DeepSeek-V4-Flash-0731: The Cheap Model Just Beat the Expensive One

DeepSeek shipped a model today that outscores its own flagship, and the only thing they changed was the post-training.

Same 284B total parameters. Same 13B active. Same 1M context window. Same $0.14 per million input tokens. Same model string in your API call. The architecture is identical to what landed in April as DeepSeek-V4-Flash-Preview — DeepSeek says so directly in the changelog. What moved is everything downstream of pretraining, and it moved far enough that the small model now beats DeepSeek-V4-Pro-Preview, a 1.6T-parameter model with nearly 4x the active parameters and 3x the price tag.

That's the headline. Below is what actually shipped, what the numbers say, and where you should be skeptical.

---

What Shipped on July 31, 2026

DeepSeek's API changelog, dated July 31, moved the official DeepSeek-V4-Flash API into public beta under the build name DeepSeek-V4-Flash-0731. The upgrade path is one line:

# Before
client.chat.completions.create(model="deepseek-v4-flash-preview", ...)

# After
client.chat.completions.create(model="deepseek-v4-flash", ...)

No base URL change, no credential change, no restructuring. If you were already pinned to deepseek-v4-flash, the model underneath your calls changed on you today — worth knowing before you go debugging why your agent's token spend shifted.

Four things came with it:

Responses API support, natively. V4-Flash now speaks OpenAI's Responses API format and has been specifically adapted for Codex. DeepSeek's integration docs claim a single config exposes the model across Codex CLI, the ChatGPT desktop app, and the Codex VS Code extension. Flash is currently the only DeepSeek model with documented Codex support. That matters more than it sounds — Responses gives you more flexible tool calls and finer output control than plain Chat Completions, which is exactly what long-horizon agent loops need.

Weights, under MIT. The checkpoint is live on Hugging Face right now, MIT-licensed, with vLLM and SGLang recipes in the card. The repo lists 304B params rather than 284B because the release ships with a DSpark speculative decoding module bolted on. You enable it with one vLLM flag.

Three reasoning effort levels. low, high, and max, controlling how much the model deliberates before answering. DeepSeek recommends a 384K max output length for high and max. Note there's no Jinja chat template in this release — DeepSeek ships an encoding folder with Python scripts instead.

Scope limits. Only the V4-Flash API changed. The V4-Pro API and the App/Web models are untouched. If your production system is pinned to Pro, nothing happened to you today. DeepSeek says the official V4-Pro release is coming, with reporting pointing to August.

---

The Benchmark Numbers

Here's DeepSeek's own table from the model card. GLM-5.2 and Opus-4.8 columns are DeepSeek's comparison set, not mine.

Benchmark	V4-Flash-0731	V4-Flash (Preview)	V4-Pro (Preview)	GLM-5.2	Opus-4.8
Terminal Bench 2.1	82.7	61.8	72.1	81.0	85.0
NL2Repo	54.2	39.4	38.5	48.9	69.7
Cybergym	76.7	38.7	52.7	—	83.1
DeepSWE	54.4	7.3	12.8	46.2	58.0
Toolathlon-Verified	70.3	49.7	55.9	59.9	76.2
Agents' Last Exam	25.2	15.8	16.5	23.8	25.7
AutomationBench Public	25.1	10.8	12.8	12.9	27.2
DSBench-FullStack †	68.7	37.0	41.8	61.8	71.6
DSBench-Hard †	59.6	25.8	31.1	54.5	71.7

† Internal DeepSeek test sets.

Nine benchmarks, nine wins over the bigger sibling. DeepSWE went from 7.3 to 54.4 — that's not an improvement, that's a model that previously couldn't do the task learning to do the task.

Now the asterisks, because they're load-bearing.

These are vendor-reported numbers. DeepSeek ran the public Code Agent tasks using the minimal mode of DeepSeek Harness — a framework that hasn't been released yet — at max reasoning effort, temperature=1.0, top_p=0.95. Two of the nine rows are internal test sets nobody else can run. Cross-provider comparisons under different harnesses tell you a direction, not a ranking. Opus-4.8 and GLM-5.2 weren't evaluated by their own labs under DeepSeek's harness.

