DeepSeek V4 Pro vs. Grok 4.6: First Real-World Tests Put Both Models in the Top Tier

Introduction

Two major AI model releases landed at almost the same time, putting DeepSeek V4 Pro and Grok 4.6 directly into the spotlight.

DeepSeek V4 Pro arrived with a sharp jump in agent and software-engineering performance, while Grok 4.6 pushed hard on coding, knowledge work, and overall capability. The source article describes both releases as a direct challenge to the current frontier models from OpenAI and Anthropic.

What makes the comparison especially interesting is that the models were not only measured on benchmark tables. Early real-world tests also asked them to build websites, create games, reproduce visual designs, and generate 3D experiences from the same prompts.

The result is less like a simple winner-takes-all contest and more like a close match in which each model has areas where it stands out.

DeepSeek V4 Pro Arrives as Grok 4.6 Enters the Ring

The source article highlights several benchmark results for DeepSeek V4 Pro.

In CyberGym, a cybersecurity-agent evaluation, DeepSeek V4 Pro reportedly scored 83.3, narrowly exceeding Fable 5 at 83.1 and Opus 4.8 at 78.3.

On AutomationBench, it reached 31.8, ahead of Fable 5 at 29.1 and Opus 4.8 at 27.2.

The closest result came from Terminal-Bench 2.1. DeepSeek V4 Pro scored 87.9, compared with 85.0 for Opus 4.8 and 88.0 for Fable 5.

The source article also reports a score of 60.0 on a tool-enabled Humanity's Last Exam setting, compared with 57.9 for Opus 4.8 and 63.0 for Fable 5.

DeepSWE was another major improvement in the report. DeepSeek V4 Pro reached 62.7, up sharply from the preview version's 12.8 and above Opus 4.8's reported 58.0, while remaining below Fable 5's 70.0.

Grok 4.6 was tested against GPT-5.6 and Fable 5 in several areas.

The source article reports a score of 61 on its comprehensive intelligence index, one point behind Fable 5's 62.

On CursorBench, Grok 4.6 reportedly scored 69.9%, ahead of GPT-5.6 at 67.2% and close to Fable 5 at 70.5%.

On FrontierCode, it reached 61.3%, again slightly ahead of GPT-5.6 at 60.6% and behind Fable 5 at 63.6%.

The knowledge-work results were even stronger in the source article. Grok 4.6 reportedly took first place in all three of the highlighted evaluations: 1,753 Elo on GDPval-AA v2, 1,577 Elo on AA-Briefcase, and 15.8% on the professional legal benchmark Harvey LAB.

The source article's main conclusion from these figures is straightforward: both models have moved into the same competitive conversation as the leading frontier systems.

The Price Difference Is Part of the Story

Performance is only half of this comparison.

The source article also emphasizes inference cost. It reports the following output-token prices per one million tokens at the time of publication:

Model	Reported output price per 1M tokens
DeepSeek V4 Pro	$0.87
Grok 4.6	$6
GPT-5.6 Sol	$30
Claude Opus 5	$25
Fable 5	$50

0.435 per million input tokens for cache misses and $0.003625 for cache hits in its current USD pricing view.

DeepSeek's documentation also notes that its V4 models support a 1M-token context window, tool calls, and both OpenAI-format and Anthropic-format API access.

The exact prices of the other models can change over time, so the source article's cross-model price comparison should be treated as a snapshot from its publication date rather than a permanent price table.

First Real-World Test: DeepSeek V4 Pro Builds a 3D Earth

The first practical test described in the article put DeepSeek V4 Pro under a fairly demanding prompt.

With a single prompt, the model generated a complete interactive 3D Earth experience in a browser.

The result reportedly supported basic interactions such as dragging, rotating, and zooming. It also included global data flows, dynamic routes, and geographic markers placed across the globe.

The visual details went further. Atmospheric scattering, clouds, day-and-night lighting, and the surrounding UI were all included in the generated experience.

The important part of this test is not simply that a model generated a webpage. It shows how far an AI coding model can go when asked to coordinate visual design, 3D rendering, interaction, and application structure in one task.

DeepSeek V4 Pro vs. Grok 4.6 in Three Application Tasks

The source article then compares the two models using the same or closely matched prompts.

