Grok 4.5, xAI's new coding and agent model released on July 8, 2026, does not have a single price. The same model costs $2 per million input tokens on the direct application programming interface (API), but significantly more through Cursor, the code editor that jointly trained the model with xAI. The pricing confusion stems from a fundamental reality: Grok 4.5 reaches buyers through three entirely different control planes, each with separate pricing, usage limits, data handling, and contract terms.

Why Does the Same Model Cost Different Amounts?

The answer lies in how xAI distributes Grok 4.5. On the direct xAI API, pricing is transparent: $2.00 for input tokens, $0.50 for cached input, and $6.00 for output tokens per million processed. But Cursor, the integrated development environment that helped train the model, bundles Grok 4.5 into its paid plans with qualitative usage language rather than numeric allowances. Cursor's fast variant lists $4 input and $18 output per million tokens, which is a different mechanism entirely from xAI's Priority Processing feature.

The confusion deepens with Grok Build, xAI's command-line agent tool. While xAI's launch page states that Grok 4.5 is the default model in Grok Build, the pricing catalog separately lists a Code API model named grok-build-0.1 with a 256,000-token context window and rates of $1 input, $0.20 cached input, and $2 output per million tokens. These pages may describe different layers of one product, but the public documentation does not clearly map which identifier appears on a team's invoice.

What Hidden Costs Sit Above the Sticker Price?

The base pricing is only the beginning. Four additional cost layers can multiply the final bill. Priority Processing doubles input, cached, reasoning, and output rates when priority service is delivered. Server-side tools add per-invocation charges. Requests above 200,000 tokens use a rate that the model page references but does not publicly display. Additionally, Grok 4.5 does not appear in xAI's batch-discount table, meaning bulk processing does not receive volume pricing.

On top of those, failed runs, retries, human repair, and Cursor's on-demand overage charges are workload-specific and do not appear in any rate card. For teams using Cursor, the advertised $20 monthly plan can trend toward roughly $60 in daily agent use, according to independent analysis.

How to Compare Grok 4.5 Across Different Access Routes

• Direct xAI API: Best for teams that want a 500,000-token context window and run their own agent loop. Offers the clearest pricing visibility and per-token billing, with contracting through X.AI LLC. Requires managing infrastructure and agent orchestration independently.
• Grok Build: Harder to evaluate because the public documentation does not clearly map which model identifier (grok-4.5, grok-build-latest, or grok-build-0.1) appears on invoices. Teams should confirm which identifier their contract uses before assuming pricing or context window specifications.
• Cursor: Ideal for existing Cursor users who value the editor and agent workflow as much as the model itself. Includes qualitative usage allowances on paid plans, but does not publish a numeric Grok-specific quota or clarify how retries consume the included allocation.

The first buying decision is not "Grok versus another model," but rather which access path fits your team's infrastructure, budget tracking, and data governance requirements.

How Does Grok 4.5 Actually Perform on Engineering Tasks?

xAI claims Grok 4.5 is "Opus-class," a phrase that lacks a published definition and functions more as positioning than measurement. The vendor's own benchmark chart tells a more nuanced story. On five engineering benchmarks shown at launch, Anthropic's Fable model outperformed Grok 4.5 on four tests: DeepSWE 1.0, DeepSWE 1.1, Terminal-Bench 2.1, and SWE-Bench Pro. Grok 4.5 led on only one benchmark, SWE Marathon.

On token efficiency, xAI reports that Grok 4.5 uses approximately 15,954 output tokens per SWE-Bench Pro task, roughly 4.2 times fewer than xAI's Opus 4.8 comparison. Independent testing by Artificial Analysis found Grok 4.5 in the Grok Build harness used fewer tokens than several competing coding-agent systems, supporting an efficiency case within the specific test harnesses that were measured. However, this evidence does not automatically translate to production cost savings, because failed runs, retries, and human repair add unpredictable overhead.

A reader who takes away only "Grok wins" or only "Grok loses" has misread a split result. The benchmarks establish that Grok 4.5 is competitive on engineering tasks, but they do not establish neutral, matched-harness parity across all workloads.

What Should Teams Verify Before Signing a Contract?

Teams evaluating Grok 4.5 should not assume that Grok Build inherits the raw grok-4.5 price and 500,000-token context window without confirming which identifier their invoice uses. Similarly, Cursor users should not assume that bring-your-own-key setups bypass Cursor's infrastructure; the platform routes all requests through its own backend regardless of key ownership.

Buyers who need confirmed European data residency, a pinned model release, or a model-specific service-level agreement should verify these terms before evaluating, as the public documentation does not yet address these requirements. The surface, plan, privacy mode, provider route, region, and contract terms still need to be named in writing before a purchase decision is final.

" }