SpaceXAI released Grok 4.5 on July 8, 2026, securing the fourth position on the Artificial Analysis Intelligence Index. The model achieved a score of 54, trailing behind Fable 5, GPT-5.5, and Opus 4.8. This release represents a 16-point improvement over the predecessor Grok 4.3, positioning SpaceXAI near the intelligence frontier alongside OpenAI and Anthropic. Grok 4.5 features 1.5T parameters, making it 3x larger than Grok 4.3, though its context window has been reduced to 500k tokens from the previous 1M.

In agentic coding tasks, Grok 4.5 within the Grok Build harness scored 76 on the Artificial Analysis Coding Agent Index. This performance is on par with GPT-5.5 using Codex and slightly below Fable 5 using Claude Code. Notably, Grok 4.5 demonstrated high efficiency, using an average of 1.9M tokens per task, compared to 7.2M for Fable 5 and 6.2M for GPT-5.5. The associated cost of $2.49 per coding task is significantly lower than the $11.80 for Fable 5 and $5.07 for GPT-5.5.

The model's pricing is set at $2 per 1M input tokens and $6 per 1M output tokens, with cache hits discounted by 75% to $0.5 per 1M tokens. Costs double for inputs exceeding 200k tokens. While Grok 4.5 shows strong agentic performance in knowledge work, terminal usage, and customer service, it also reflects a common trend in larger models: accuracy on the AA-Omniscience Index rose from 35% to 52%, but the hallucination rate simultaneously increased from 25% to 54%. Overall, Grok 4.5 maintains a cost-effective profile, costing $0.31 per task on the Intelligence Index, which is 5x lower than Claude Sonnet 5 while outperforming it on intelligence benchmarks.

SpaceXAI released Grok 4.5 on July 8, 2026, securing the fourth position on the Artificial Analysis Intelligence Index. The model achieved a score of 54, trailing behind Fable 5, GPT-5.5, and Opus 4.8. This release represents a 16-point improvement over the predecessor Grok 4.3, positioning SpaceXAI near the intelligence frontier alongside OpenAI and Anthropic. Grok 4.5 features 1.5T parameters, making it 3x larger than Grok 4.3, though its context window has been reduced to 500k tokens from the previous 1M.

In agentic coding tasks, Grok 4.5 within the Grok Build harness scored 76 on the Artificial Analysis Coding Agent Index. This performance is on par with GPT-5.5 using Codex and slightly below Fable 5 using Claude Code. Notably, Grok 4.5 demonstrated high efficiency, using an average of 1.9M tokens per task, compared to 7.2M for Fable 5 and 6.2M for GPT-5.5. The associated cost of $2.49 per coding task is significantly lower than the $11.80 for Fable 5 and $5.07 for GPT-5.5.

The model's pricing is set at $2 per 1M input tokens and $6 per 1M output tokens, with cache hits discounted by 75% to $0.5 per 1M tokens. Costs double for inputs exceeding 200k tokens. While Grok 4.5 shows strong agentic performance in knowledge work, terminal usage, and customer service, it also reflects a common trend in larger models: accuracy on the AA-Omniscience Index rose from 35% to 52%, but the hallucination rate simultaneously increased from 25% to 54%. Overall, Grok 4.5 maintains a cost-effective profile, costing $0.31 per task on the Intelligence Index, which is 5x lower than Claude Sonnet 5 while outperforming it on intelligence benchmarks.