DeepSeek V4 Flash Official Release (0731): Agent Benchmarks Jump Past V4 Pro — Live on TokenPAPA

#ai #api #deepseek #llm

DeepSeek V4 Flash Official Release (0731): Agent Benchmarks Jump Past V4 Pro

On July 31, 2026, DeepSeek officially released DeepSeek-V4-Flash to public beta. The API calling convention is unchanged — set model to deepseek-v4-flash and you get the latest version.

The headline: agent capabilities are dramatically stronger, with benchmark scores that beat DeepSeek-V4-Pro-Preview across the board.

And the good news: TokenPAPA already supports the update. No code changes needed.

What's New in V4 Flash 0731

Aspect	Detail
Release	Official version, public beta
Model Name	`deepseek-v4-flash` (unchanged)
Architecture	Same structure/size as V4 Flash Preview
Change	Post-training refresh only
API	Chat Completions + native Responses API
Codex	Specifically adapted ✅

DeepSeek-V4-Flash-0731 keeps the same model structure and size as DeepSeek-V4-Flash-Preview — only the post-training was redone. The API interface was upgraded for V4 Flash only; V4 Pro API and APP/WEB models are unchanged, and the V4 Pro official release is coming soon.

Agent Benchmark Scores: Flash Beats Pro-Preview

The official release focuses on agentic coding and tool-use. Here are the published scores:

Benchmark	V4 Flash (0731)	vs V4 Pro Preview
Terminal Bench 2.1	82.7	Far ahead ✅
NL2Repo	54.2	Far ahead ✅
Cybergym	76.7	Far ahead ✅
DeepSWE	54.4	Far ahead ✅
Toolathlon (verified)	70.3	Far ahead ✅
Agent Last Exam	25.2	Far ahead ✅
Automation Bench (Public)	25.1	Far ahead ✅
DSBench-FullStack (internal)	68.7	Far ahead ✅
DSBench-Hard (internal)	59.6	Far ahead ✅

Note: For public Code Agent benchmarks, DeepSeek used the DeepSeek Harness minimal mode (upcoming release) with max settings, top_p=0.95, temperature=1.0. DSBench-FullStack is an internal full-stack dev test set; DSBench-Hard is an internal hard Coding Agent test set.

What this means: for agentic workloads — terminal automation, repo generation, cybersecurity tasks, full-stack development — the Flash tier now delivers flagship-class agent ability at Flash pricing.

What This Means for Developers

Cheapest agent model just got stronger — V4 Flash was already the budget workhorse ($0.14/1M input, $0.28/1M output, $0.0028/1M cache-hit input). Now its agent skills rival the Pro tier.
Responses API + Codex ready — native support means you can drop it into modern OpenAI tooling and Codex-style agent loops without translation layers.
No migration cost — same model name, same endpoint. Existing code keeps working.
Pro official version incoming — expect the V4 Pro refresh soon; Flash users get the upgrade today.

Try It on TokenPAPA Right Now

TokenPAPA already supports the updated deepseek-v4-flash — same API, same model name:

from openai import OpenAI

client = OpenAI(
    api_key="your-tokenpapa-key",
    base_url="https://tokenpapa.ai/v1"
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",  # Latest official version
    messages=[{"role": "user", "content": "Write a Python script that scans a repo and summarizes TODO comments."}]
)
print(response.choices[0].message.content)

Using the native Responses API

response = client.responses.create(
    model="deepseek-v4-flash",
    input="List all open issues from this GitHub repo and suggest fixes.",
)
print(response.output_text)

Using it with Codex-style tool loops

# TokenPAPA is OpenAI-compatible, so Codex and similar agents work directly:
# 1. Set base_url to https://tokenpapa.ai/v1
# 2. Set model to deepseek-v4-flash
# 3. The agent handles function/tool calls natively

Why Access V4 Flash Through TokenPAPA?

One key, 30+ models — switch between DeepSeek, GPT, Claude, Gemini, Kimi, Qwen, Mimo with a single line change
No VPN / no Chinese phone required for overseas developers
OpenAI, Claude, and Gemini native formats all supported
Instant upgrade — the 0731 release is already live

Getting Started

Sign up at tokenpapa.ai — free credits on registration
Get your API key from the console
Call deepseek-v4-flash — you're on the latest official version today

from openai import OpenAI
client = OpenAI(base_url="https://tokenpapa.ai/v1", api_key="your-key")

resp = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Explain your agent capabilities in one sentence."}]
)
print(resp.choices[0].message.content)