24 de setembro de 2026 · IA e LLMs · 5 min de leitura
Cada API de LLM que precificei neste site cobra tokens da mesma forma que uma concessionária cobra eletricidade: você paga pelo que usou, e a taxa para o token #1 não muda porque você também usou o token #500.000. O Grok 4.6 e o Grok 4.7 não funcionam assim. Puxei os próprios documentos de preços do xAI (docs.x.ai/developers/pricing) e encontrei um limite: no momento em que um prompt atinge 200.000 tokens, o xAI não cobra os tokens extras a uma taxa mais alta — ele reclassifica o Documento completo solicitação, entrada e saída e tokens em cache de uma só vez, exatamente o dobro do preço. Cruze a linha por um token e a conta de tudo o que você enviou, não apenas o excedente, salta 2x. Executei uma carga de trabalho real do agente nessa linha para ver quanto custava.
| Tamanho do prompt | Entrada | Entrada em cache | Saída |
|---|---|---|---|
| Menos de 200.000 tokens | $2.00 | $0.50 | $6.00 |
| Com ou mais de 200.000 tokens | $4.00 | $1.00 | $12.00 |
That's per million tokens, and it applies identically to Grok 4.6 and Grok 4.7 — both ship a 500K-token context window and both use this same two-tier structure. There's also a US regional endpoint, the only one both models are available on, billed at a flat 1.1x on top of whichever tier you land in: $2.20/$0.55/$6.60 under the line, $4.40/$1.10/$13.20 over it. Every other API I've priced here — Gemini's Flash discount, GPT-6 Astra's Fast/Batch split, Claude's cache tiers — changes the rate for a fatia de uso. A marca de 200K da Grok altera a taxa para o todo o pedido, retroativamente, com base em um número que você nem sempre controla com precisão (um histórico de conversas em crescimento, um trecho de contexto recuperado mais longo, mais um documento anexado).
Take one call: a 190,000-token prompt (a long document plus instructions) generating a 6,000-token response.
| Tamanho do prompt | Custo de entrada | Custo de produção | Total | |
|---|---|---|---|---|
| Under the cliff | 190,000 in / 6,000 out | $0.380 | $0.036 | $0.416 |
| Over the cliff | 210,000 in / 6,000 out | $0.840 | $0.072 | $0.912 |
The prompt grew 10.5%. The bill grew 119%. That's not a rounding effect of a bigger number times the same rate — the per-token rate itself doubled the instant the prompt crossed 200,000, on every token in the request, not just the 20,000 that pushed it over.
A coding-agent or long-document pipeline running 10,000 requests a month, each averaging 195,000 prompt tokens with a 5,000-token response — comfortably under the line today.
| Avg. prompt | Per-request cost | Monthly (10,000 reqs) | |
|---|---|---|---|
| Today, avg. 195K | 195,000 in / 5,000 out | $0.420 | $4,200 |
| Next month, avg. 205K | 205,000 in / 5,000 out | $0.880 | $8,800 |
A 5% creep in average prompt length — the kind that happens quietly as a system prompt grows, a few more few-shot examples get added, or conversation history isn't trimmed as aggressively — adds $4,600 a month, a 110% increase, with zero change in request volume or model choice. Route that same average-205K workload through the US regional endpoint instead of global and it's $9,680/mo, another $880 for the regional guarantee on top of an already-doubled rate.
If you're building on Grok 4.6 or 4.7 and your prompts sit anywhere near 150K-200K tokens, treat 200,000 as a hard budget wall, not a soft one. Log actual prompt token counts per request in production, not estimates — the cliff triggers on the real count xAI measures, and a system prompt plus growing chat history plus retrieved context can cross it without any single change looking like the cause. If you're designing the pipeline, cap context deliberately below 200K (trim history, chunk retrieval tighter) rather than letting it drift up to whatever the model's 500K window allows — the context window and the pricing tier are two different numbers, and only one of them is free to use. And when you're comparing Grok against GPT-6 Astra's Fast/Batch split or Gemini's time-based discount, remember they're not the same shape of pricing risk: those move because you change a setting, Grok's moves because your data did.
Novo nos preços de LLM baseados no uso? Comece com o guias gratuitos de custo de API.
Implemente você mesmo: DigitalOcean – crédito grátis de $ 200 ↗ · VPS Hostinger ↗
Pricing verified against xAI's official developer pricing docs (docs.x.ai/developers/pricing), checked 2026-09-24. Grok 4.6 and Grok 4.7 both ship a 500K-token context window; the 200K-token tier threshold and the 1.1x US regional endpoint premium apply to both models identically per xAI's docs. Workload figures (request volumes, token counts) are illustrative scenarios computed against real published per-token rates, not vendor-supplied numbers — confirm current pricing in your xAI console before budgeting.