DeepSeek's $0.87 and Grok's $6 Landed the Same Day. The Gap Is a Budget Line, Not a Benchmark

DeepSeek shipped the production build of V4 Pro on August 12 at $0.87 per million output tokens, and within hours SpaceXAI released Grok 4.6 at $6 per million, according to Cryptopolitan ^[1][2]. Both were aimed at the same workload, the long-horizon agent loop where a model keeps calling tools, editing code, inspecting its own results and returning a finished artifact, which makes the price gap an operating line item rather than a leaderboard footnote ^[25].

The arithmetic is the story. At $6 per million, a hundred million output tokens a month costs $600 on Grok 4.6; the same volume runs about $87 on V4 Pro ^[4][26]. Cryptopolitan has previously put the same workload at roughly $2,500 on Claude Opus 4.7 and $3,000 on GPT-5.5 ^[5]. The publication also puts V4 Pro's output rate at about a thirty-fifth of OpenAI's GPT-5.6 Sol and a fifty-seventh of Claude Fable 5 ^[23]. Input matters too for agents that re-read large contexts: DeepSeek lists $0.435 per million on a cache miss and $0.003625 on a hit, against Grok's $2, with Grok's fast tier double and requests over 200,000 tokens priced at $4 in and $12 out ^[7][6].

Cheap is only interesting if the work lands. On the agent evaluations DeepSeek and OpenRouter cite, V4 Pro scored 83.3 on CyberGym against 83.1 for Fable 5 and 78.3 for Claude Opus 4.8, 31.8 on AutomationBench ahead of both, and 87.9 on Terminal-Bench 2.1, above Opus 4.8's 85.0 and 0.1 behind Fable 5 ^[11][12][13]. OpenRouter, which has V4 Pro 0813 live, reports large gains over the preview, including DeepSWE 62.7, up 49.9 ^[21]. The model is a mixture of experts with 1.6 trillion total parameters and 49 billion active per token, a million-token context and a 384,000-token output ceiling, with open weights downloaded over 1.4 million times on Hugging Face in a month ^[8][9][10].

SpaceXAI's own tests put Grok 4.6 on top of GDPVal-AA v2 at 1,753 Elo, up from Grok 4.5's 1,526, and at 15.8 percent on the Harvey legal benchmark where GPT-5.6 Sol scored 2.5 percent ^[14][15]. On the independent Artificial Analysis Intelligence Index, Grok 4.6 ties GPT-5.6 Sol at 61, a point behind Fable 5 ^[16]. Token efficiency cuts against the headline price: in a Flappy Bird build, V4 Pro spent more than 20,000 tokens for $0.019 while Grok used about 5,000 for $0.03, so Grok cost about 58 percent more in dollars on a quarter of the tokens ^[17][27]. V4 Pro also botched a pelican animation, rendering the bird moving backwards ^[18].

The floor is not load-bearing. DeepSeek got to $0.87 in May by making a 75 percent cut permanent, down from $3.48, and has since said a significant increase in overall API pricing is coming ^[19][20].

What to watch: whether DeepSeek's announced price rise arrives before operators finish migrating, how many additional providers OpenRouter brings online for the 0813 build, and whether Musk's August 12 claim to staff that AI revenue would exceed all other SpaceX lines combined in September holds without Grok 4.6 defending its price ^[20][21][22].