Grok 4.5 deep dive: xAI’s flagship for coding and STEM

What it is

Grok 4.5 is the hosted model xAI (catalog copy says SpaceXAI) currently labels as its smartest, aimed at coding, knowledge work, and STEM. It offers a 500K-token context window, accepts text, image, and file inputs with text output, and exposes agent-friendly parameters including reasoning, tools, and structured_outputs. Versus sibling Grok 4.20 (2M context, lower unit price), 4.5 trades a shorter window and higher price for the flagship SKU—decide first whether you want maximum context or the newest flagship.

Spec sheet at a glance

By the numbers: a 500K-token context window; input modalities text, image, and file; output text only. It exposes 15 API parameters—frequency_penalty, presence_penalty, temperature, top_p, seed, stop, max_tokens, logprobs/top_logprobs, reasoning and include_reasoning, tools and tool_choice, structured_outputs, and response_format—so it drops into agent and function-calling stacks cleanly. Knowledge cutoff is not disclosed in this snapshot, and the OpenRouter primary listing does not record a batch or cached-input tier. Vendor: xAI.

Pricing and where to run it

Snapshot pricing is $2 input and $6 output per 1M tokens from the OpenRouter primary listing (is_primary_listing=1). Because output costs 3× input, long generations and agent loops inflate the bill—estimate with your real input/output ratio. This snapshot only verifies the OpenRouter path; check the Token page for any additional hosts.

How it compares on price and specs

Grok 4.5’s blended cost is about $4 per 1M tokens (the $2/$6 mean). Within the family: Grok 4.20 / 4.20 Multi-Agent ~$1.88 (2.0M ctx), Grok 4.3 ~$1.88 (1.0M). Across vendors: GPT-5.4 ~$8.75 (1.05M), Gemini 3.1 Pro Preview ~$7 (~1.05M), Claude Fable 5 ~$30 (1.0M), Llama 4 Scout just $0.20 (10M). 4.5 is outside AI Hippo’s composite Top 20 in this snapshot—that score is a context-per-dollar proxy, not a capability benchmark—so the longer, cheaper 4.20 Multi-Agent lands at #2. Judge 4.5 on flagship fit, not on that proxy rank.

Cost in practice: two scenarios

Model the real bill with your input/output split. An agentic coding turn of 100K input + 20K output costs about $0.32 on Grok 4.5 ($0.20 input + $0.12 output), about $0.18 on Grok 4.20, about $0.55 on GPT-5.4, and about $2.00 on Fable 5. A window-filling long-document pass of 500K input + 5K output runs about $1.03 on 4.5, about $0.64 on 4.20, and about $1.33 on GPT-5.4. The takeaway: 4.5 costs roughly 1.8–2× sibling 4.20, but far less than Fable 5; if 500K still is not enough, look at the 2M 4.20 line. These are spec and cost comparisons, not capability benchmarks.

When to choose it—and when not to

Pick Grok 4.5 when you want xAI’s current flagship SKU, need multimodal inputs plus reasoning/tools, and 500K context is enough. If you prioritize stuffing a whole repo or a huge document in one call—or hard cost control—sibling Grok 4.20 (2M, $1.25/$2.50) is usually the better buy. On a tight budget with openness to open/cheap tiers, look at Llama 4 Scout or DeepSeek V4 Flash. Decide with AI Hippo’s model, Token, and compare pages—not the composite rank alone.

Sources

• Grok 4.5 on AI Hippo
• OpenRouter models API
• xAI