Treat the Flash-0731 vs Flash-Preview vs Pro-Preview columns as the trustworthy part — those are apples-to-apples inside one lab's setup. Treat the rightmost two columns as marketing until someone independent reruns them.

---

What the Independent Numbers Say

Artificial Analysis published their evaluation the same day, and it's the useful cross-check.

Intelligence Index: 50, up 10 points from the previous V4 Flash's 40. That puts it 6 points ahead of DeepSeek V4 Pro on the same index — independent confirmation of the "small model beats big model" claim, using a harness DeepSeek doesn't control.

Where that lands it competitively:

• One point behind GPT-5.6 Luna (max, 51)
• Within one point of GLM-5.2 (max, 51)
• Level with Gemini 3.6 Flash (50)
• Seven points behind the open-weights frontier, Kimi K3 (max, 57)

On GDPval-AA v2, an agentic real-world-work eval, Elo jumped from 1189 to 1559. Once weights circulate widely that's the second-highest open-weights score, behind Kimi K3 (1687) and ahead of GLM-5.2 (1510).

Two details from AA that deserve attention:

Terminal-Bench 2.1 came in at 79%, not 82.7%. AA measured a 17-point rise to 79. DeepSeek reported 82.7 on its own harness. Neither is wrong — it's a harness gap, and it's a clean illustration of why you shouldn't port vendor benchmark numbers into your own capacity planning.

Hallucination, not accuracy, drove the knowledge gains. AA-Omniscience improved 7 points to -16, and AA is explicit that overall accuracy percentage was unchanged. The model didn't learn more facts. It learned to stop making them up. For agent work that's arguably the more valuable fix, but don't read it as a knowledge upgrade.

---

Pricing, and the Thing That Actually Makes It Cheap

Straight from DeepSeek's pricing page as of today:

	deepseek-v4-flash	deepseek-v4-pro
Input, cache hit	$0.0028 / 1M	$0.003625 / 1M
Input, cache miss	$0.14 / 1M	$0.435 / 1M
Output	$0.28 / 1M	$0.87 / 1M
Context	1M	1M
Max output	384K	384K
Concurrency limit	2500	500

Look at the cache hit row. $0.0028 against a $0.14 cache-miss rate is a 98% discount, where most of the industry sits at 90%. For agent loops — which replay enormous shared prefixes over and over — that single line item is doing more work than the headline rate. AA's blended figure at a 7:2:1 cache-hit/input/output ratio comes to roughly $0.08 per million tokens.

OpenAI cut GPT-5.6 Luna's price by 80% the day before this launch. Even after that cut, AA measures Flash-0731's cost per task on DeepSeek's first-party API at about 60% below GPT-5.6 Luna (max) — a model one Intelligence Index point ahead of it. The cache discount is the main reason.

One catch you need to budget for: this model is verbose. AA burned 210M output tokens running the Intelligence Index on it, against a median of 62M for comparable reasoning models. Total eval cost was $237. Reasoning tokens bill at the output rate, so a cheap per-token price on a chatty model doesn't automatically produce a cheap per-task bill. Measure total tokens generated on your workload, not the rate card.

If you're on OpenRouter, its listed rates differ from DeepSeek's first-party card — check which one you're actually paying.

---

List Your AI Tool on SaaSCity

Shipping inference infrastructure, an agent harness, or a dev tool built on open-weight models like V4-Flash? Get it in front of the people making exactly these model-selection calls.

SaaSCity is a free startup directory where your product doesn't sit in a list — it gets a building on an interactive 3D city map, plus a permanent indexed page.

• 🆓 Free to list: submit in under 2 minutes, no credit card
• 📈 Dofollow backlink that actually moves your domain rating over time
• 🗺️ 3D map visibility instead of row 400 of a spreadsheet
• 🎯 The right audience: founders, tooling engineers, and early adopters comparing models this week

Submit your product for free →

---

Running It Locally

Unsloth had GGUFs up within hours, and the quantization story here is unusually good.

