SpaceXAI Debuts Grok 4.5: A Cheap Cursor-Trained AI That Can't Quite Beat Rivals

The training of this new model relied on a massive database of real-world developer interactions, courtesy of Cursor's user base. Instead of just reading static code like a college student cramming before an exam, Grok 4.5 was fed trillions of tokens showing how actual humans write, debug, and desperately fix their broken software. The engine running this massive operation is the V9 fundamental model, packing roughly 1.5 trillion parameters.

To get this monster up and running, the team burned through electricity using tens of thousands of NVIDIA GB300 GPUs. The result of all this expensive silicon-melting is a Mixture-of-Experts architecture that is surprisingly built for speed and budget rather than pure academic dominance.

When it comes to actual performance, the benchmark results are a mixed bag of corporate excuses. On the Terminal Bench 2.1, Grok 4.5 hit 83.3%, barely losing to GPT-5.5 at 83.4% and Fable 5 at 84.3%. However, in the more complex SWE Bench Pro coding test, the model scored 64.7%, easily beating OpenAI's 58.6% but falling behind Claude's 69.2%.

The real selling point here is the price tag rather than raw brainpower. SpaceXAI is pricing input tokens for Grok 4.5 at more than four times cheaper than its high-end competitors. It seems the strategy is to flood the market with cheap, 'good enough' intelligence rather than winning prestigious AI beauty pageants.

For those eager to try it, the model is already integrated into the Cursor editor across all paid tiers and is available via the Grok Build agent platform. The only people left out of the party are developers in the European Union, who get to sit back and watch due to local regulatory blockades.

Spending sixty billion dollars on a text editor just to release a budget-friendly chatbot is the peak of modern tech hubris. SpaceXAI is betting that developers care more about saving a few pennies per API call than having the absolute smartest assistant on the planet. It remains to be seen if the coding community will embrace this bargain-bin genius or stick to the premium brains of Anthropic.