Grok 4.7: What xAI's New Model Changes (Pricing, Benchmarks, API)

Grok 4.7 at a glance

xAI released Grok 4.7 on September 21, 2026. The company (whose site and developer docs now carry the SpaceXAI name) positions it as its most capable model for coding and knowledge work, built on a larger base model than Grok 4.6 with extended reinforcement learning on hard tasks. The official Grok 4.7 announcement is the source for the launch claims and benchmarks below; API numbers come from xAI's developer docs.

The short version for anyone who builds on the API: it costs exactly the same as Grok 4.6, has the same 500K context window and the same rate limits, and scores higher on xAI's coding benchmarks. For most teams that makes it a model-ID swap, not a migration project.

What's new in Grok 4.7 compared with Grok 4.6

xAI's launch post names four changes. Read them as the vendor's claims; independent evaluations usually take a few weeks to appear.

• Bigger base model. Grok 4.7 is not a fine-tune of 4.6; it starts from a larger pretrained model.
• More RL on difficult tasks. Extended reinforcement learning aimed at long, multi-step coding and knowledge-work problems.
• Self-verification and long-context management. The model is trained to check its own work and keep track of state across long sessions, which is exactly where agent loops tend to fall apart.
• Native Grok Bot harness understanding. Better conversational behavior inside xAI's own agent harness.

What did not change matters just as much for production planning:

Same price, same limits, same parameters, better score. There is no cost argument for staying on 4.6; the only reason to hold back is if your own evals show a regression on your specific prompts.

Grok 4.7 benchmarks: what the numbers say

All figures below are published by xAI in the launch post. None has been independently replicated at the time of writing.

How to read this: the 5.9-point jump on CursorBench is the one number with a like-for-like 4.6 baseline, and it is roughly a 15% relative improvement on the task type most API customers actually run. The DeepSWE score is reported at high effort, which is also the API default, so you get that configuration without touching any parameter. Treat the rest as directional until third-party leaderboards catch up, and run your own eval set before trusting any vendor chart with a production decision.

Grok 4.7 API pricing, including the 200K-token cliff

The headline price is $2 input / $6 output per million tokens, but the xAI API pricing page has a second tier most launch coverage skips: once a prompt reaches 200K tokens, every rate doubles.

Server-side tools are billed per call on top of tokens: web search, X search and code execution each cost $5 per 1,000 calls. The US regional endpoint (us.api.x.ai) runs at 1.1× the global token rates, which matters if you need data residency.

Worked example: one agent step

Take a typical coding-agent turn: a 150K-token prompt (system prompt, tool schemas, repo context) and 8K output tokens.

Two takeaways. First, cache hits halve the bill here, so keep your stable prefix (system prompt, tool definitions, pinned files) byte-identical at the front of every request. Second, the 500K window is real but expensive: sending the same 250K tokens in one call costs about 84% more than splitting it into two calls under the line. If your agent can retrieve or summarise instead of stuffing the whole repo in, stay under the line.

Grok 4.7 Fast: same model, twice the speed, twice the price

xAI also ships Grok 4.7 Fast. According to the pricing docs it is the same Grok 4.7 model served on faster infrastructure, at twice the standard token rates ($4 input / $1 cached / $12 output per million below 200K). The catch: it is only available inside Cursor and Grok Build, not on the public API. If you are building your own product, standard grok-4.7 is the only option, and you tune latency with the reasoning-effort setting rather than a faster tier.

How to call Grok 4.7 from the API

xAI's REST API follows the OpenAI specification, so the official OpenAI Python SDK works by pointing it at xAI's base URL. Switching an existing Grok 4.6 integration is a one-line change to the model ID.

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["XAI_API_KEY"],
    base_url="https://api.x.ai/v1",
)

resp = client.chat.completions.create(
    model="grok-4.7",
    messages=[
        {"role": "system", "content": "You are a senior Django reviewer."},
        {"role": "user", "content": "Find the N+1 query in this view: ..."},
    ],
)

print(resp.choices[0].message.content)

# Reasoning tokens are billed as output tokens; track them per call
usage = resp.usage
print("prompt:", usage.prompt_tokens,
      "completion:", usage.completion_tokens,
      "reasoning:", usage.completion_tokens_details.reasoning_tokens)

Three production notes the quickstart does not spell out:

• Log reasoning tokens. At the default high effort, hidden reasoning can easily outweigh the visible answer, and it bills at the output rate. The reasoning_tokens field is how you find out which endpoints are quietly expensive.
• Drop effort for cheap tasks. Classification, extraction and routing rarely need high. Test low or medium on those paths and keep high/xhigh for code generation and multi-step planning.
• No Batch API. Grok 4.7 does not support batch processing, so bulk offline jobs (backfills, large-scale tagging) run at full synchronous prices. Budget for that, or route those jobs to a model that offers batch discounts.

If you are wiring Grok 4.7 into agents, document pipelines or internal tools and want the cost controls built in from day one, that is the kind of work Netalith does under AI automation and workflow development.

Should you switch to Grok 4.7?

Grok 4.7 is an incremental release, not a new generation, and that is its strength for builders: a measurable coding gain with zero pricing or integration churn. The real work is the same as with any model upgrade: keep an eval set, watch reasoning-token spend, and keep prompts under the 200K cliff unless the task truly needs more.

Not sure whether your product or content is ready for AI assistants like Grok to find and cite it? Book a free AI visibility consultation with Netalith. No cost, no account needed.