Grok 4.7 is SpaceXAI’s model for coding and knowledge work, released on 21 September 2026. It is available through the xAI API, Cursor and Grok Build, with a gradual rollout in GitHub Copilot. Standard direct API pricing starts at $2 per million input tokens and $6 per million output tokens. For UK web and marketing teams, it is worth testing on research, client reporting, content review and development. Independent professional-work evaluations support that trial; better advertising copy, search rankings and commercial results remain unproven. Launch announcement, GitHub rollout.
Based on public documentation checked on 21 September 2026. AIWIZ has not run a hands-on model comparison for this article. Availability findings describe published access routes, not tests of individual account entitlements.
When did Grok 4.7 launch?
SpaceXAI, which also uses xAI branding, announced Grok 4.7 on 21 September 2026. Its API release notes confirm the model identifier grok-4.7. Earlier release-date articles may describe the position before launch; use the dated announcement and current documentation when checking access. API release notes
The company describes a larger base model, longer reinforcement-learning training on harder tasks and better checking of its own work. These are vendor descriptions. The comparison section below separates the published launch results from other evaluations.
Where you can use Grok 4.7
Availability checked: 21 September 2026. A subscription to one Grok product does not establish access through another. The tables distinguish documented listings from routes where model-specific access remains unconfirmed.
Documented access routes
| Route | Status | What to know |
|---|---|---|
| xAI API | Available | Use grok-4.7 with the Responses or Chat Completions API. Model details |
| Cursor | Paid access documented | Grok 4.7 draws from the Cursor Models usage pool. Fast is the default on Pro and above. India’s Start plan uses fixed medium effort at standard speed. SpaceXAI says “all plans”, but the checked Cursor pages do not establish Hobby entitlement. Cursor model page, SpaceXAI guide |
| Grok Build | Available; documented default | Free access is advertised, with Fast excluded. The pages checked did not state a numerical free allowance. Developer guide |
| GitHub Copilot | Rolling out gradually | GitHub lists Grok 4.7 as generally available, with staged rollout to Pro, Pro+, Max, Business and Enterprise. Supported surfaces include VS Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, JetBrains, Xcode and Eclipse. Organisation model policies can affect access. GitHub announcement |
| OpenRouter | Listed | Model identifier x-ai/grok-4.7. Prices and tier conditions are shown below. OpenRouter listing |
| Vercel AI Gateway | Listed | Model identifier spacexai/grok-4.7. Vercel advertises launch pricing ending on 27 September 2026. Vercel listing |
Routes not confirmed in the checked sources
| Route | What the evidence establishes |
|---|---|
| Cloudflare AI Gateway | SpaceXAI names Cloudflare as a gateway. Cloudflare documents xAI routing, but its checked model catalogue did not list Grok 4.7. No request was tested. Cloudflare provider guide, catalogue |
| grok.com, iOS and Android apps; Grok on X | The consumer pages checked did not establish a dated Grok 4.7 rollout or model-specific subscription entitlement. Consumer FAQ, overview |
| Grok Bot in Cursor | Cursor staff say the serving model is not published and can change. Linking SuperGrok or X Premium+ grants Bot usage; it does not upgrade the Cursor plan. Staff discussion, Bot plans |
| Google Cloud partner models | Grok 4.7 was not listed in the checked partner-model catalogue. Google catalogue |
| Amazon Bedrock | The checked model cards list Grok 4.3 and 4.6, without 4.7. AWS catalogue |
| Microsoft Foundry | The checked Grok guide lists 4.6 preview and older models, without 4.7. Foundry guide |
An absent catalogue entry does not prove that no account-specific preview exists. GitHub Copilot, Microsoft Copilot, Copilot Studio and Microsoft Foundry are separate products; a listing in one does not establish access in another.
Seeing a model in a selector also does not establish permission to use every feature. Cursor’s Cloud Agent access depends on the plan. For GitHub organisation accounts, an administrator’s model policy may affect availability, although default model enablement can make new models available automatically. Cursor plans, GitHub rollout policy
Context, inputs and UK data handling
The direct API specification lists a 500,000-token context window, text and image inputs, and text output. Reasoning effort can be low, medium, high or xhigh, with high as the default. Batch API requests are not supported. These are direct API specifications; host products can impose their own limits. A large context window is a capacity measure, not a guarantee that the model will extract every fact correctly. Model specification
The documentation checked did not identify a UK-specific exclusion or rollout restriction. GitHub’s general staged rollout can still affect UK accounts. Access and data location are separate considerations: xAI documents a global endpoint without a guaranteed processing location, and a US endpoint at a 10% token-price premium. Its regional-endpoint page does not document a UK or EU endpoint. The US guarantee has exclusions, including Files, Collections, server-side tools and network transit. These direct API terms should not be assumed to describe another distributor’s handling of data. Regional endpoints
GitHub now explicitly includes Grok 4.7 in its supported-models and xAI hosting lists. The hosting page describes xAI’s zero-data-retention API policy for those models; this is a retention commitment, not a UK processing guarantee. Supported models, Copilot model hosting
What does Grok 4.7 cost by access route?
