SpaceXAI released Grok 4.5 on Wednesday. Elon Musk describes it as an Opus-class model, offering a cheaper and more efficient alternative to other powerful AI models.
SynthePulse Insight · AI deep reading
Version 1 · 3 sources
SpaceXAI releases Grok 4.5, priced at a fraction of competitors, yet still trailing top models on key coding benchmarks. Its real competitiveness may lie not in performance, but in extremely low cost and efficient token usage.
Grok 4.5 demonstrates near-frontier performance on some benchmarks. On Terminal Bench 2.1, it scores 83.3%, just 0.1 percentage points behind GPT-5.5 and 1 point behind Fable 5. However, on the more challenging DeepSWE 1.1, Grok 4.5 scores only 53%, trailing GPT-5.5 (67%) and Fable 5 (70%) by over 10 percentage points. On SWE Bench Pro, it scores 64.7%, below Fable 5's 80.4% and Opus 4.8's 69.2% (max setting).
On Artificial Analysis' Intelligence Index, Grok 4.5 ranks fourth, behind Fable 5, GPT-5.5, and Opus 4.8, but improved by 16 points over Grok 4.3. On the Coding Agent Index, Grok 4.5 scores 76 in Grok Build, tied with GPT-5.5 in Codex and just 1 point below Fable 5 in Claude Code. Elon Musk calls Grok 4.5 an 'Opus-class model' and claims internal evaluations find it roughly on par with Opus 4.7, but faster.
Grok 4.5's pricing is aggressive: $2 per million input tokens and $6 per million output tokens. In comparison, Opus 4.8 is $5/$25, GPT-5.5/5.6 is $5/$30, and Fable 5 is $10/$50. xAI claims Grok 4.5's token efficiency is twice that of other models, using 4.2 times fewer tokens than Opus 4.8 on SWE Bench Pro tasks.
On the Coding Agent Index, Grok 4.5 costs only $2.49 per task in Grok Build, compared to $5.07 for GPT-5.5 in Codex and $11.80 for Fable 5 in Claude Code. Grok 4.5 consumes an average of 1.9 million tokens per task, far less than GPT-5.5's 6.2 million and Fable 5's 7.2 million. This pricing strategy mirrors that of Chinese vendors (e.g., Zhipu and DeepSeek): win on price when performance is close.
Grok 4.5 is SpaceXAI's (formerly xAI) first enterprise-focused model, targeting coding, agent tasks, and knowledge work. The model was co-trained with Cursor, a code editor acquired by SpaceXAI for $60 billion in stock, using trillions of tokens of Cursor data. Grok 4.5 is now available via Grok Build, Cursor, and the SpaceXAI console, and plugins for Word, PowerPoint, and Excel have been released.
Analysts note that Grok 4.5's release marks SpaceXAI's strategic shift from consumer models on X (formerly Twitter) to the enterprise market. Omdia analyst Lian Jye Su believes this move aims to drive Cursor adoption among developers. Futurum Group analyst Bradley Shimmin says more competition will give enterprises more choices. However, Su also notes that token cost is unlikely to be a decisive factor for enterprises choosing Grok 4.5; companies accustomed to Claude are unlikely to switch solely due to price.
Despite improvements in accuracy (from 35% to 52%), Grok 4.5's hallucination rate jumped from 25% to 54%. This means the model knows more but is also more confident when wrong. This weakness could affect its reliability in enterprise-critical tasks.
Additionally, Grok 4.5 is not yet available in the EU; xAI plans to launch there in mid-July. This limits its availability in the European market. The model was trained on tens of thousands of Nvidia GB300 GPUs, employing extensive data filtering, deduplication, and domain-specific selection strategies, with reinforcement learning covering hundreds of thousands of tasks, primarily from software engineering.
This article is based on reports from THE DECODER, AI Business, and TechCrunch, as well as independent benchmark data from Artificial Analysis. All performance data and pricing information come from xAI's official releases or third-party evaluators. Hallucination rate data is from Artificial Analysis, an independent source. Analyst opinions are from Omdia and Futurum Group.
Grok 4.5 enters the enterprise market with extremely low pricing and efficient token usage, approaching frontier levels on some coding benchmarks but still lagging on complex tasks. Its real competitiveness lies in cost advantage rather than absolute performance. However, a doubled hallucination rate and regional restrictions may hinder enterprise adoption.
Primary report
Primary source
Also covered by 2 sources