AI creator Xiangyang Qiaomu first ran three small tasks with DeepSeek V4 Pro and then passed two of the same prompts to Grok 4.6.

Building and Deploying a Website with Three Skills

The first task asked the model to use three Skills to develop and deploy a website.

DeepSeek V4 Pro reportedly completed the full workflow successfully. The source article describes the page design and overall completeness as stable.

This type of test is particularly relevant to agentic coding because the model has to coordinate multiple tools or capabilities rather than simply generate a block of code.

Reproducing 60 Design Styles in Bento Cards

The second task asked the model to reproduce 60 different design styles and present them as a collection of Bento cards.

DeepSeek V4 Pro reportedly differentiated the typography, color combinations, and layouts across the designs.

When the same prompt was given to Grok 4.6, the latter gained an advantage. The source article says some of Grok's pages had more mature layout, color, and visual hierarchy, producing a more polished final presentation.

Building a 3D Brick-Breaker Game

The third task was to create a 3D brick-breaker game from scratch.

DeepSeek V4 Pro produced a playable game with a 3D environment, background music, dynamic sound effects, and interactive feedback.

Grok 4.6 performed at a similar level in this round. The source article describes the two results as nearly tied in visual quality, scene completeness, and playability.

The Flappy Bird Test Shows a Different Trade-Off

A more detailed test came from developer Jun Song, who gave DeepSeek V4 Pro and Grok 4.6 exactly the same prompt to build a Flappy Bird-style game from scratch.

The two models took very different approaches.

DeepSeek V4 Pro used more than 20,000 tokens, but the reported API cost was only $0.019.

Grok 4.6 used about 5,000 tokens, while the reported cost was $0.03.

So Grok was more token-efficient in this particular run, but DeepSeek produced the stronger final result according to the source article.

The DeepSeek version reportedly included a layered mountain background, clouds, gradient and volumetric-looking pipes, and a floating “+1” animation when the character passed a pipe.

The source article's takeaway is useful: lower token consumption does not automatically mean a better application result, and higher token consumption does not necessarily mean higher cost.

Another Front-End Test Also Favored DeepSeek

Developer Hamza ran another front-end generation test, and the source article reports another win for DeepSeek V4 Pro.

The generated page was described as having better overall completeness and visual quality than the Grok 4.6 result.

This reinforces the pattern seen in the earlier tasks: the two models are close, but their strengths can shift depending on the specific application and prompt.

V4 Pro Still Has Weak Spots

DeepSeek V4 Pro did not win every visual-generation test.

In one comparison involving a pelican, the V4 Pro version reportedly had stronger overall visual completion than the Flash version and produced a more appealing elephant illustration, but the pelican was moving in the wrong direction.

That is a small example, but it highlights a familiar limitation of generative coding and visual systems: a result can look polished while still containing a basic semantic or motion error.

Three.js Cherry Tree: The Gap Becomes More Visible

The source article then increased the difficulty by asking DeepSeek V4 Pro, GPT-5.6 Sol, and Claude Opus 5 to generate the same cherry tree using Three.js.

This time, the difference was more obvious.

The article reports that DeepSeek V4 Pro fell behind the other two models in areas such as trunk and branch detail, foliage, lighting, depth of field, and overall atmosphere.

That result is important because it prevents the comparison from becoming a simple victory narrative. DeepSeek V4 Pro can be extremely strong in end-to-end coding tasks, but there are still specialized visual-generation cases where other frontier models produce more refined results.

Overall: DeepSeek V4 Pro and Grok 4.6 Are Close

After several rounds of testing, the source article's overall judgment is that DeepSeek V4 Pro and Grok 4.6 have reached a similar competitive level.

Neither model wins every task.

DeepSeek V4 Pro looks particularly strong in some end-to-end coding and application-building tasks, while Grok 4.6 can be more token-efficient and showed an advantage in some visual-design and knowledge-work comparisons.

That makes the comparison more useful than a single leaderboard score. For developers, the better model may depend on whether the priority is coding quality, agent reliability, visual polish, token efficiency, or API cost.

Two AI Models Are Closing the Gap With the Frontier

The source article describes the simultaneous release of DeepSeek V4 Pro and Grok 4.6 as an unusual moment for the AI market.

Rick De Oliveira's reaction, quoted in the original article, was essentially that the two models were both powerful and affordable.

