DeepSeek V4 Flash: Official Release of the Lightweight Model That Beats V4-Pro on Agent Tasks
On July 31, 2026, DeepSeek took the V4 Flash API out of preview: build 0731, a re-trained 284B-parameter MoE whose agent benchmarks beat V4-Pro-Preview, is now in public beta, and the model weights were published the same day.
What DeepSeek Announced
From Preview to Public Beta: Build 0731
According to the official API changelog, dated July 31, the official release of DeepSeek-V4-Flash is now available in public beta under build DeepSeek-V4-Flash-0731. This is not a minor tweak: the model has been fully re-trained, and its results on agent benchmarks beat those of V4-Pro-Preview, the company's flagship model.
A Surgical Change: Only the V4-Flash API
The update affects only the V4-Flash API. Neither the app, nor the web, nor the V4-Pro API was touched, so teams already consuming V4-Flash get the improvement without forced migrations. The model weights were published the same day, so anyone can self-host it instead of depending on the API.
What DeepSeek V4 Flash Is
A Re-Trained 284B MoE
V4 Flash launched on April 24, 2026, as the lightweight member of the V4 family: reasoning close to V4-Pro, on-par performance on simple agent tasks, smaller size, faster responses, and a much lower price. With build 0731, DeepSeek re-trains that 284B-parameter model with a mixture-of-experts (MoE) architecture and raises the bar exactly where it matters most: agent tasks.
Agent Benchmarks Above V4-Pro-Preview
The most striking fact is that the lightweight model outperforms the flagship on agent benchmarks. In practice, that means for multi-step flows —tool use, code execution, autonomous tasks— V4 Flash can be the most cost-effective choice without giving up results, a direct argument for teams building agents in production.
Pricing and the V4 Family Picture
$0.14/$0.28 per Million Tokens
On DeepSeek's official pricing page, V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens. On OpenRouter, the 0731 snapshot drops to $0.09 and $0.18 respectively, with a 1,048,576-token (1M) context window and up to 65,536 output tokens. Those are among the lowest prices on the market for a model with that context.
V4-Pro and the Permanent 75% Discount
The release fits DeepSeek's aggressive pricing strategy: V4-Pro has kept a permanent 75% discount since May 22, 2026, leaving it at $0.435 per million input tokens and $0.87 per million output tokens, with no expiration date. The V4 family is now tiered by cost: V4-Pro for complex tasks, V4-Flash for volume.
For Developers: Responses API, Codex, and a 1M Context
V4-Flash-0731 adds native support for the Responses API and Codex adaptation, two key pieces for integrating it into agents and coding tools. With a 1M-token context and that price per token, the model is positioned for bulk workloads: long-document processing, data pipelines, and agents that need to keep extended conversations going without inflating the bill. A minimal call example using the Responses API:
from openai import OpenAI
client = OpenAI(api_key="...", base_url="https://api.deepseek.com")
r = client.responses.create(
model="deepseek-v4-flash",
input="Summarize this changelog in three bullet points"
)
print(r.output_text)The snippet is illustrative: each SDK adapts the parameter names, but the model slug and endpoint are the ones recorded in the official documentation.
Conclusion
DeepSeek V4 Flash 0731 turns the lightweight model into the economical option for agents and bulk workloads, with a price per token that is hard to match and a 1M context window. If you work with AI in production, now is a good time to compare your current cost per token against this alternative. Keep following the blog for more technology and artificial intelligence news.
