How Much Does Grok 4.5 API Cost?

Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. Those are the sticker numbers from xAI’s July 8 announcement, and they undercut every model in its claimed capability class.

But sticker price is the least interesting part of this launch. Grok 4.5’s economics rest on a different number: output tokens per solved task, where xAI claims a 4.2x advantage over Claude Opus 4.8. This breakdown covers the full price sheet, the parts that aren’t confirmed yet, the odd trade-off against xAI’s own cheaper model, and what real workloads cost.

The full price sheet

From the xAI docs pricing page, as of July 9, 2026:

Note the last row. xAI offers 20% batch discounts on grok-4.3 and the grok-4.20 family, but no batch rate is listed for grok-4.5. If your workload is overnight bulk processing, that changes the comparison against xAI’s own catalog.

Tool-call pricing matters for agentic use: a research agent making heavy use of web and X search pays $5 per thousand invocations on top of token costs.

The cached-input question

Several launch-day posts cited a $0.50 per million cached-input rate, and OpenRouter’s listing surfaced similar figures. As of July 9, neither the release page nor xAI’s docs pricing page confirms a cached-input discount for grok-4.5.

Until the console or docs state it plainly, don’t build a cache-heavy cost model on that number. Prompt caching routinely cuts agent costs by half or more when it’s real, so this is worth re-checking weekly; we’ll update when xAI documents it.

The oddity: Grok 4.5 costs more than Grok 4.3

xAI’s own catalog makes an unusual shape. The newer model is 60% more expensive on input, 140% more on output, and holds half the context:

So “upgrade to the new one” isn’t automatic. Long-context retrieval, document analysis, and bulk batch jobs belong on grok-4.3. Grok 4.5’s case is capability per token on agentic and coding work. Our API guide includes a model-picker table for this exact decision.

Effective cost: why token efficiency beats sticker price

Here’s the number that changes the math. On SWE Bench Pro, xAI reports:

• Grok 4.5: 15,954 output tokens per resolved task, on average
• Opus 4.8 (max): 67,020 output tokens per resolved task

Price that out per task:

A 4.2x sticker gap on output price becomes roughly 17x per completed task, because you’re multiplying a cheaper rate by fewer tokens. That’s the whole pitch: “the highest intelligence per unit of time and cost,” in xAI’s words.

Handle this with appropriate skepticism. The token counts are vendor-reported, from one benchmark, and Opus’s “max” effort mode is deliberately verbose because extended reasoning is how it earns its higher scores on the harder evals; it wins SWE Bench Pro 69.2% to 64.7%. You’re not comparing identical work. But even discounting heavily, terse-and-cheap compounds: shorter outputs cost less, stream faster at Grok’s 80 tokens per second, and shorten every loop in an agent session.

The head-to-head capability picture is in Grok 4.5 vs Claude Opus 4.8.

What real workloads cost

Rough scenarios at list prices, assuming typical prompt/response shapes:

A support chatbot (500 input + 300 output tokens per exchange, 100k exchanges/month): 50M input + 30M output = $100 + $180 = $280/month.

An agentic coding pipeline (200 tasks/day at SWE-Bench-like shapes, ~30k input + ~16k output per task): about 180M input + 96M output per month = $360 + $576 ≈ $936/month. The same volume at Opus 4.8 output verbosity and rates lands well into five figures.

A document-heavy RAG service stuffing 100k-token contexts: input dominates at $2/M, and this is where grok-4.3 at $1.25/M with a 1M window, or cheaper open-weight rivals, deserves the comparison instead.

How Grok 4.5 pricing compares to the market

Against Opus 4.8’s pricing, Grok 4.5 costs 40% as much on input and 24% on output while splitting xAI’s published benchmarks 2-2. Against Fable 5, the model that tops all four of those charts, it’s an 80%+ discount for a real capability gap. That’s the honest framing: you’re buying near-Opus capability at a deep discount, not the frontier.

Worth remembering: launch prices are marketing. xAI wants developers to switch, and week-one pricing reflects that. Lock-in decisions deserve a margin of safety for future increases.

Keeping your Grok 4.5 bill under control

• Measure before migrating. In Apidog, run your real prompts against grok-4.5 and your current model, and assert on the usage object in responses. If Grok’s outputs on your prompts aren’t shorter, the efficiency argument doesn’t transfer to your workload.
• Watch tool calls. $5 per 1,000 search invocations is invisible in token dashboards. Log them separately.
• Mock during development. Every UI iteration against a live endpoint is money. Generate a mock server from the response schema in Apidog and point your dev environment there; spend tokens only on prompt work. Download Apidog free to set both up in an afternoon.
• Use the free windows while they last. Grok Build and Cursor include free Grok 4.5 usage right now; here’s every current path.

For the model itself, benchmarks and all, start at what is Grok 4.5.

FAQ

How much does the Grok 4.5 API cost? $2 per million input tokens, $6 per million output tokens, with web/X search tool calls at $5 per 1,000 and priority processing at 2x rates.

Does Grok 4.5 have cached input pricing? Unconfirmed. Secondary sources report $0.50/M, but xAI’s own pricing page doesn’t list a cached rate for grok-4.5 as of July 9, 2026. Check the console.

Is Grok 4.5 cheaper than Claude Opus 4.8? Yes, at both the sticker level ($2/$6 vs $5/$25) and more so per completed task once token efficiency is factored in.

Why is Grok 4.5 more expensive than Grok 4.3? xAI prices it as the premium capability model. You pay more per token for higher benchmark scores and 2x token efficiency, but you give up half the context window and the batch discount.