xAI put out Grok 4.5 on July 8, 2026, and the framing was different from past releases. Instead of leading with a benchmark chart, Elon Musk described it on X as "roughly comparable to Opus 4.7, but much faster" and cheaper to run. That's a strange way to sell a flagship model, but it turns out to be the accurate one.
What actually shipped
Grok 4.5 sits on a 1.5-trillion-parameter base (internally called V9) and was trained with help from Cursor, tuned specifically for long-running coding and agent tasks: multi-repo jobs, hundreds of sequential tool calls, minimal hand-holding. That partnership shows up directly in the benchmarks.
Where it lands on price and performance
On the Artificial Analysis Intelligence Index, Grok 4.5 ranks 4th out of 168 tracked models with a score of 54, behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. Rank alone undersells it, though. Here's the pricing next to those three:
| Model | Input ($/1M tokens) | Output ($/1M tokens) | Intelligence Index |
|---|---|---|---|
| Grok 4.5 | $2.00 | $6.00 | 54 |
| Claude Opus 4.8 | $5.00 | $25.00 | higher |
| GPT-5.5 | not disclosed | not disclosed | higher |
Grok undercuts Opus 4.8 by roughly 3x on input pricing and over 4x on output. It also runs at around 80 tokens/second, which is fast for a model this size.
Where it really separates itself is agentic tool-use: Grok 4.5 posted the best agentic tool-calling result of any model on the Artificial Analysis board, and it matches GPT-5.5 on the Coding Agent Index (76) inside the Codex harness while burning 3 to 4 times fewer tokens per task. If your workload is "run this agent loop until the task is done" rather than "write me the single best proof," that efficiency compounds fast.
It also topped the SWE Marathon coding benchmark outright, which specifically tests long-running, multi-step engineering work rather than one-shot code snippets. That's exactly the use case xAI tuned it for.
The catch
Two things are worth knowing before you switch. Grok 4.5 has a smaller context window than Grok 4.3 and costs more per token than its predecessor, so it isn't a free upgrade in every sense. And an independent factual-reliability benchmark found it hallucinates more confidently than before even as accuracy improves overall. The wrong answers sound more certain, which is arguably worse than the model hedging.
There's also a non-technical wrinkle. The loudest thread on Grok 4.5's Hacker News launch post wasn't about capability at all. It was about trust, specifically concerns that Musk has nudged the model's outputs on political questions. Pushback showed up in the same thread too, with people who'd tested it saying Grok felt more cautious in practice than GPT or Gemini on sensitive topics. Worth forming your own opinion here rather than taking either side at face value.
Is it worth switching to
If your workload is genuinely agentic (long tool-calling chains, multi-repo refactors, background coding jobs), Grok 4.5's price-to-performance is hard to beat right now, and Cursor's own team has reportedly made it their daily driver. If you need the single best answer to a hard one-shot reasoning question, Opus 4.8 or GPT-5.5 still rank higher and that gap is real, not noise.
For teams already inside [Claude](https://questloops.com/tools/claude) or ChatGPT workflows, Grok 4.5 is a serious enough option on cost alone to justify running a side-by-side test on your actual agent tasks before committing either way.