Token rates describe metered usage. Subscriptions may include an allowance against which that usage is charged. Cached input means input reused under the provider’s caching rules; repeating a prompt does not necessarily make it eligible.
All rates below are USD per million tokens, checked on 21 September 2026. These tables describe the stated usage-based tiers. Legacy plans or separately agreed terms may differ.
| Route and condition | Input | Cached input | Output |
|---|---|---|---|
| xAI global API, prompt below 200,000 tokens | $2.00 | $0.50 | $6.00 |
| xAI global API, prompt of 200,000 tokens or more | $4.00 | $1.00 | $12.00 |
| xAI US endpoint, prompt below 200,000 tokens | $2.20 | $0.55 | $6.60 |
| xAI US endpoint, prompt of 200,000 tokens or more | $4.40 | $1.10 | $13.20 |
| GitHub Copilot, up to 200,000 input tokens | $2.00 | $0.50 | $6.00 |
| GitHub Copilot, above 200,000 input tokens | $4.00 | $1.00 | $12.00 |
| Cursor standard, up to 256,000 input tokens | $2.00 | $0.50 | $6.00 |
| Cursor Fast, up to 256,000 input tokens | $4.00 | $1.00 | $12.00 |
| Cursor standard, above 256,000 input tokens | $4.00 | $1.00 | $12.00 |
| Cursor Fast, above 256,000 input tokens | $6.00 | $1.50 | $18.00 |
| OpenRouter catalogue, prompt below 200,000 tokens | $1.60 | $0.40 | $4.80 |
| OpenRouter catalogue, prompt of 200,000 tokens or more | $3.20 | $0.80 | $9.60 |
Sources: xAI pricing, US premium, GitHub pricing, Cursor pricing, OpenRouter catalogue.
Notice the exact boundaries: xAI’s higher tier starts at 200,000 prompt tokens, GitHub’s starts above 200,000 input tokens, and Cursor’s starts above 256,000. xAI and Cursor explicitly apply the applicable tier to the full request. The GitHub pricing page checked lists the thresholds but does not explain that allocation in the same detail, so its billing rules should be checked before budgeting a long session.
OpenRouter’s table rows describe its standard catalogue. Its public listing also shows separate Priority endpoints starting at $3.20 input, $0.80 cached input and $9.60 output per million tokens. Those starting prices happen to match the standard long-context row; they are different service conditions. Check the selected endpoint before budgeting. OpenRouter providers
Vercel separately advertises headline launch rates of $1.20 input, $0.30 cached input and $3.60 output per million tokens, with the promotion ending on 27 September. Its listing has additional price tiers, so those headline figures should not be used for every request. Vercel pricing
Direct API priority processing is separate from Fast mode. It doubles standard token rates only when the response confirms priority service; a response served at the default tier uses standard rates. Priority pricing
Fast mode and tool charges
Fast uses the same model on faster infrastructure. SpaceXAI advertises twice the output speed and documents Fast for Cursor and Grok Build, excluding Build’s free tier. It does not document a public grok-4.7-fast API model. Developer guide
The price premium depends on the tier. In Cursor, Fast costs twice the corresponding standard rate up to 256,000 input tokens, but 1.5 times the corresponding standard rate above that boundary. Use Cursor’s own table for Cursor billing. Cursor model pricing
SpaceXAI also publishes a Fast card for Cursor and Grok Build: $4/$1/$12 below 200,000 prompt tokens and $6/$1.50/$18 above it, for input/cached input/output per million tokens. The long-context premium is therefore 1.5 times the matching standard rate, despite the accompanying blanket 2× description. Its below/above wording leaves exactly 200,000 tokens unspecified for Fast; confirm that boundary if it affects your budget. Cursor’s separate 256,000-token boundary still governs its own published rates. SpaceXAI Fast pricing
API tool fees are additional to token charges. Web search and code execution cost $5 per 1,000 calls; attachment search costs $10 per 1,000 calls. From 20:00 BST on 21 September 2026, X Search costs $5 per 1,000 posts fetched and $10 per 1,000 user profiles fetched, replacing its former per-call rate. The dated notice takes precedence over the older X Search row still visible on the pricing page. Tool pricing