DeepSeek-V4-Flash is quantization-aware trained. The official checkpoint stores its routed experts — 96% of the model — natively in MXFP4, with everything else in FP8 or BF16. GGUF's MXFP4 is that same format, so Unsloth repacks the experts bit-for-bit and dequantizes FP8 into BF16 with no rounding. Their claim: all 1,328 tensors bit-identical to DeepSeek's weights, KL-divergence near zero, 100% top-token agreement.

Practical numbers:

• UD-Q8_K_XL: 162 GB, lossless
• UD-Q4_K_XL: only ~7 GB smaller than Q8, same bit-exact experts
• 3-bit: 103 GB, runs on a 110 GB RAM machine

That Q8/Q4 gap being 7GB is the interesting part — when the experts are already 4-bit natively, "quantizing" barely saves you anything. Just run Q8.

Unsloth's note is worth repeating: GGUFs converted without these paths deviate from the official weights. Source matters here.

For server deployment, vLLM with --speculative-config '{"method":"dspark",...}' turns on the speculative decoding module. SGLang works too. Sampling recommendation is temperature=1.0, with top_p=0.95 for agentic work and 1.0 otherwise.

If 162 GB is out of reach, the smaller local models worth running haven't gone anywhere — this is a server-class checkpoint, not a laptop one.

---

Should You Switch?

Switch now if you're running agentic coding, terminal agents, SWE pipelines, or high-volume tool-calling loops where cost per task decides whether the product works. The cache economics on repeated-prefix workloads are hard to argue with, and the concurrency limit is 5x Pro's.

Wait if you need broad single-shot knowledge or reasoning strength. AA's data says accuracy didn't move — hallucination rate did. And official V4-Pro is close.

Don't switch if any of these apply: you need an SLA (this is public beta, DeepSeek hasn't published enterprise terms), you can't send data to China-based servers, or you need multimodal input — this is text-only.

Before you route production traffic: run your own eval set. DeepSeek has previously introduced peak-hour surge pricing, so confirm current rates rather than trusting a blog post — including this one. Check beta-tier rate limits, which are often tighter than GA. And smoke-test the Responses API path specifically if you're integrating through Codex, since that support is one day old.

If you're comparing across providers before committing, our breakdown of AI agent coding token plans covers what the per-seat and per-token options actually cost at volume.

---

The Part That Should Actually Change Your Thinking

Strip away the benchmark table and the price war framing, and one fact is left standing: DeepSeek took a model, changed nothing about its size or architecture, redid the post-training, and got a system that beats a sibling model with 4x the active parameters.

That's a statement about where the returns are right now. Post-training and harness design are producing gains that parameter scaling used to produce, at a fraction of the compute. If a 13B-active model can be re-trained into beating a 49B-active one, the question every lab is now asking isn't "how much bigger" — it's "how much better can we make what we already have."

The weights are MIT-licensed and on Hugging Face. Which means everyone gets to find out at once.

---

Advertise Your Startup on SaaSCity

Cheap frontier-class inference means the wrapper you were priced out of building last quarter is viable this quarter. When you ship it, distribution is the next problem.

SaaSCity puts your product on the map — literally. Free listings, a permanent indexed page, and a launch queue that gets you in front of builders instead of into an archive.

Submit your startup → · Browse the directory →

---

All figures verified against DeepSeek's official API changelog and pricing page, the DeepSeek-V4-Flash-0731 model card on Hugging Face, Artificial Analysis's July 31, 2026 evaluation, and Unsloth's deployment documentation, as of July 31, 2026. Vendor-reported benchmarks are labeled as such. Pricing and beta terms are subject to change — confirm at the source before committing spend.

SaaSCity.io covers AI model releases, benchmarks, and the tooling decisions behind them. Explore the directory or list your own product.