On one side, Liang Wenfeng dropped the official DeepSeek V4 Pro release in the early hours of the morning. On the other, Elon Musk unveiled his next-generation flagship, Grok 4.6. Two AI heavyweights, launched at virtually the same time — and the rivalry is intense.

Late on August 12, Beijing Time, DeepSeek’s official API documentation was quietly updated, with the model version switching from the preview release to DeepSeek-V4-Pro-0813. Almost simultaneously, Musk’s xAI officially released Grok 4.6.

Both models are targeting the same frontier: enabling AI to handle long-running tasks by continuously calling tools, modifying code, validating results, and ultimately delivering a genuinely usable product.

Matching Fable 5: DeepSeek Comes Within 0.1 Points of the Global Lead

The most striking results from the official DeepSeek V4 Pro release came from two benchmarks, where it took the top spot outright.

On CyberGym, a benchmark for cybersecurity agents, DeepSeek scored 83.3, edging past Fable 5 at 83.1. On AutomationBench, which evaluates automated task execution, it scored 31.8, ahead of Fable 5’s 29.1.

The closest battle came on Terminal-Bench 2.1. DeepSeek scored 87.9, just 0.1 points behind Anthropic’s Fable 5, which leads the global leaderboard at 88.0.

The jump from the preview version is particularly striking: DeepSeek went from 72.1 to 87.9.

Its software-engineering agent, DeepSWE, also surged from 12.8 in the preview version to 62.7, nearly five times its previous score.

Grok 4.6 has also entered the top tier. It scored 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol Max and placing third globally.

On GDPval-AA v2, a benchmark for knowledge-work tasks, Grok 4.6 achieved an Elo score of 1,753, surpassing GPT-5.6 Sol Max at 1,728 and Claude Fable 5 at 1,741.

On the professional legal-task benchmark Harvey LAB, Grok 4.6 scored 15.8%, compared with 11.3% for Fable 5.

A 57× Cost Gap: $0.87 vs. $50

Performance is only half the story. Pricing is where this showdown gets truly explosive.

For every one million output tokens, Grok 4.6 costs $6, GPT-5.6 Sol costs $30, Claude Opus 5 costs $25, and Fable 5 costs $50.

DeepSeek V4 Pro, by comparison, costs just $0.87 — roughly one-seventh the price of Grok, one-thirty-fourth the price of GPT-5.6 Sol, and one-fifty-seventh the price of Fable 5.

For input tokens, DeepSeek V4 Pro charges RMB 3 per million tokens for cache misses, while output costs RMB 6 per million tokens. Grok 4.6 charges $2 per million input tokens and $6 per million output tokens.

Musk’s Grok 4.6 has already pushed pricing down to roughly half that of some other frontier models.

DeepSeek, however, has taken the price war all the way to the floor.

Some analysts have described the shift this way: developers once had to choose between models that were too expensive and models that were not capable enough.

Now, DeepSeek is challenging that trade-off with frontier-level performance at a fraction of the cost, while Grok is following closely with an aggressive price-to-performance ratio.

Top-tier AI intelligence is no longer necessarily out of reach.

First Real-World Tests: Same Prompt, Two Models Go Head-to-Head

Beyond benchmark scores, the first wave of real-world testing is already emerging.

AI creator Xiangyang Qiaomu gave the same prompts to both models.

In a Flappy Bird development test, DeepSeek V4 Pro consumed more than 20,000 tokens, but cost just $0.019. Grok 4.6 used only around 5,000 tokens, yet cost approximately $0.03.

But when it came to the final product, DeepSeek appeared to have the edge.

The game featured a mountain landscape and layered clouds, while the pipes included gradients and a stronger sense of depth. When the character passed through the pipes, the game even displayed a floating “+1” animation.

In another frontend development test, DeepSeek V4 Pro again outperformed Grok 4.6.

That said, V4 Pro is far from flawless.

In one pelican comparison test, its elephant design was considered more visually appealing, but the pelican’s movement direction was completely reversed.

After several rounds of real-world testing, one thing is becoming increasingly clear:

DeepSeek V4 Pro and Grok 4.6 are now operating on remarkably similar performance ground — and the gap between frontier models is getting harder to ignore.

战争远未结束

马斯克已经提前预告，Grok 4.7马上就来，目标直指超越所有顶尖AI。DeepSeek这边还有一张底牌未出——备受关注的Harness团队公众号已于7月注册，内测用户称参与深度体验两天后愿称之为宇宙最强。

身处这个以小时为单位进化的时代，智能不再是少数巨头的特权。梁文锋的DeepSeek和马斯克的Grok，在同一天共同改写了一条规则：高性能与低成本，可以同时存在。属于所有人的应用大爆发，才刚刚开始。