Video: "Grok 4.5 IS REALLY GOOD! Opus & GPT Level BUT Faster, Cheaper, & Smarter! (Fully Tested)" by Julian Goldie on YouTube.

What Grok 4.5 actually is

Grok 4.5 is SpaceXAI's most capable model to date — a V9 architecture at 1.5 trillion parameters, launched publicly on 9 July 2026 after a short private beta inside SpaceX and Tesla. The notable thing about this release is where the training data came from. Most large language models learn from static code repositories and technical documentation. Grok 4.5 was trained partly on Cursor interaction data — real signals from how engineers actually write, revise, and debug code during live sessions. Whether that produces a meaningfully better model is something the benchmarks only partially answer, but it is a different approach worth watching.

It is available now through the SpaceXAI console, the Grok Build agent interface, and inside Cursor itself. EU availability is expected in mid-July.

What the benchmarks say — and what they do not

On the Artificial Analysis Intelligence Index, Grok 4.5 scores 54. That puts it fourth overall: Fable 5 leads at 60, followed by Claude Opus 4.8 at 56, GPT-5.5 at 55, and then Grok 4.5. In other words, it is close to the frontier but not at it. The picture in coding-specific tests is more mixed. Grok 4.5 leads Opus 4.8 on the provider-run DeepSWE 1.0 and on Terminal Bench 2.1. It trails Opus 4.8 on the independently-run DeepSWE 1.1 and on SWE Bench Pro.

That gap between provider-run and neutral benchmarks is worth flagging. SpaceXAI announced the model using the tests where it scores best — that is standard practice, but it is a reason to wait for third-party testing before committing to it as a primary model. Julian Goldie ran his own evaluation using a purpose-built testing tool and found the results genuinely competitive on structured coding tasks, with some inconsistency on open-ended builds.

The cost case is the strongest argument for it

At $2 per million input tokens and $6 per million output tokens, Grok 4.5 costs substantially less than its nearest rivals. Claude Opus 4.8 is priced at $5/$25. For coding tasks specifically, per-task cost estimates put Grok 4.5 at $2.49 versus $5.07 for GPT-5.5 in Codex and $11.80 for Fable 5 in Claude Code. Cached input comes in at $0.50 per million — a 75% discount, which matters if you are running repeated tasks against the same codebase.

If the quality holds up in real use — and that is a genuine if — those economics matter for UK businesses running AI-assisted development at any real scale. This is not a beginner model; it is designed for coding, agentic workflows, and knowledge work, and the pricing reflects that it is competing on value for professional use rather than on raw benchmark position.

One honest knock: the context window went down

Grok 4.3 had a 1-million-token context window. Grok 4.5 has 500K. SpaceXAI has not explained the trade clearly. To be fair, 500K tokens is still roughly 375,000 words — more than enough for most codebases and certainly enough for typical agentic workflows. But if you specifically need the larger context (say, processing an entire large monorepo in a single pass), this is a step backwards from the previous version, and worth knowing before switching.

Pricing also doubles for inputs over 200K tokens, which blunts the cost advantage for longer-context tasks.

Where this fits in practice for UK businesses

For teams already using AI in their development workflows, Grok 4.5 is worth testing as an alternative to Opus or GPT-5.5, particularly for structured coding tasks where cost matters and benchmark position near the frontier is more relevant than being strictly at it. The Cursor integration makes it easy to drop in as a secondary model. The EU availability gap (expected mid-July) means UK teams using EU-hosted infrastructure may need to wait a few weeks.

Worth knowing: this model is not available through the usual ChatGPT or Claude interfaces. You access it through the SpaceXAI console, the Grok Build agent, or Cursor. If your team is not already using one of those, there is a setup step before you can compare it meaningfully against what you have now.

Where this connects to NordSys

If tracking this kind of AI-agent news makes you wonder whether one could actually run inside your business, that's exactly what our AI Agents do — named, briefed and managed for you, no setup fee, from £6 a day.

See our AI Agents →