Grok 4.7 is a reasoning model for coding, research, and work that takes several steps to complete. Released on September 21, 2026, it is available through the xAI API as grok-4.7, alongside coding products such as Cursor and Grok Build. The official announcement emphasizes more persistent problem solving and stronger self-checking.
For developers, the useful questions are practical: what does a request cost, how do you call the model, and when is the extra reasoning worth paying for? This guide covers those decisions, with a small API example and a workflow for adding external evidence.
What changed in Grok 4.7?
The release uses a larger base model and additional training on difficult, extended tasks. That makes repository changes and document-based analysis sensible evaluation targets. It does not mean every answer will improve, or that every task should become a long-running agent session.
Artificial Analysis reports an Intelligence Index score of 46 at xhigh reasoning effort. Its coding evaluation also improves with Grok Build. However, the same report finds high output-token consumption: approximately 81,000 tokens per Intelligence Index task, compared with 36,000 for Grok 4.6 at high effort. Those are different effort settings, not a controlled prediction for your application.
Treat published benchmarks as a shortlist filter. Your acceptance test should measure whether the model finishes your task correctly, how much review it needs, and what the complete run costs.
Grok 4.7 API pricing and context limits
The model specification lists a 500,000-token context window, text and image inputs, and text output. It supports function calling, structured outputs, and four reasoning levels: low, medium, high, and xhigh. High is the default.
The table below uses the global xAI API's published USD rates, checked September 23, 2026. Prices are per million tokens. The API pricing page applies long-context rates to the entire request once the prompt reaches 200,000 tokens.
| Billing category | Prompt below 200K tokens | Prompt at least 200K tokens |
|---|---|---|
| Uncached input | $2.00 / million | $4.00 / million |
| Cached input | $0.50 / million | $1.00 / million |
| Output | $6.00 / million | $12.00 / million |
For an illustrative short-context request with 10,000 uncached input tokens and 2,000 billed output tokens, the token cost is $0.032: $0.02 for input plus $0.012 for output. This is arithmetic, not a measured run. Extra reasoning, retries, and separately priced tools can increase the bill.
Do not apply one provider's thresholds to another product. Cursor's pricing documentation uses an input threshold above 256,000 tokens for its long-context tier. Check the service that actually bills you.
How to call Grok 4.7 through AIsa
AIsa supports Grok 4.7 through its OpenAI-compatible API. Create an API key in the AIsa console and set AISA_API_KEY in your environment. Following the AIsa integration guide, send a chat completion request with grok-4.7 as the model:
curl --fail-with-body https://api.aisa.one/v1/chat/completions \
-H "Authorization: Bearer ${AISA_API_KEY}" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-4.7",
"messages": [
{
"role": "user",
"content": "Review this JavaScript function for bugs: function median(a) { a.sort(); return a[a.length / 2]; } Explain the problems, propose a fix, and suggest edge-case tests."
}
]
}'
The endpoint, authentication, and message format follow AIsa’s documentation; this example has not been executed as a paid inference request. It asks for a review rather than granting access to a repository or running generated code.
Before expanding it into an agent, decide what counts as success. For this example, check numeric sorting, even-length arrays, empty inputs, and whether the implementation mutates the original array. Run your own tests rather than accepting the model's assurance that its fix works.
Using Grok 4.7 with external data and AIsa
A model cannot evaluate evidence it has never received. For a competitor brief, retrieve current pages or public social posts first, preserve their URLs and timestamps, and then ask the model to separate sourced observations from its own conclusions.
AIsa provides model routing and data APIs through one API key. Its custom-agent guide is a useful starting point for adding capabilities to an existing application.
A practical design is to use AIsa's API catalog for retrieval, normalize the responses in your application, and send only relevant excerpts to Grok 4.7. Use the same AIsa API key for the data requests and the Grok model call above, so the workflow does not require a separate xAI account or key.
Check the AIsa model catalog for current model pricing before deploying the workflow. The earlier pricing table is an xAI reference, not an AIsa billing quote. Start with one small, traceable research task before adding scheduled runs.
Watch a Grok 4.7 review before choosing a workflow
Pat Simmons compares Grok 4.7 with other models in a hands-on review. Use it to identify tasks worth reproducing yourself, rather than treating one creator's results as a universal ranking.
Video by Pat Simmons. The preview above opens the embedded player; the title links to the original YouTube video.
Five practical tips for evaluating Grok 4.7
- Define the deliverable. Request a patch, a cited brief, or a structured extraction with clear acceptance criteria. “Do some research” leaves both quality and cost hard to judge.
- Test a small representative set. Include routine tasks, ambiguous requests, and known failure cases. Reuse the same inputs when comparing models so changes are easier to explain.
- Track completed-task cost. Record token charges, tool calls, retries, and review time. A lower token price is useful only when the workflow delivers acceptable results.
- Provide relevant context. Start with the files or evidence needed for the decision. A large context window is capacity, not an instruction to send every document you have.
- Separate suggestions from execution. With function calling, validate requested arguments in your application. Require review for consequential actions and give the agent a clear stopping condition.
Five terms in the Grok API documentation
- Context window: The token capacity available for the request and its conversation context. It is not a guarantee that every detail will be used correctly.
- Reasoning effort: A setting controlling how much work the model devotes to an answer. Compare its effect on quality, latency, and spending together.
- Cached input: Previously processed prompt content that qualifies for discounted input billing. Do not budget for cache hits until your usage records show them.
- Function calling: A way for the model to request a named tool with arguments. The application remains responsible for validating and executing that request.
- Agent harness: The surrounding software that supplies tools, manages state, and handles execution. Different harnesses can produce different results with the same model.
Grok 4.7 FAQ
Is Grok 4.7 available now?
Yes. It launched on September 21, 2026, and is also available through AIsa. Use your AIsa API key with the chat completions endpoint and grok-4.7 as the model, as shown above.
How much does the Grok 4.7 API cost?
The global xAI API starts at $2 per million uncached input tokens and $6 per million output tokens. Long-context requests and additional tools can cost more. Use the billing table above for the standard token tiers.
Can I use Grok 4.7 Fast through the public API?
No. The developer guide lists Fast as a faster serving option in Cursor and Grok Build, rather than a public xAI API model. Confirm the selected speed tier before comparing costs.
Does a 500K context window mean Grok remembers everything?
No. It describes capacity, not perfect recall or permanent memory. Provide relevant evidence, verify the answer against it, and explicitly manage any information your application needs to retain.
Is Grok 4.7 better than Grok 4.6 for coding?
Published coding evaluations show improvements, but the outcome depends on the task, harness, and reasoning setting. Test both against your own acceptance criteria and compare total cost per accepted result.
For your first experiment, choose one task you already know how to verify. Add external evidence only where it improves the answer, then use the AIsa API catalog and model catalog to plan the next integration step.