DeepSeek V4 Pro Official Version Deep Dive: From “Price Slayer” to “Performance Powerhouse”

On August 13, 2026, DeepSeek quietly promoted V4 Pro from “Preview” to “Official” without any launch event. The preview had been online for 111 days since April 24. During those 111 days, Kimi K3 stole the open‑source spotlight with its high‑profile release, V4 Flash went official first, and a pending API price hike loomed over developers. When the final version finally arrived, everyone asked the same question: was it worth the wait?

I. Core Specifications: Architecture Unchanged, Capabilities Transformed The official V4 Pro shares the exact same architecture as the preview version – a MoE (Mixture of Experts) structure with 1.6 trillion total parameters and 49 billion activated parameters, supporting a 1‑million‑token context window and a maximum single output of 384,000 tokens.

The real change lies in post‑training. The company used one striking number to illustrate the magnitude of this improvement: the DeepSWE (Software Engineering Agent benchmark) score skyrocketed from 12.8 in the preview to 62.7 in the official release – a nearly 5‑fold increase. Cybergym rose from 52.7 to 83.3, Terminal Bench 2.1 from 72.1 to 87.9, and DSBench‑Hard more than doubled to 67.2.

What these benchmarks have in common is that they test not just “can it write a piece of code,” but “can it work persistently like a real engineer” – autonomously entering terminals, installing environments, running commands, troubleshooting errors, reading entire codebases, tackling complex engineering problems, and executing multi‑step tool calls over long chains. The preview version was most prone to failures in such long‑horizon tasks, and the official version has achieved a qualitative leap precisely in this area.

Another major update is multimodal capability. The preview was a pure text model; the official version natively supports image reasoning for the first time. The DeepThink engine can now analyze images, interpret screenshots, and process mixed‑format documents within the same reasoning flow.

Efficiency optimisations are also worth noting: under the 1‑million‑token context scenario, V4 Pro’s per‑token inference FLOPs are only 27% of those of the previous‑generation V3.2, and KV cache usage is just 10%. This dramatically lowers the cost of processing ultra‑long documents.

II. V4 Pro vs. V4 Flash: Division of Labour Between Flagship and Lightweight The DeepSeek V4 series adopts a dual‑version strategy. The most intuitive difference lies in parameter scale:

Dimension V4 Pro V4 Flash Total parameters 1.6 trillion 284 billion Activated parameters 49 billion 13 billion Context window 1 million tokens 1 million tokens Max output 384,000 tokens 384,000 tokens API price (output per million tokens) 6 RMB 2 RMB Concurrency limit 500 2500 V4 Pro is positioned as the “flagship top‑tier” – aimed at enterprise‑grade complex research, heavy‑duty development, and ultra‑difficult logical tasks. Its strength shines brightest in complex Agent tasks, where it leads across key Agent benchmarks: Terminal Bench 2.1 (87.9 vs. 82.7), DeepSWE (62.7 vs. 54.4), Cybergym (83.3 vs. 76.7), and DSBench‑Hard (67.2 vs. 59.6).

V4 Flash, by contrast, focuses on efficiency and economy. Despite its much smaller total parameter count, it performs on par with Pro on simple Agent tasks. More importantly, it offers outstanding cost‑performance: the API price is only one‑third of Pro’s, while its concurrency limit is five times higher (2500 vs. 500). Flash also pioneered support for the Responses API and Codex integration.

Think of it this way: V4 Pro is an “armoured professional vehicle”; V4 Flash is an “everyday car” – the former tackles the hardest problems, the latter serves the most people.

For most developers, if you are only doing chat, lightweight coding, or everyday Q&A, Flash is clearly the better value. Only when you need ultra‑long contexts, complex Agent tasks, or higher output quality does it become worth paying triple the price for Pro.

It is worth noting that both V4 Pro and Flash support three‑level thinking intensity adjustment (low/high/max), allowing users to flexibly choose based on task complexity. Both also support both thinking and non‑thinking modes, with thinking mode enabled by default.

III. Advantages over Other Models: A 0.1‑Point Gap, a 60‑Fold Price Difference What truly shook the industry is V4 Pro’s comparison with the world’s top closed‑source models.