That combination is what makes the releases notable.

DeepSeek's official documentation confirms that V4 Pro is designed for tool calls, long-context workloads, and both thinking and non-thinking modes.

xAI's official Grok materials likewise position its recent Grok generation around coding, agentic tasks, and knowledge work. The official Grok 4.5 announcement, for example, describes the model as built for coding, agentic tasks, and knowledge work and reports real-world engineering evaluations.

The broader trend is clear even when individual benchmark numbers change: frontier-level AI is becoming more accessible to developers, and price-performance is becoming an increasingly important part of model competition.

What Happens Next?

The source article closes by pointing to the next round of competition.

It reports that xAI has already teased Grok 4.7, while OpenAI and Anthropic are expected to continue responding with their own model updates.

That means today's benchmark leader may not stay there for long.

For developers, the more durable lesson is to evaluate models on the tasks that actually matter to their workflow. A model that scores well on a benchmark but struggles with your repository, deployment process, UI requirements, or toolchain may still be the wrong choice.

常见问题

What is DeepSeek V4 Pro?

DeepSeek V4 Pro is DeepSeek's V4 flagship model for reasoning, coding, tool use, and long-context workloads. DeepSeek's official documentation lists a 1M-token context window and support for tool calls, along with OpenAI-format and Anthropic-format API access.

How does DeepSeek V4 Pro compare with Grok 4.6?

The source article reports that the two models are close across several agent and coding tests, with each model winning different tasks. DeepSeek V4 Pro performed especially well in several end-to-end application-building tests, while Grok 4.6 showed strong results in coding, knowledge work, and some visual-design tasks.

Is DeepSeek V4 Pro cheaper than Grok?

According to the source article's August 14, 2026 comparison, DeepSeek V4 Pro's reported output price was $0.87 per million tokens versus $6 for Grok 4.6. DeepSeek's official pricing page currently confirms $0.87 per million output tokens for V4 Pro, although API prices can change.

Can DeepSeek V4 Pro build websites and games?

The source article reports successful tests in which DeepSeek V4 Pro generated a 3D Earth experience, websites, Bento-style design collections, a 3D brick-breaker game, and a Flappy Bird-style game. These are reported real-world test results rather than a guarantee that every prompt will produce the same quality.

Is DeepSeek V4 Pro good for AI coding agents?

It is designed for tool calls and agentic workloads, and the source article reports strong results on several agent and software-engineering evaluations. DeepSeek's official documentation also lists tool-call support for V4 Pro.

Does DeepSeek V4 Pro always beat other frontier models?

No. The source article includes tests where GPT-5.6 Sol and Claude Opus 5 produced better results, particularly in a Three.js cherry-tree generation comparison. The results vary significantly by task.

What should developers compare besides benchmark scores?

Developers should also compare API cost, latency, context length, tool use, coding reliability, output quality, and how well a model fits their existing agent workflow. For production systems, consistency and total cost can matter more than a small difference on a benchmark.

相关工具

• DeepSeek API: Official API endpoint for accessing DeepSeek models.
• DeepSeek API Documentation: Official documentation for models, pricing, tool calls, and API integration.
• xAI API: Official developer resources for building with Grok models.
• Grok: xAI's user-facing Grok service for interacting with its models.
• Three.js: JavaScript 3D library relevant to the Three.js generation test described in the article.

Related Links

• DeepSeek Models & Pricing: Official DeepSeek API model specifications and current token pricing.
• DeepSeek V4 Documentation: Official DeepSeek API documentation and integration guides.
• xAI Grok 4.5 Announcement: Official xAI announcement covering the current Grok generation's coding and agent capabilities.
• xAI API Documentation: Official documentation for using Grok through xAI's developer platform.
• Three.js: Official Three.js project site and documentation.
• DeepSeek API GitHub: Official DeepSeek GitHub organization.

Summary

DeepSeek V4 Pro and Grok 4.6 are pushing frontier AI in two directions at once: stronger real-world agent and coding performance, combined with much more aggressive price-performance.

The early tests do not produce a universal winner. DeepSeek V4 Pro performed strongly in several end-to-end application-building tasks, while Grok 4.6 showed competitive coding, knowledge-work, and visual-design performance.

The bigger shift is that developers now have more capable models to choose from without automatically paying frontier-level prices.