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

What it is

An XBOW blog post evaluating Grok 4.5 (released "just yesterday" relative to the post) on their offensive-security vulnerability-discovery benchmark, positioning it against GLM 5.2, GPT-5.5, and Mythos. The summary already covers the solve-rate and dollar-cost numbers; this document carries the two things it cannot: why Grok's cost advantage exists, and how XBOW actually deploys models in practice.

What to do (beyond the headline numbers)

The one non-obvious, decision-relevant finding: Grok 4.5's cost advantage comes from pricing, not token efficiency. On a token-budget basis Grok runs close to GPT-5.5 through the middle of the curve. The dollar advantage is therefore a function of how xAI prices/caches/deploys the model — not a capability edge — and the post explicitly warns that a provider can change this at any time while the underlying model stays identical. Treat the $1–$10 sweet spot as a pricing artifact that could move, not a durable property.

Practical selection guidance the page states:

• Squeezing the first few cents: GLM 5.2 begins producing results at lower spend.
• Moderate budget, want highest observed solve rate: Grok 4.5.
• Willing to overspend for the last few points: GPT-5.5 eventually reaches a slightly higher maximum.
• Speed / strong early progress matters more than cost: Mythos, or GPT-5.5 — both beat Grok on shorter traces.

Caveats

• Grok does not win at every price point; it loses at minimal budgets (GLM) and at the highest spend (GPT-5.5), and trails on short traces.
• XBOW does not run one model — they maintain a fleet, using different models for exploit crafting, validation, and reconnaissance ("what's best at exploit crafting is not best at validation"). The takeaway isn't "switch to Grok," it's "Grok fills the middle band well."
• The author repeatedly flags limited domain expertise in the framing ("stretching an analogy").
• Numbers are approximate and drawn from XBOW's own benchmark; no methodology, iteration counts, or per-model pricing tables are given in the text.

flowchart TD
    A[What's your budget?] --> B{Few cents}
    A --> C{$1-$10}
    A --> D{Much higher}
    A --> E{Speed / early progress}
    B --> F[GLM 5.2]
    C --> G[Grok 4.5 ~75% at $1, ~90% at a few $]
    D --> H[GPT-5.5 slightly higher max]
    E --> I[Mythos / GPT-5.5]
    G -. advantage is pricing, not token efficiency .-> H