Grok 4.5: SpaceXAI suddenly delivers a serious coding model
Near-frontier benchmark levels, a fraction of the cost, and a surprisingly honest mishap. What the new model can do and where it reaches its limits.
SpaceXAI was recently viewed more as a GPU vendor than a serious model provider. With Grok 4.5, that changes abruptly. The model competes near the very top of coding benchmarks while costing only a fraction of its competitors.
According to SpaceXAI, Grok 4.5 is the company's most powerful model to date, built for coding, agentic tasks, and knowledge work. It was trained jointly with Cursor on tens of thousands of NVIDIA GB300 GPUs, with a strong focus on data quality rather than raw token volume. The RL training included hundreds of thousands of tasks, centered on multi-step software development.
The numbers are not marketing smoke. On DeepSWE 1.0, Grok 4.5 ranks third behind Fable and GPT 5.5, well ahead of Opus 4.8. On Terminal Bench 2.1, all three are practically tied. And on SWE Marathon, Grok 4.5 even leads the pack.
| Model | DeepSWE 1.0 | Terminal Bench 2.1 |
| Fable (max) | 66.1% | 84.3% |
| GPT 5.5 (xhigh) | 64.31% | 83.4% |
| Grok 4.5 | 62.0% | 83.3% |
| Opus 4.8 (max) | 55.75% | 78.9% |
The real argument, however, is the price: $2 per million input tokens, $6 per million output tokens. Fable charges $10 and $50. On top of that comes token efficiency that leaves everything else looking outdated: Grok 4.5 solves SWE Bench Pro tasks using an average of around 16,000 output tokens; Opus 4.8 requires over 67,000. That is well over four times fewer, at an output speed of 80 tokens per second.
One downer regarding the price: from 200,000 tokens of context onward, the costs double, with a maximum context window of 500,000 tokens. Charging double for this comparatively modest context gain feels contrived. Even so, the model remains cheaper than the direct competition.
The transparency surrounding a genuine mishap is remarkable. An earlier snapshot of the Cursor codebase inadvertently ended up in the training data. As a result, the proprietary Cursor Bench is worthless for Grok 4.5, and Cursor communicates this openly instead of promoting the skewed result. That deserves respect. Mistakes happen; how they are handled is what matters.
How does the model perform in practice? A developer with 24 hours of early access reported running a security audit on their cloud product shortly before launch: the model analyzed pull requests, identified precisely what was resolved and what was not, created two separate PRs in a single run, processed review comments, and even accurately implemented requested changes from a simple screenshot. Even with deliberately sloppy instructions, it stayed on track, where other models tend to get stuck on old messages.
Price-Performance
Around 31 cents per task in the Artificial Analysis test suite, compared to $2.75 for Fable. Achieved an Intelligence Index score of 54 and fourth place overall.
3D Surprise
When converting a 2D game to 3D with three.js, Grok 4.5 modeled the environment and creatures better than any model tested so far. Not perfect, but usable for the first time.
Orchestration Weakness
When delegating to sub-agents, it lacks the finesse of the latest generation. The model occasionally got stuck in hanging processes and failed to clean up cleanly afterward.
Not Strong Everywhere
In individual reasoning benchmarks, it finished as the weakest frontier model and proved relatively expensive because it reasons very extensively.
The most honest appraisal comes from a gaming analogy in the review: Grok 4.5 is the greatest PS2 game of all time, two months after the PS3 was released. It does not quite match the orchestration capabilities of Fable or GPT 5.6. But what has been extracted here from the previous technology generation is remarkable.
Grok 4.5 is available immediately in Grok Build, in Cursor across all plans, and via the SpaceXAI API. The model is not yet available in the EU; clearance is expected in mid-July.
Real Competition at Last
Grok 4.5 is not a Fable killer. It is something more important: a very good, very affordable standard model for coding that puts price pressure on OpenAI and Anthropic. When the result is faster, cheaper, and better models for everyone, we users are the ones who win. And that is exactly what it is all about.