Claude Opus 4.8 (88.6%).
- Limitations remain: Text-only (no multimodal), trails frontier models on PhD-level reasoning, and lags V4-Pro on long-horizon multi-step agent tasks.
- V4-Pro GA is next: Today's update only touches the Flash API; the official V4-Pro release is described as coming "ASAP."
DeepSeek just pushed DeepSeek-V4-Flash-0731 into public beta, and the headline isn't the model itself , it's what changed under the hood. The architecture and parameter count are identical to the preview version, but the agent capabilities have been so dramatically improved that DeepSeek claims the Flash now outperforms the V4-Pro-Preview on agentic benchmarks. That's a meaningful claim: Flash is the smaller, cheaper sibling of the family.
What actually shipped
The update delivers significantly enhanced agent capabilities with benchmark results described as far exceeding V4-Pro-Preview. The DeepSeek API now supports V4-Pro and V4-Flash, available via both the OpenAI ChatCompletions interface and the Anthropic interface. To switch to the new version, you don't need to change anything except confirm you're pointing at the right model name.
The API calling method remains unchanged , simply set the model name to deepseek-v4-flash to use the latest version.
There's no migration required if you were already on the preview.
Two other additions matter for developers building coding agents specifically:
- Native Responses API support: V4-Flash now speaks the OpenAI Responses API format natively, which is the protocol used by tools like OpenAI Codex CLI.
- Codex integration: Codex now sends OpenAI Responses requests to Moon Bridge, and Moon Bridge routes them to DeepSeek V4. This makes V4-Flash a drop-in backend for Codex-based workflows without needing custom adapters.
The model underneath
DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts model with 284B total parameters and 13B activated parameters, supporting a 1M-token context window. MoE (Mixture of Experts) means only a fraction of the network activates per token , here, 13B out of 284B , keeping inference fast and cheap while retaining the capacity of a much larger model.