Subscriptions and included usage
Cursor’s individual Pro, Pro Plus and Ultra subscriptions are $20, $60 and $200 a month respectively, before tax. They include a Cursor Models usage pool; the checked plan page does not give a fixed numerical Grok 4.7 allowance. Optional on-demand usage is charged separately at the published rates. The India-only Start plan fixes Grok 4.7 at medium effort and standard speed. Cursor plans
GitHub’s current usage-based allowances are:
| Copilot plan | Monthly price | Included AI credits |
|---|---|---|
| Pro | $10 | 1,500: 1,000 base plus 500 flex |
| Pro+ | $39 | 7,000: 3,900 base plus 3,100 flex |
| Max | $100 | 20,000: 10,000 base plus 10,000 flex |
| Business | $19 per user | 1,900 per user, pooled at billing-entity level |
| Enterprise | $39 per user | 3,900 per user, pooled at billing-entity level |
One AI credit represents $0.01. Individual flex allowances can change. Existing annual subscriptions using legacy request-based billing have separate rules. Copilot Free uses automatic model selection only, and GitHub does not list it for selectable Grok 4.7 access. Individual billing, plan comparison, model billing
What changed compared with Grok 4.6?
SpaceXAI reports higher scores while retaining the standard direct API token price. Its launch table shows CursorBench 4.0 increasing from 40.4% to 46.3%, DeepSWE v1.1 from 65.2% to 71.0%, and Terminal-Bench 4.0 from 20.3% to 38.0%. These are vendor-published results. The table compares Grok 4.7 at xHigh with Grok 4.6 at High, except that the 4.7 DeepSWE result uses High. The differences therefore cannot all be attributed solely to the model change. Launch benchmark table
Cursor’s own Insights page offers a comparison at the same named effort level: Grok 4.7 High scores 43.9% on CursorBench 4.0 at an estimated average cost of $4.69 per task, against 40.4% and $5.20 for Grok 4.6 High. Cursor publishes these results and is commercially involved with the model. It also cautions that small score differences may not be statistically meaningful. The cost estimates describe benchmark tasks, not a forecast for your agency’s work. Cursor Insights
Artificial Analysis gives Grok 4.7 xhigh a score of 46 on Intelligence Index v4.3.2. It reports approximately 81,000 output tokens per task, versus 38,000 for Grok 4.6 xhigh. These are evaluation averages, not forecasts for an agency assignment. Higher token use can increase task cost despite unchanged unit prices; it should not be equated with a longer visible answer. Independent evaluation
The more relevant results for research and reporting are the professional-work evaluations. At matching xhigh effort, Grok 4.7 scores 1,657 against Grok 4.6’s 1,555 on AA-Briefcase v1.1, and 1,695 against 1,632 on GDPval-AA v2.1. These are relative Elo ratings, not success percentages. AA-Briefcase, GDPval-AA
AA-Briefcase tests substantial business tasks using supplied source files and tools to create deliverables. It assesses requirements, analysis and presentation; it does not test live web research. In a separate comparison of 4.7 xhigh with 4.6 high, Artificial Analysis reports stronger analytical quality but a slightly lower presentation score. That is not a same-effort comparison or a verdict on brand prose. Methodology, evaluation breakdown
These results support trials of reports and document workflows. They do not establish better AIWIZ copywriting, rankings, conversions or advertising returns.
What web and marketing teams could test
These are proposed applications, not results from an AIWIZ trial. Each assumes approved access to the supplied material. The model can use Search Console, GA4, advertising accounts or client systems only if access is granted through the host product’s tools, connectors or an integration.
Research-led writing and editorial review
Supply an article, an approved source pack and the brand’s writing guidelines. Ask for unsupported claims, contradictions and omissions before requesting a revised draft. Check that edits preserve meaning, use natural UK English and retain accurate citations. Record useful corrections, introduced errors and editing time.
