.png)
April 21, 2025
Competition in the field of Artificial Intelligence (AI) continues to intensify, especially regarding the cost of AI models. xAI, Elon Musk's AI company, has introduced a new language model, Grok 3 Mini, which is causing a stir due to its combination of performance and low cost. Grok 3 Mini is part of the Grok 3 family, which includes the main model Grok 3, a faster variant, Grok 3 Fast, and four versions of Grok 3 Mini. The latter are available in slow and fast versions, each with low or high reasoning capabilities.
xAI emphasizes the speed and cost-effectiveness of Grok 3 Mini. In contrast to the larger Grok 3, which operates without explicit reasoning, Grok 3 Mini has an integrated reasoning process. According to xAI, Grok 3 Mini achieves top scores in benchmarks for university-level math, programming, and science, while being up to five times cheaper than other reasoning models. In some areas, it is even said to outperform more expensive flagship models. Its performance in mathematics is particularly highlighted, where Grok 3 Mini achieved a top score of 93% in the AIME 2024.
A special feature of Grok 3 Mini is the provision of a complete "Reasoning Trace" with each API response. This is intended to give developers more insight into the model's behavior. However, it should be noted that these apparent "thought processes" can sometimes be misleading, according to current research.
While Grok 3 Mini is the latest addition to the model lineup, both Grok 3 and Mini are now accessible to developers via the xAI API. Integration into existing toolchains is intended to facilitate adoption. Grok 3 continues to target demanding tasks that require deep world knowledge and expertise. xAI describes it as the most powerful available model without a dedicated reasoning component.
The team at Artificial Analysis has benchmarked the Grok 3 family and highlighted Grok 3 Mini Reasoning (high) due to its price-performance ratio. According to their "Artificial Analysis Intelligence Index," Grok 3 Mini Reasoning (high) outperforms models like Deepseek R1 and Claude 3.7 Sonnet (64k Reasoning Budget) – at significantly lower costs.
With a price of $0.3 per million input tokens and $0.5 per million output tokens, Grok 3 Mini is almost an order of magnitude below models like OpenAI's o4-mini or Google's Gemini 2.5 Pro. For applications requiring higher speed, a faster version is available for $0.6 and $4 per million tokens, respectively. Grok 3 Mini achieves a value of approximately 67 on the Intelligence Index at low cost.
The results focus on the metric "intelligence," which combines six different benchmarks. A detailed breakdown for each individual benchmark is in progress. However, it should be noted that benchmark results do not necessarily reflect real-world performance. In particular, smaller models can deliver impressive numbers that are not always confirmed in everyday use.
In terms of pure speed, Grok 3 surpasses its more reasoning-focused Mini counterpart: On standard endpoints, Grok 3 generates 500 tokens in about 9.5 seconds, while Grok 3 Mini Reasoning takes 27.4 seconds.
Artificial Analysis places Grok 3 and Grok 3 Mini Reasoning (high) in the top 5 of their respective categories – Non-Reasoning and Reasoning – and notes that xAI has firmly established itself among the leading providers in the current AI model landscape with these releases. The increasing price pressure in the AI field, especially after Google's recent cost reduction with Gemini 2.5 Flash, is further intensified by Grok 3 Mini.
Bibliographie: https://the-decoder.com/grok-3-mini-turns-up-the-heat-as-ai-price-wars-push-model-costs-even-lower/ https://twitter.com/theaitechsuite/status/1913896505687101575 https://the-decoder.com/grok-3-the-new-model-from-musks-xai-wins-praise-from-openai-founder/ https://www.facebook.com/THEDECODERAI/posts/one-of-the-most-compelling-results-in-recent-o3-benchmarks-comes-from-its-perfor/638996225620123/ https://x.ai/blog/grok-3 https://medium.com/@cognidownunder/the-ai-showdown-grok-3-vs-o3-mini-a-tale-of-two-titans-fdaebe4dc8da https://www.reddit.com/r/singularity/comments/1isishj/grok_3_not_performing_well_in_real_world/ https://www.linkedin.com/pulse/gen-ai-business-52-we-1-eugina-jordan-adw9e https://garymarcus.substack.com/p/grok-3-hot-take https://apidog.com/blog/grok-3-grok-3-mini-api-free