DeepSeek and Grok launch new AI models as pricing war intensifies

DeepSeek and Elon Musk's SpaceXAI released flagship AI models on August 12 that target agent-based long-horizon work, a category focused on letting models call tools, edit code, verify outputs and deliver finished products rather than simple answers. The simultaneous launches of DeepSeek V4 Pro and Grok 4.6 point to a sharp drop in the cost of using frontier AI. DeepSeek priced V4 Pro at $0.87 per million output tokens, while Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens. The article says that puts DeepSeek at about one-seventh the cost of Grok 4.6 on output pricing, one thirty-fifth the cost of GPT-5.6 Sol and one fifty-seventh the cost of Claude Fable 5, while Grok 4.6 is about half the price of other frontier models on published API rates. Benchmark results in the article place both systems in the top tier, but with different strengths. DeepSeek V4 Pro posted 83.3 in CyberGym, ahead of Fable 5 at 83.1 and Claude Opus 4.8 at 78.3, and scored 31.8 in AutomationBench versus 29.1 for Fable 5 and 27.2 for Opus 4.8. In Terminal-Bench 2.1, V4 Pro reached 87.9, above Opus 4.8's 85.0 and just below Fable 5's 88.0. Grok 4.6 led in several knowledge-work tests, scoring 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and trailing Fable 5 by one point, while reaching 1753 Elo in GDPVal-AA v2 against 1741 for Fable 5 and 1728 for GPT-5.6 Sol. In Harvey LAB, a legal-task benchmark, Grok 4.6 scored 15.8% versus 11.3% for Fable 5 and 2.5% for GPT-5.6 Sol. The piece says early developer testing suggests the practical gap between the two models is narrow. In tests collected from tech communities, DeepSeek V4 Pro built an interactive 3D Earth in a browser from a single prompt, including drag, rotation, zoom, global data streams, dynamic flight routes and geographic markers. In a 3D brick-breaker game task, the two models were effectively tied, while Grok 4.6 showed a slight advantage in a 60-variant Bento card design task. In a Flappy Bird build test by developer Jun Song using identical prompts, DeepSeek V4 Pro used more than 20,000 tokens and cost $0.019, while Grok 4.6 used about 5,000 tokens and cost $0.03. The article says V4 Pro delivered better scene layering, more detailed interaction feedback and stronger end-to-end completion in that test. It also notes weaknesses: V4 Pro reversed a pelican animation's movement direction and lagged GPT-5.6 Sol and Claude Opus 5 in a Three.js comparison involving a cherry blossom tree. Grok 4.6 is presented as part of a wider push by Musk into AI products and infrastructure. SpaceXAI launched Grok Bot on August 11 with Cursor, describing it as an AI agent product that runs on its own cloud computer, logs into existing tools including services without API access, completes work across applications and inboxes, and can operate in parallel through multiple Bots. The article links that launch to SpaceX's $60 billion agreement in June to acquire Cursor and says Cursor's real-world programming usage data likely helped improve the Grok family. Grok 4.6 has already been integrated into Cursor and Grok Build and is distributed through OpenRouter, Vercel and Cloudflare. The article also highlights Musk's commercial projections. In an internal all-hands meeting address released early on August 12, he said AI revenue will exceed all other SpaceX revenue combined in September and surpass it substantially in the fourth quarter. He projected 10 gigawatts of AI compute capacity by the end of next year, which he said would imply annual revenue of $300 billion to $500 billion based on estimated value of $30 to $50 per watt. Musk said, "Five years from now, AI will be 99% of SpaceX's value," adding that "SpaceX's value will be an astronomical number." He also said Grok 4.7 will arrive within weeks and is expected to have 2.1 trillion parameters. The broader implication is that a frontier-model price war is moving into the agent era. Gavin Baker, Chief Investment Officer at Atreides Management, said Grok 4.6 performs roughly at the level of Fable 5 Max while offering input token prices 80% lower and output token prices 88% lower. Kim Monnis argued Grok 4.6 is cheaper than Opus 5 and GPT-5.6 Sol and even below Sonnet 5 while still operating at a top-tier level. The article says that dynamic could affect rivals including Anthropic, which is reportedly planning an IPO in September or October, and OpenAI, which has confidentially filed for an IPO. It also flags implications for Microsoft, which holds about 27% of OpenAI valued at nearly $135 billion and recorded a $3.2 billion gain from its Anthropic investment in the fourth quarter of fiscal 2026. The piece cautions that published API prices do not translate directly into final task costs because token use, inference intensity and completion paths differ by model. Still, it argues that as capability gaps narrow, the trade-off between performance and inference cost is becoming the decisive factor for developers and enterprises.

本网站上的信息是使用AI生成的,我们无法保证其准确性。 请仅作为参考信息使用。
DeepSeek and Grok launch new AI models as pricing war intensifies - CoinPost Terminal