On Terminal Bench 2.1 (which measures an AI agent’s ability to complete real‑world terminal tasks), V4 Pro scored 87.9, just 0.1 points behind the global leader Anthropic Fable 5’s 88.0. Three months ago, the preview version was only at 72.1.

Even more striking are two overtakes: on Cybergym (AI security agent testing), V4 Pro surpassed Fable 5 with 83.3 vs. 83.1; on AutomationBench (automation task benchmark), it won 31.8 vs. 29.1. In software engineering capability (DeepSWE), 62.7 exceeded Anthropic’s previous flagship Opus 4.8’s 58.0.

On the price side, the disparity is even more dramatic. Fable 5 charges $50 per million output tokens, GPT‑5.6 Sol Max $30, Grok 4.6 $6, while DeepSeek V4 Pro costs just $0.87 – roughly one‑fifty‑seventh of Fable 5’s price. Cache‑hit input is even lower at $0.0036 per million tokens.

A 0.1‑point performance gap, a 57‑fold price chasm. This equation forces developers to rethink their technology choices. In the past, using DeepSeek meant “saving costs on simple tasks”; now it means “performance approaching the top tier while still saving a huge amount of money.”

In coding ability, V4 Pro achieves 89.2% on HumanEval, surpassing GPT‑5.5’s 88.7% and Claude Opus 4.7’s 87.5%. On Codeforces programming tasks, the model scored 3206, exceeding GPT‑5.4’s 3168. On SWE‑bench Verified, it scored 80.6%, only 0.2 percentage points behind Claude Opus 4.6.

Of course, objectively speaking, gaps remain. In pure knowledge reasoning, ultra‑complex full‑stack software engineering, and other scenarios, V4 Pro still lags slightly behind the world’s absolute top models. DeepSeek founder Liang Wenfeng candidly admitted earlier: “We may be about 12 months behind the US. Given the still order‑of‑magnitude gap in total compute, it is unrealistic to claim comprehensive superiority.” At the same time, V4 Pro’s multimodal capabilities are only just beginning.

But the gap is steadily narrowing. From April to August, V4 Pro took 111 days to complete a transition from “cheap and good enough” to “powerful and still cheap.”

IV. The Price Hike Signal: An Industry Inflection Point Alongside the official release of V4 Pro, DeepSeek announced a peak‑tier pricing mechanism: during peak hours (weekdays 9:00‑12:00 and 14:00‑18:00), V4 Pro’s output price will reach 27 RMB per million tokens – a roughly 350% increase from the current 6 RMB. V4 Flash prices will also rise.

This is not an isolated case. So far this year, Zhipu AI has raised API prices three times, Tencent Cloud significantly increased prices for its Hunyuan series core models in March, and Alibaba Cloud followed suit with price hikes on AI compute cards. The price war that lasted nearly two years seems to be coming to an end.

A Morgan Stanley research note on August 9 pointed out that DeepSeek’s price increase is a positive signal of improving pricing discipline in the industry, underpinned by strong model demand that bolsters pricing power. A recent Goldman Sachs report also significantly raised its 2026 annualised recurring revenue forecasts for Chinese AI models.

Data from the global model API aggregation platform OpenRouter shows that in the week of August 3‑9, models developed by Chinese companies occupied the top four spots in calls, with DeepSeek‑V4‑Flash official ranking first. Chinese models took seven of the top ten positions.

V. Conclusion The official release of DeepSeek V4 Pro marks a transformation for a Chinese AI company from “price slayer” to “performance powerhouse.” With its 1.6‑trillion‑parameter MoE architecture, a standard 1‑million‑token context window, Agent capabilities approaching global No. 1, and prices that are a fraction of its competitors’, it redefines the ceiling of “cost‑performance” in large language models.

The division of labour between V4 Pro and V4 Flash is now clear: Pro tackles the hardest tasks; Flash serves the most people. For developers, the choice is no longer “can I afford it?” but “where should I use it?”

Competition in the large model industry is shifting from “bigger” to “smarter,” and from “price war” to “value war.” DeepSeek V4 Pro stands as one of the most compelling footnotes in this transition.