SpaceXAI (the company most people still call xAI) launched Grok 4.5 on July 16, 2026, and aimed it squarely at developers instead of casual chat users. It was trained alongside [Cursor](https://cursor.com), and it now shows up as an option in Cursor on every plan, in Grok Build (SpaceXAI's CLI), and through the API console.
Where Grok 4.5 actually wins
On Artificial Analysis's Intelligence Index, Grok 4.5 ranked #4 out of 168 tracked models with a score of 54, and it posted the best agentic tool-use result of any model on the board. On SpaceXAI's own SWE Marathon benchmark it resolved 29.0% of tasks on the first pass, ahead of [Claude](https://www.questloops.com/tools/claude) Opus 4.8 (max) at 26.0%.
It's not a clean sweep, though. On SWE Bench Pro, a competing model still leads at 80.4% and Opus 4.8 follows at 69.2%, with Grok 4.5 at 64.7%. SpaceXAI's real pitch is efficiency: the company says Grok 4.5 resolves SWE Bench Pro tasks using about 4.2 times fewer output tokens than Opus 4.8 (max), averaging 15,954 tokens per task versus 67,020.
Pricing: this is where it gets interesting
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, with a 75% cache discount that brings cached input down to $0.50 per million. Here's how that lines up against the two models it's most often compared to:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Grok 4.5 | $2.00 | $6.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| [GPT-5.6](https://www.questloops.com/blog/openai-s-gpt-5-6-explained-sol-vs-terra-vs-luna-and-which-one-to-use) Sol | $5.00 | $30.00 |
Grok 4.5 undercuts both on sticker price, and SpaceXAI's efficiency claims mean the real-world gap is likely bigger than the table shows, since it's also using fewer tokens to finish the same task.
How to try Grok 4.5 for free
SpaceXAI is currently offering free Grok 4.5 usage for a limited time, and this is the easiest way to test it before paying for API access:
- Install Grok Build with `curl -fsSL https://x.ai/cli/install.sh | bash`, which gives you free access to Grok 4.5 through the CLI while the promotion lasts.
- If you already use [Cursor](https://cursor.com), select Grok 4.5 from the model picker; it's included on every Cursor plan rather than gated behind a higher tier.
- For API testing beyond the free window, cached input at $0.50 per million tokens is the cheapest way to iterate on the same prompt repeatedly.
There's no fixed end date published for the free window, so treat it as a limited-time offer rather than a permanent freebie.
The catch nobody's benchmark chart shows
Grok 4.5's Hacker News launch thread spent more energy debating trust than benchmarks. The loudest complaint was about Elon Musk allegedly steering the model's answers on political questions, and it dominated the discussion ahead of any capability comparison. A counterpoint showed up in the same thread: one commenter argued that in their own testing, Grok has actually been more cautious on politically charged prompts than GPT or Gemini. Cursor's CEO, for what it's worth, reportedly called Grok 4.5 his team's daily driver.
If your use case is straightforward coding and agentic tool-calling, none of that matters much. If you're building something where perceived neutrality is part of the product, it's worth reading that thread yourself before you commit.
Frequently asked questions
**Is Grok 4.5 free?** Not permanently. SpaceXAI is running a limited-time free usage window in Grok Build and Cursor; the API itself is metered at $2/$6 per million tokens.
**Is Grok 4.5 better than GPT-5.6 or Claude Opus 4.8?** On price and token efficiency, yes. On raw benchmark scores it trades wins and losses depending on the task, with SWE Marathon favoring Grok 4.5 and SWE Bench Pro favoring other models.
**Do I need an X or Twitter account to use it?** No. Grok Build, the API console, and Cursor's integration don't require an X account, only grok.com and the X app do.
