Is DeepSeek V4 Flash or V4 Pro better for coding agents?

Flash, since the 0731 build. It scores 82.7 on Terminal Bench 2.1 and beats the V4 Pro Preview on several code and agent benchmarks while costing about a third as much and supporting the Responses API format that Codex style tools expect. Pro has not received an equivalent agent focused update yet.

How much cheaper is V4 Flash than V4 Pro?

On the official API, Flash is $0.14 in and $0.28 out per million tokens versus $0.435 in and $0.87 out for Pro, so output is about 3.1x cheaper on Flash. Both get extreme cache hit discounts on input.

Can either model run locally?

Flash yes, Pro no. Flash is about 155 GB at 4 bit and runs on a 192 GB Mac Studio or a big RAM workstation. Pro is a 1.6T parameter model that approaches 900 GB at 4 bit, which is multi GPU server territory.

When is V4 Pro worth the premium?

When raw reasoning depth is the bottleneck: hard math, difficult proofs, long multi document analysis where the 49B active parameters and larger expert pool earn their keep. For everyday chat, coding, and agent loops, Flash 0731 is the better value and often the better model.

DeepSeek V4 Flash vs V4 Pro: Which One and When

DeepSeek's V4 family splits into two very different animals: Flash (284B MoE, 13B active), which got its big 0731 release today, and Pro (1.6T MoE, 49B active), the reasoning heavyweight still in preview on the API. Same 1M context, same MIT open weights policy, very different jobs. Here is the practical breakdown.

Spec Sheet

	V4 Flash (0731)	V4 Pro (Preview)
Total parameters	284B	1.6T
Active per token	13B	49B
Context	1M in / 384K out	1M in / 384K out
API price / M tokens	$0.14 in, $0.28 out	$0.435 in, $0.87 out
Responses API / Codex	Yes, native (0731)	No
Open weights	MIT, on HF	MIT, on HF
4 bit GGUF size	~155 GB	~900 GB (server only)
Runs at home	Yes, 192 GB memory class	No

The Plot Twist: Flash Beats Pro at Agent Work

Until this morning the story was simple: Pro is the smart one, Flash is the cheap one. The 0731 re post training broke that. On Terminal Bench 2.1, Flash 0731 posts 82.7; on DeepSWE it jumped from 7.3 to 54.4; and across several code and agent benchmarks it now outscores the V4 Pro Preview, a model activating nearly four times the parameters per token. DeepSeek trained the small model into a specialist and, for agent loop workloads, the specialist wins.

Pro keeps the crown where breadth and depth dominate: hard math, long proofs, dense multi document reasoning. The 1.6T expert pool is not decoration; on deep reasoning evals Pro Preview remains ahead, and a future Pro release trained with the same agent recipe would presumably retake the lead. But that model is not here today, and Flash 0731 is.

Cost in Practice

The sticker gap says 3.1x on output, but agent reality widens it: agent loops burn output tokens (Flash's cheap side) and resend context (where cache hits cost fractions of a cent on both). A day of heavy agent work that costs single digit dollars on Flash lands in the tens on Pro. On flat plans the gap disappears from your bill: LU Labs carries both models in the Hosted Pro plan at 49€/mo against one quota, which makes switching per task free, in the way that produces the honest comparison: use both, notice which one you stop switching away from.

The Local Angle

This one is not close. Flash fits in 155 GB at 4 bit and runs on a 192 GB Mac Studio or big RAM workstation at interactive speeds (full math in the hardware post, setup in the local guide). Pro at 4 bit approaches 900 GB and wants a multi GPU node (details in the Pro guide). If owning the model matters to you, the family decision is already made, and the community agrees: the abliterated Flash build is the most downloaded uncensored model of 2026.

Verdict by Workload

Workload	Pick
Coding agents, terminal work, repo tasks	Flash 0731, it benchmarks ahead and speaks Responses API
Everyday chat and writing	Flash, the quality gap does not justify 3x
Hard math, proofs, research grade reasoning	Pro
Very long document analysis at full 1M context	Pro (hosted; nobody runs 1M locally)
Running on your own hardware	Flash, by forfeit