Grok 4.7 is xAI’s latest announced model for coding and knowledge work. The September 21, 2026 launch announcement, published under the SpaceXAI name, emphasizes longer task execution, stronger self-checking, better context management, and improved document and presentation work.
For developers, the decision is practical: does it complete a difficult task with fewer corrections, at an acceptable total cost? This guide covers pricing, the Fast variant, free access through Grok Build, and a small evaluation plan you can use on your own work.
Verified September 26, 2026 using official sources. We have not run Grok 4.7, tested Grok Build, or reproduced the published benchmarks.
What changes with Grok 4.7?
The company says it trained a larger base model with longer reinforcement learning on harder tasks, including problems taking many hours. It also reports a new safeguard stack with stronger jailbreak resistance. These are vendor-reported improvements, not a guarantee that an agent will finish every long task correctly.
The developer guide lists a 500,000-token context window, text and image input, and text output. The public API model identifier is grok-4.7.
A large context window gives an application room to supply more material. It does not establish that the model will notice every contradiction or preserve every instruction. The useful test is whether it can locate the relevant facts and use them correctly while completing a task.
Grok 4.7 pricing: the starting rate is only one tier
The official pricing page lists these standard global API rates in US dollars per million tokens:
| Prompt size | Input | Cached input | Output |
|---|---|---|---|
| Below 200,000 tokens | $2.00 | $0.50 | $6.00 |
| 200,000 tokens or more | $4.00 | $1.00 | $12.00 |
The long-context threshold applies to all tokens in that request, not only the portion beyond the threshold. Server-side tools can add separate charges. Check current account pricing before planning a large agent run.
For a simple illustration, suppose one short-context request uses 100,000 uncached input tokens and 10,000 billed output tokens. The calculation is $0.20 + $0.06 = $0.26, before tool charges or other applicable additions. This is arithmetic using the published rates, not a measured task cost.
An agent may send many requests and repeatedly include conversation history. Estimate the whole session, including retries and failed approaches. The length of the final answer alone does not describe how much work the system performed.
Grok 4.7 Fast: access and pricing caveats
According to the developer guide, Fast uses the same model on faster infrastructure. It is available through Cursor and Grok Build, but not through the public xAI API, and it is excluded from Grok Build’s free tier. The launch post claims twice the output speed.
For short prompts, the pricing page lists $4 input, $1 cached input, and $12 output per million tokens. Its long-context Fast table conflicts with its general “twice the standard rates” description. Confirm that tier with the provider before relying on a long-context Fast estimate.
Choose speed based on the part of the workflow that actually delays you. Faster token output may help an interactive coding session, but it does not automatically halve time spent running tests, waiting for tools, or reviewing a patch.
How to read the Grok 4.7 benchmark claims
The following are selected company-published results from the announcement, not our own measurements:
| Benchmark | Reported Grok 4.7 result |
|---|---|
| CursorBench 4.0 | 46.3% |
| DeepSWE v1.1 | 71.0% — marked high effort |
| EEBench | 64.0% |
| AA Briefcase v1.1 | 1,657 |
| Terminal-Bench 4.0 | 37.6% |
| HealthBench Professional | 56.7% |
The launch comparison generally labels Grok 4.7 as xHigh, with the DeepSWE exception noted above. These benchmarks use different tasks and scoring scales; the AA Briefcase number is not a percentage.
Use benchmark results to choose what to investigate. For a repository task, ask whether the agent produces a correct patch and meaningful tests. For office work, inspect whether the result faithfully represents the source material. A strong score in one domain does not settle performance in another.
Can you try Grok 4.7 for free?
The Grok Build page advertises a free trial of its coding agent powered by Grok 4.7. This is separate from public API billing. We did not establish a numeric free quota or guaranteed trial duration from the reviewed pages; check the allowance shown for your account.
The Build setup guide explains installation, launching grok inside a project, and browser authentication on first launch. It also supports API-key authentication. Do not assume an API-key session inherits a free account allowance.
For a first attempt, choose a project you understand and can restore easily. Start by asking the agent to explain a small area of the codebase. That gives you a quick way to check whether it has identified the relevant files before asking it to make changes.
A practical coding test for Grok 4.7
Here is an original evaluation exercise: use a small application with a reproducible filtering bug. Give the agent the observed behavior, expected behavior, and command for running the existing tests. Ask it to diagnose the cause before proposing a fix.
Investigate why searching for a customer by email misses entries when the query contains leading or trailing spaces. Identify the relevant code path, propose a minimal fix, and add a regression test. Preserve the existing handling of empty queries. Explain what you changed, which tests you ran, and any remaining uncertainty.
Choose an actual bug with this behavior or adapt the example to a known issue in your test project. The prompt does not imply that your application has such a defect.
- Diagnosis: does the explanation match the real code path?
- Patch scope: does the change address the bug without unrelated rewrites?
- Regression test: would the test fail on the original implementation?
- Evidence: are test results actually available, or merely claimed?
- Review effort: how much time does a human need to accept the result?
Then add one realistic complication, such as case sensitivity or an existing normalization rule. Observe whether the agent revises its approach consistently. This is more informative than immediately giving it a sprawling task with no clear completion criterion.
Test documents and presentations with a source packet
For knowledge work, prepare a short fictional packet containing meeting notes, a small budget table, and a list of unresolved questions. Ask for a one-page brief and a five-slide outline. Supply the numbers yourself so the result can be checked.
Require the brief to separate confirmed facts from proposals, preserve all totals, and identify missing information. For the slide outline, check whether each slide advances the argument and whether the conclusion follows from the packet. If your chosen tool produces files, inspect the rendered files too; correct text can still have unusable layout.
Score the deliverable before reading the model’s explanation of its work. A confident self-review should not influence whether a missing cost item or invented deadline is considered acceptable.
Measure completion quality alongside time and cost
For each trial, save the model and effort setting, task description, elapsed time, billed usage where available, and required corrections. Define success before the run. For coding, that might mean an accepted patch with a passing regression test. For a document, it might mean no unsupported claims and no arithmetic errors.
Repeat with a few representative tasks rather than relying on one showcase result. Our AI tool testing checklist provides a broader scoring approach. For interface projects, our Vibe Coding design brief guide can help make the request more precise.
Grok 4.7 is a candidate for longer coding and knowledge-work tasks. The useful outcome is evidence that it improves your own workflow, with clear limits on cost and correction effort. Start small, inspect the result, and expand the task only after the first test succeeds.
Official sources and verification
- Launch announcement — reported improvements and benchmark results.
- Developer guide — API identifier, context, and Fast availability.
- API pricing — token tiers, tools, and Fast pricing.
- Grok Build — advertised free trial.
- Build documentation — installation and authentication.
The coding and document exercises are editorial suggestions. No paid requests, account sign-ins, or independent benchmark runs were performed for this article.