Grok 4.5 Poised To Take Over the Middle of the AI Security Market

Grok 4.5 performs exceptionally well when you have more than a shoestring budget but still care about cost. It does not win at every price point, but across the broad middle-to-medium-high of the range, it is currently the strongest option we tested.

In our recent analysis of affordable AI models, like GLM and Muse Spark, I argued that asking which model is “best” is usually the wrong question. The more useful question is which model is best for the budget, time, and work you actually have.

We have now added Grok 4.5 (released just yesterday) to the comparison, making the playing field more crowded, and making the point stronger.

Grok 4.5 Owns the Middle

At a fixed number of iterations, Grok 4.5 performs like a frontier model. It rapidly improves through the early and middle stages of a run, eventually solving roughly 93% of the vulnerabilities in our benchmark. That puts it slightly ahead of Mythos and GPT-5.5 within the tested range, although on shorter traces either of those two are more effective.

But the more interesting result appears when we account for cost.

At very small budgets, there are cheaper ways to make progress. GLM 5.2, for example, begins producing results at a lower spend. If you are trying to squeeze every possible result out of the first few cents, Grok is not the obvious choice.

Once the budget reaches the middle of the range, however, Grok 4.5 takes over. Between roughly $1 and $10, it consistently delivers the highest observed solve rate of the models we tested. Around $1, it reaches approximately 75%, compared with about 65% for the nearest alternatives. At a few dollars, it approaches 90%.

That is a meaningful advantage in the part of the curve where many practical workloads are likely to operate.

The Sensible Sports Car

In the last post, I compared model selection to choosing a car. You do not always need the Ferrari. Sometimes you need the Toyota that gets you there cheaply and reliably.

Grok 4.5 may be something in between. A sports car that is surprisingly practical, perhaps. (Again, I am stretching an analogy in a field where I have very little expertise. I do know enough to state that it's probably not a Cybertruck, though. For one, it appears to be usable in practice.)

It is not the cheapest vehicle off the lot. If the only money you have is loose change from underneath the seat, GLM is still attractive. But Grok gives you considerably more performance without immediately entering the most expensive part of the market.

And unlike some models that are optimized for a particular budget band, Grok does not fall apart when you let it run longer. Its advantage narrows at higher spend, where GPT-5.5 eventually achieves a slightly higher maximum solve rate, but Grok remains highly competitive.

What the Token Results Tell Us

On a token-budget basis, Grok runs relatively close to GPT-5.5, particularly through the middle of the curve.

Its cost advantage therefore, appears to come substantially from pricing rather than from a radical improvement in token efficiency.

Model capability and model economics are related, but they are not the same thing. A provider can change the practical value of a model through pricing, caching, or deployment efficiency even when the underlying model stays exactly the same.

Today, Grok 4.5 occupies a particularly useful combination of capability and price.

So Which Model Should You Use?

The answer remains annoying: it depends.

If you are pinching pennies and can tolerate a model working more slowly or less consistently, GLM 5.2 may still be the rational choice.

If you want the highest solve rate we observed across a moderate budget, Grok 4.5 is currently the model to beat.

If you are prepared to spend much more to chase the final few percentage points, GPT-5.5 eventually pulls slightly ahead. Mythos also remains extremely capable, particularly when speed and strong early progress matter more than minimizing spend.

The important point is not that Grok 4.5 is now the universally “best” model. There is no universally best model. Not in general, not for cybersecurity, not even for one single cybersecurity company: Like every credible contender in a complex space, XBOW maintains a fleet of models and uses them for different purposes: what’s best at exploit crafting is not best at validation, and reconnaissance is best served by yet another model.

Grok 4.5 fills the middle of the market unusually well for technical agentic tasks. It offers near-frontier offensive capability at a price where running it repeatedly and at scale begins to look practical.

That makes it a very good model—and, for a large range of budgets, perhaps the most rational one.