Why DeepSeek's Cheapest Model Is Suddenly Its Best One

DeepSeek shipped the official public beta of its V4-Flash API today, moving it out of preview status. That alone would be a minor footnote. Here's what is: the retrained V4-Flash now beats DeepSeek's own flagship, V4-Pro, on agent benchmarks, despite running at a fraction of the size and cost.

We spent the afternoon pulling this apart. Here's what changed, what stayed the same, and what it means if you're already running DeepSeek on GMI Cloud

V4-Pro stays exactly as it was. Same 1.6 trillion parameter model GMI Cloud has served since April 24. It's the same model you already know.

V4-Flash got quietly retrained, same size, same architecture, just a new post-training pass, and now beats Pro on every agent benchmark that matters. One number tells the story: on DeepSWE, Flash jumped from 7.3 to 54.4. That's a 645% improvement, with zero new parameters.

The numbers, side by side

Flash goes beyond just closing the gap with Pro here. It wins outright, on every listed category. That challenges the usual 'bigger model wins' assumption most evaluation pipelines are quietly built on.

Why now

DeepSeek's own changelog frames this as Flash graduating from preview: same architecture, retrained, with the old deepseek-chat and deepseek-reasoner aliases set to retire in three months. V4-Pro's official (non-preview) release is still "coming soon," so this week is deliberately a Flash-first move.

The timing lines up with the rest of the market too. This dropped one day after OpenAI's own price cuts, in the same week Alibaba and MiniMax pushed their own updates. The technical story stands on its own, this timing just explains why your feed is full of 'big AI news' threads right now.

Why this is a capability story

It's tempting to file this under 'got cheaper.' The price stayed exactly the same. V4-Flash's price on GMI Cloud hasn't moved: still $0.14 input / $0.28 output per million tokens. What changed is capability per dollar, and that's a different thing to test for.

If your evaluation setup defaults to Pro for anything "agentic," this is the week to re-run it against Flash:

1. Coding agents and terminal automation – Flash's Terminal Bench and DeepSWE scores now lead Pro outright. Test this first.

2. Tool-calling and multi-step workflows – Toolathlon and NL2Repo both jumped. Check it against your own tool schemas, not just the public numbers.

3. Cost-per-completed-task – At roughly a tenth of Pro's per-token price, a Flash win changes the unit economics of any high-volume agent pipeline.

A benchmark table tells you where to look. Your own prompts will tell you the rest.

Try it on GMI Cloud right now

Both models are live, same OpenAI-compatible endpoint, so switching is a one-line change:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.gmi-serving.com/v1",
    api_key="YOUR_GMI_API_KEY",
)

response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4-Flash-0731",
    messages=[
        {
            "role": "user",
            "content": "Find the failing test in this repo and explain the root cause."
        }
    ],
)
print(response.choices[0].message.content)

Want to see it work inside a real coding loop instead of a single completion? OpenCode is a strong harness for V4-Flash, and GMI Cloud is a native provider inside it. Connect your API key from the terminal, swap models mid-session, no config files.

One thing still catching up

DeepSeek is still preparing a -0731-suffixed weights repo for Hugging Face as of this afternoon. The public MIT-licensed repos still reflect the April preview generation. GMI Cloud's live endpoint runs DeepSeek's official API release, which is ahead of the open-weight publish. If you specifically need raw weights for local or dedicated deployment rather than API access, that piece is still on its way.

Do this today

• Re-run your agent and coding benchmarks against V4-Flash-0731 before defaulting to Pro for cost reasons alone

• Track cost-per-completed-task, not just token price, they tell different stories this week

• Point OpenCode (or your own harness) at GMI Cloud, the model swap is one line

• If you need open weights instead of API access, give it a little more time, the -0731 Hugging Face repo is on its way

Get started

DeepSeek-V4-Pro and DeepSeek-V4-Flash-0731 are both live now through GMI Cloud's OpenAI-compatible API. Start in the Playground to test