Grok 4.7, 4.6, and 4.5: The Top Tier

Lesson rewritten on November 3, 2026 based on the Models page, the grok-4.7 spec sheet, the Reasoning guide, and xAI best practices: Grok 4.7 (September 2026) is the model xAI recommends for all use cases, ahead of Grok 4.6 (August) and Grok 4.5 (July), previously presented as the flagship model; max_output_tokens caps only the visible response, not reasoning, and the cache reduction is 75% on grok-4.7, not 90%. Sources: Models · Reasoning · Best practices.

The Top of the Grok Lineup

Three models make up xAI’s high end, all with 500,000 tokens of context, text and images as input, and text as output. Grok 4.5, announced on July 16, 2026, kicked off the series; Grok 4.6 followed on August 12, then Grok 4.7 on September 21. xAI presents Grok 4.7 as its state-of-the-art model for code, agentic tasks, and knowledge work, and recommends it for anything that isn’t image, video, or voice: it’s the starting point for any new project. The other two remain in the catalog at the same base rates.

Whether you’re building a complex agent, analyzing large documents, or solving math problems, turn to grok-4.7 when response quality takes priority over cost.

Adjustable Reasoning

All three models reason continuously—reasoning cannot be disabled—but you control the effort: the reasoning.effort parameter in the Responses API (reasoning_effort in Chat Completions) ranges from low to high, the default value, and goes up to xhigh on Grok 4.6 and 4.7. On Grok 4.5, according to the Reasoning guide, a request for xhigh is treated as high.

The benefit is as much architectural as it is economic: a single integration covers simple tasks and hard tasks, and the cost/quality trade-off happens per request rather than by routing between two models. On a classification task, low effort responds quickly and consumes little; on a math proof, high or xhigh mobilize the full depth of the model. xAI advises setting the effort explicitly for every request and logging it with every result.

The previous generation keeps its two separate variants—grok-4.20-0309-reasoning and grok-4.20-0309-non-reasoning—still available in the catalog.

Encrypted Reasoning in Practice

Grok’s reasoning is said to be “encrypted”: xAI keeps internal reflections confidential, but you can retrieve an encrypted version via the Responses API and inject it into subsequent requests so the model picks up where it left off. With Grok 4.7, this encrypted content is returned with every response; on other models, request it via the include parameter:

response = client.responses.create(
    model="grok-4.7",
    input="Prove that the square root of 2 is irrational",
    reasoning={"effort": "high"},
    include=["reasoning.encrypted_content"],
)

Reasoning Tokens

Tokens consumed by reasoning appear separately in the response usage:

{
  "usage": {
    "input_tokens": 32,
    "output_tokens": 9,
    "output_tokens_details": {
      "reasoning_tokens": 110
    },
    "total_tokens": 151
  }
}

In this example from the API reference, the total adds up input, visible output, and reasoning. Reasoning tokens are part of the billed tokens: on a complex request, they can represent a significant portion of the total cost, and this is precisely what the effort setting lets you control.

max_output_tokens caps only the visible output: reasoning and function calls don’t count toward it. Without a value, it defaults to 128,000 tokens; a limit that’s too tight cuts off the response, which the Responses API signals with an incomplete status.

Recommendation: prefer streaming, and allow at least 900 seconds per request, as xAI best practices advise—deep reasoning on a hard problem can take several minutes, and examples in the Reasoning guide go up to 3,600 seconds.

Tiered Pricing

The high end follows the rule common to the lineup: the base rate applies as long as the prompt stays under 200,000 tokens, and doubles beyond that, for all tokens in the request. For all three models, $2.00 input and $6.00 output per million become $4.00 and $12.00 once the prompt crosses this threshold.

Cache significantly reduces the bill: input tokens already seen are billed at $0.50 per million instead of $2 on Grok 4.7 and 4.6 (75% reduction), and $0.30 on Grok 4.5 (85%). On an assistant that returns the same system prompt and reference context every turn, the difference is considerable—hence the importance of keeping a stable prompt prefix. Requests sent via the Batch API also get a 20% discount.

Grok 4.20 Multi-agent

xAI also offers grok-4.20-multi-agent-0309, a variant dedicated to orchestrating multiple agents, with one million tokens of context and 4.20 generation pricing. On this model, reasoning.effort doesn’t control reasoning depth but the number of agents mobilized: 4 or 16.

Multimodal Input

All three models accept text and images as input. Images must be in JPG or PNG format, with a maximum size of 20 MiB, and there is no limit on the number of images.

Output remains text-only—these models do not generate images. For visual generation, you will use the dedicated Grok Imagine models, presented later in this course.

Pricing verified on November 3, 2026—prices change regularly: before any budget calculation, check the current grid on the official xAI models and pricing page.

Key Takeaways

• Grok 4.7 (September 2026) is the model xAI recommends; Grok 4.6 and 4.5 remain in the catalog, with the same 500,000-token context and same base rates
• Reasoning cannot be disabled: reasoning.effort ranges from low to high (default), and up to xhigh on 4.6 and 4.7
• Encrypted reasoning is retrieved and reinjected; Grok 4.7 returns it with every response
• max_output_tokens counts only the visible response; allow at least 900 seconds for requests
• Tiered pricing: $2.00 / $6.00 under 200,000 prompt tokens, $4.00 / $12.00 beyond; 75% cache reduction on 4.7
• grok-4.20-multi-agent-0309 remains the variant dedicated to orchestration, where effort controls the number of agents