On August 12, DeepSeek released V4 Pro priced at $0.87 per million output tokens. On the same day, SpaceXAI released Grok 4.6 priced at $6 per million output tokens.
DeepSeek V4 Pro and Grok 4.6 Smartphones Launched on the Same Day
Both releases were just hours apart.
Yesterday, the company's founder, Liang Wenfeng, released the working version of V4 Pro. The preview version lasted about four months.
SpaceXAI, the AI division led by Elon Musk, released Grok 4.6 on the same day.
Both companies targeted their models for long-term work as agents. The model continues to utilize tools, modify code, analyze results, and deliver the finished product.
The performance of V4 Pro is approximately one-seventh that of Grok 4.6, one thirty-fifth of OpenAI's GPT-5.6 Sol, and one fifty-seventh of Claude Fable 5.
SpaceXAI prices Grok 4.6 at $2 per million input tokens and $6 per million output tokens. The fast version costs twice as much, and requests with long contexts exceeding 200,000 tokens cost $4 and $12 respectively.
DeepSeek surpasses this by indicating a cost of $0.435 per million for input processing on a cache miss and only $0.003625 on a cache hit.
V4 Pro is a mixture of experts system, with a total parameter count of 1.6 trillion, 49 billion active tokens per token, trained on over 32 trillion tokens.
The context window contains a million tokens, and the output ceiling is 384,000 tokens. In a month, the open weights were downloaded from Hugging Face over 1.4 million times.
V4 Pro leads in security and automation. In the CyberGym Agent test, it scored 83.3 points, compared to 83.1 for Fable 5 and 78.3 for Claude Opus 4.8. In the AutomationBench test, V4 Pro scored 31.8, outperforming both others.
V4 Pro scored 87.9 points in the Terminal-Bench 2.1 test. This result is higher than Opus 4.8 (85.0) but 0.1 point less than Fable 5.
SpaceXAI tests show that Grok 4.6 outperforms GDPVal-AA v2 with an Elo score of 1753, which is higher than Grok 4.5 (1526), and scored 15.8% in the Harvey benchmark, where GPT-5.6 Sol scored 2.5%.
Grok 4.6 shares 61st place with GPT-5.6 Sol in the AI Index, trailing Fable 5 by one point.
In building a Flappy Bird game, the V4 Pro version used over 20,000 tokens for $0.019 and provided higher detail. Grok 4.6 used about 5,000 tokens but cost $0.03.
The V4 Pro version also failed to reproduce the pelican animation. The bird seemed to move backwards.
DeepSeek V4 Pro 0813 is live on OpenRouter.@deepseek_ai reports significant performance gains for agents vs the V4 Pro preview: DeepSWE 62.7 (+49.9), CyberGym 83.3 (+30.6), NL2Repo 61.5 (+23.0), and Terminal Bench 2.1 87.9 (+15.8).
New providers joining the network soon
Try it out now: https://t.co/1PYoEFbvyK
— OpenRouter (@OpenRouter) August 12, 2026
DeFi Agents Can Operate Uninterrupted for $87 Per Month
Agents constantly executing multi-step transactions or monitoring DeFi position movements.
Cryptopolitan reported that an application generating 100 million tokens per month costs approximately $87 on the V4 Pro version compared to roughly $2,500 on Claude Opus 4.7 or $3,000 on GPT-5.5.
The service cost of V4 Pro allows making money on strategies that are loss-making in Western AI systems.
In May, DeepSeek reached the $0.87 mark, then permanently lowered the price by 75%, dropping from $3.48 to below-market levels, as Cryptopolitan reported at the time.
DeepSeek stated a "significant increase" in overall API prices is coming, so the current rock-bottom price level may be unsustainable.
In a message to employees on August 12, Musk stated that AI revenues in September will surpass all other SpaceX business lines combined and will account for 99% of the company's value in five years.
