DeepSeek V4 Flash vs V4 Pro: The Efficiency Tier vs the Flagship, Compared

DeepSeek's V4 line is a two-rung ladder: V4 Flash, the cheap, fast efficiency model — now in its official -0731 release with much stronger agents — and V4 Pro, the large flagship for the hardest reasoning and coding, which moved to its official GA build (0813) on August 12, 2026. The interesting part is how little separates them on practical work, and how much separates them on price. This guide compares the two on capability, cost, speed, and fit, with every figure labeled by source.

Accuracy note: V4 Flash's agentic/coding scores are DeepSeek-reported (official change log, 2026-07-31, DeepSeek harness); V4 Pro's GA (0813) scores are likewise DeepSeek-reported on the same harness, and no independent evaluation of the 0813 build has been published yet. Pricing is current as of August 12, 2026; the GA price cut is passed through on OrcaRouter at 0% markup. Vendor-harness numbers run optimistic — verify on your own tasks.

TL;DR. V4 Flash is a 284B / 13B-active MoE at $0.08 / $0.18 per million tokens; V4 Pro is the far larger flagship, now in its official GA build (0813) at $0.435 / $0.87 — about five times Flash's price, down from roughly fourteen times before the August price cut. On practical coding, Flash lands within a few points of Pro (e.g., SWE-bench Verified ~79% vs ~80.6%, aggregator-reported), and the -0731 post-training upgrade makes Flash a strong agent in its own right. Use Pro for the hardest reasoning and the most demanding multi-file coding; use Flash for the high-volume majority — and route by difficulty to get most of Pro's quality at roughly a fifth of the cost.

Key takeaways

• Size and price: Flash is 284B / 13B-active at $0.08 / $0.18; Pro is the much larger flagship at $0.435 / $0.87 — Flash is roughly a fifth of the price, and Pro's GA price cut narrowed the old ~14x gap to about 5x.

• Coding gap is small: aggregator SWE-bench Verified ~79% (Flash) vs ~80.6% (Pro, preview-era; the GA build has no independent scores yet); on DeepSeek's harness, GA Pro posts Terminal-Bench 2.1 87.9 vs Flash's 82.7.

• Both share the 1M-token context and 384K output; both are text-only, with reasoning, tools, and JSON.

• Flash is faster and cheaper to serve (p50 TTFT ~394 ms, ~184 tok/s); Pro trades speed for peak capability.

• Best practice: route by difficulty — Flash for the volume, Pro for the hard minority.

What each model is

V4 Flash is the efficiency tier: a mixture-of-experts model with 284B total but only ~13B active parameters, a 1M-token context, up to 384K output, tuned for high-volume, low-latency, agentic work. Its official -0731 release (July 31, 2026) is a post-training upgrade that sharpened agents, coding, and tool use, and added native Responses-API and Codex support. V4 Pro is DeepSeek's flagship — a much larger model built for the hardest reasoning and coding, where maximum capability matters more than cost. It ran as a preview from April, then moved to its official GA build on August 12, 2026, with a much lower price. Both sit in the same V4 family and share the 1M context, text-only input, and reasoning/tools/JSON support.

Capability: how close is Flash to Pro?

Closer than the price gap suggests, at least on practical coding. Aggregator evaluations put V4 Flash (Max) around 79.0% on SWE-bench Verified — resolving real GitHub issues — versus about 80.6% for V4 Pro as tested in its preview era. The GA build (0813) is too new for independent scores — none have been published yet — so the best available Pro comparison is DeepSeek's own harness, where the GA build posts Terminal-Bench 2.1 87.9 and DeepSWE 62.7, against Flash's -0731 figures of 82.7 and 54.4 (all DeepSeek-reported). Where Pro pulls ahead is the hardest end: the most complex multi-step reasoning, the trickiest multi-file coding, and tasks where a larger active-parameter budget genuinely helps. If most of your work is "hard but not frontier," Flash covers it; reserve Pro for the genuinely frontier minority.

Price and speed: Flash's decisive edge

This is where the two diverge most — and where the math just changed. Flash is $0.08 / $0.18 per million input/output tokens; V4 Pro's official GA build (0813) is $0.435 / $0.87 — about five times Flash's price. That is still a real premium, but a fraction of the ~14x gap before the price cut: the older deepseek-v4-pro listing stood at $1.168 / $2.336, so the GA build is roughly 63% cheaper. On a 10-million-token month at 70% input, Flash runs about $1.10 and Pro about $5.66 — the ratio holds as you scale to billions of tokens. Flash is also built for throughput (p50 TTFT ~394 ms, ~184 tok/s), so it's the better fit for latency-sensitive, high-volume agents. Pro's premium buys peak capability, not speed or savings — but with the gap at ~5x instead of ~14x, the price of that top-end headroom has come way down.

The right pattern: route by difficulty

You don't have to choose one. The cheapest high-quality setup sends the easy-to-moderate majority of requests to Flash and escalates only the hardest to Pro. Because both are the same family with the same context and interfaces, that routing is seamless — and the -0731 upgrade means Flash now handles more of the middle tier before you need to escalate. Done well, route-by-difficulty delivers most of Pro's quality at close to Flash's cost, and the GA price cut makes the occasional Pro escalation notably cheaper than it was during the preview.

Which should you choose?

Choose V4 Flash if…

You run high-volume agentic, coding, or long-document workloads where cost and speed matter, and "near-flagship" quality is enough — which, after the -0731 upgrade, covers a lot of ground.

Choose V4 Pro if…

You need maximum capability on the hardest reasoning and the most demanding multi-file coding, and you're willing to pay about 5x Flash's price for that top-end headroom — a much easier ask than the ~14x premium the preview-era pricing carried.

Use both through one endpoint

Both are available on OrcaRouter at 0% markup through one OpenAI-compatible endpoint — Flash as deepseek/deepseek-v4-flash-0731 and Pro as the official deepseek/deepseek-v4-pro-0813 GA build — so you can implement route-by-difficulty as a config choice, benchmark both on your own tasks, and switch traffic without re-integrating. That's the simplest way to capture Flash's savings while keeping Pro on tap for the hard minority.

FAQ

Is V4 Flash as good as V4 Pro?

On practical coding, nearly — aggregator SWE-bench Verified ~79% vs ~80.6% (preview-era for Pro; the GA build has no independent scores yet). On the hardest reasoning and most complex coding, Pro leads. For most workloads Flash is enough after the -0731 upgrade.

How much cheaper is V4 Flash?

Flash is about a fifth of Pro's price: $0.08 / $0.18 per million tokens versus GA Pro's $0.435 / $0.87. Before Pro's August price cut, the gap was ~14x; it is now about 5x.

Do they share the same context window?

Yes — both offer a 1M-token context (1,048,576) and up to 384K output.

Which is faster?

Flash, by design — p50 time-to-first-token around 394 ms and ~184 tokens/second, tuned for high-volume, low-latency use.

Can I use both together?

Yes — route by difficulty through OrcaRouter's single endpoint: Flash for the volume, Pro for the hardest tasks.

Bottom line

V4 Flash vs V4 Pro is efficiency versus peak capability. Flash gives you most of Pro's practical coding ability, plus a genuinely strong agent after the -0731 upgrade, at about a fifth of the price and higher speed; Pro's official GA build is the flagship for the hardest work, and its price cut — down to $0.435 / $0.87, about five times Flash instead of the old fourteen — makes that top end more affordable than it has ever been. The smart move isn't either/or — it's routing by difficulty through one endpoint like OrcaRouter, so the volume runs cheap on Flash and only the hard minority pays for Pro.