Grok 3 mini API pricing vs Grok 3 full

When you move Grok from a demo to a line item, the Grok 3 mini vs Grok 3 pricing spread forces a real decision. xAI lists Grok 3 at $3 per million input tokens and $15 per million output, while Grok 3 mini sits at $0.30 and $0.50—a 10x gap that dwarfs most model-tier differences. But the cheaper token rate means nothing if the mini model drops accuracy on your task or blows up your latency budget in a different way.

Capabilities: What you actually get

Grok 3 full is xAI’s flagship dense model, tuned for multi-step reasoning, tool-augmented agents, and long-context synthesis. Grok 3 mini is a separately trained smaller variant that uses distillation from the teacher; it trades parameter count for throughput.

In our load tests, mini parses tightly specified JSON schemas with slightly lower reliability than full, which is near flawless on the same prompts. On multi-hop retrieval questions, full leads noticeably. For classify, extract, or short chat replies, mini is competent. For planning agentic loops or generating code from ambiguous specs, full earns its keep.

Price and cost model

xAI bills per token with no flat fees. Published rates:

• Grok 3: $3.00 / M input, $15.00 / M output
• Grok 3 mini: $0.30 / M input, $0.50 / M output

That defines the Grok 3 mini vs Grok 3 pricing conversation. Both support prompt caching via cache_control on system blocks; repeated prefixes get steep discounts. For a long static system prompt, cached input on full can approach mini’s headline rate.

resp = client.chat.completions.create(
    model="grok-3",
    messages=[
        {"role": "system", "content": "You are a strict validator.",
         "cache_control": {"type": "ephemeral"}},
        {"role": "user", "content": payload}
    ],
    max_tokens=100,
)

Gateway access forwards these hints, so cached cost math holds.

Latency and throughput

In streaming traces from a US-east client, mini’s time-to-first-token is typically 2–3x lower than full, and its decode throughput is markedly higher. Under concurrency, full degrades more sharply; mini sustains higher queries per second on the same key tier.

If you disable streaming and wait for full completion, mini still wins on wall clock for short outputs. For long generations, full’s quality may reduce retry cascades, narrowing the effective gap.

Ergonomics and API surface

Both speak OpenAI-compatible chat completions. Differences:

• Full exposes reasoning_effort to trade latency for chain-of-thought depth.
• Mini accepts the param but ignores it.
• JSON mode is stable on both; mini occasionally needs stricter validation.
• Function calling: full binds arguments more reliably; mini drops optional params more often.

tools = [{"type": "function", "function": {
    "name": "get_weather",
    "parameters": {"type": "object", "properties": {"city": {"type": "string"}},
                   "required": ["city"]}}}]
client.chat.completions.create(
    model="grok-3-mini",
    messages=[{"role": "user", "content": "Weather in Austin?"}],
    tools=tools,
    tool_choice="auto",
)

Error shapes match OpenAI: 429 on rate limit, 400 on schema violation.

Ecosystem: xAI direct vs gateway access

Direct xAI gives lowest path latency and native cache hints. You manage keys, fallback, and multi-region yourself. To use another vendor’s model, you sign another contract.

A gateway like n4n.ai exposes one OpenAI-compatible endpoint addressing 240+ models, including both Grok tiers, with automatic fallback when xAI is rate-limited or degraded. It honors client routing directives and forwards cache-control, so the Grok 3 mini vs Grok 3 pricing advantage stays transparent. Added median latency is a trivial TLS hop.

client = OpenAI(base_url="https://api.n4n.ai/v1", api_key=GW_KEY)
# model="grok-3" or "grok-3-mini"

Limits and quotas

xAI per-key defaults on free tier are tighter for full: roughly 60 RPM and 10k TPM versus 300 RPM and 50k TPM for mini. Paid tiers raise these substantially. Both models share a 128k context window.

Gateways pass provider 429s transparently and may add a global pool. Monitor x-ratelimit-remaining headers.

Head-to-head comparison

Which to choose

Production RAG on legal/finance docs: Use Grok 3 full. Accuracy on clause extraction justifies the premium; cache the boilerplate.

Real-time moderation or routing: Grok 3 mini. Higher RPM and faster decode handle firehose volume; heuristic fallback covers edge errors.

Agentic coding assistant: Full. Mini’s tool binding drops critical args.

Bulk data enrichment (millions of rows): Mini. At $0.30/M input, cost dominates; parallelize workers.

Multi-model product with failover: Gateway plus both. Default to mini, escalate to full when complexity crosses a threshold.

The Grok 3 mini vs Grok 3 pricing gap is the headline, but the real decision is error cost. If a wrong answer triggers a retry or human review, the cheap model isn’t cheap.