SpaceXAI describes blog posts, reports, white papers and iterative editing as Grok use cases. These are general product claims, not Grok 4.7-specific writing tests. They do not establish superiority in brand voice or persuasion. Long-form writing
Client reporting and campaign briefs
Provide approved campaign exports, metric definitions, date ranges and account notes. Ask for checked calculations, an executive summary and recommendations tied to the evidence. Reconcile every total and distinguish observed changes from explanations that still need testing. For a positioning brief, require sources for competitor claims and separate those observations from proposed messaging.
Grok 4.7 supports web search, X search and code execution when configured. Use calculation tools for the figures and check the cited material behind the commentary. Retrieved social posts are examples of discussion, not a representative customer survey. Developer guide
Copy variants and content repurposing
Start with approved product facts and existing copy. Ask for an email sequence, landing-page alternatives or social posts adapted from a finished article. Review factual fidelity, voice, format and whether each variant offers a distinct angle. Test campaign performance separately after editorial approval.
SpaceXAI documents channel-specific copy and repurposing as intended Grok workflows. That supports trying them, not claiming the new model will produce higher-converting copy. Marketing copy use case
WordPress and JavaScript development
Supply the relevant PHP, JavaScript or TypeScript source, error logs and a bounded specification. Ask for a diagnosis, a patch and appropriate tests. A developer should review and run the tests on staging, then check for regressions. Generated tests can pass while missing the specification.
Measure time to an accepted fix, retries and review minutes. Include the WordPress and PHP versions where relevant; version-specific changes such as those in our WordPress 7.0 guide can affect the answer.
Technical SEO and analytics investigation
Supply a crawl export and affected HTML to investigate canonical tags, or a GTM export and tracking specification to review event names and consent settings. Ask for evidence supporting each finding and proposed correction.
A specialist should verify SEO findings against the live site and Search Console, and investigate tracking with Preview, consent state and network requests. A configuration export does not prove what fired in a live session. Count confirmed findings and false alarms; observe ranking effects separately after implementation.
Long specifications and research packs
Ask for obligations, deadlines and open questions with page references. Verify critical figures and inspect the source for omissions. Check the host’s context limit, leave room for instructions and output, and account for long-context pricing. A large window does not guarantee reliable extraction.
How to test it fairly
Start with five to ten representative tasks and a written acceptance criterion for each. This is a practical pilot, not a statistically representative benchmark. Use the same files, prompts and permitted tools across Grok 4.7 and the models you already use, then repeat the important tasks.
Record the model version, host product, context supplied and reasoning effort: low, medium, high or xhigh where supported. Record Fast separately. Matching effort labels across providers does not establish equal compute budgets, so compare the settings you would realistically use at work.
Set a spending limit before the trial. Score correctness, completeness, time to an accepted result, review minutes and retries. Keep generation speed separate from total completion time: a quick draft can still require a long correction process.
For cost, account for token and tool usage across every attempt, plus human review and rework. Distinguish included subscription usage from additional charges so the same usage is not counted twice. Measure token consumption in your own runs; a lower rate per token does not necessarily produce a cheaper accepted result.
Keep client material within approved access and data-handling arrangements. Use synthetic trial material until the selected route and its tools have been checked against the client’s requirements. The direct API’s documented endpoints do not establish UK or EU processing. Regional documentation
For writing, compare drafts without showing reviewers which model produced them. Score factual accuracy, voice and editorial usefulness separately. Measure commercial outcomes later: faster drafting or maintenance does not itself establish better rankings or conversions.
Should your team test Grok 4.7?
Grok 4.7 is a reasonable candidate for a small, budgeted trial on research, reporting, document review and development, particularly for teams that already have access through their existing tools. The evidence supports evaluating it; it does not establish that it should replace your current default.
For brand copy and campaign strategy, keep existing defaults until a comparison on representative work shows a useful improvement. Include the time spent checking factual accuracy and correcting output. The May 2026 training cut-off also means newer information must come from material you supply or authorised retrieval tools. Model catalogue
If you would like help evaluating AI in your content, reporting or web workflows, get in touch with AIWIZ.
Frequently asked questions
Is Grok 4.7 free to use?
SpaceXAI advertises free access through Grok Build, with Fast excluded. The pages checked did not state a numerical free allowance. The direct API is metered, and the documented selectable Copilot routes are paid plans. Cursor’s Hobby entitlement was not established by the sources checked. Developer guide, GitHub rollout
What is the difference between Grok 4.7 and Grok 4.7 Fast?
Fast uses the same model on faster infrastructure. SpaceXAI advertises twice the output speed, but the price premium depends on the route and context tier. It is documented for Cursor and Grok Build, rather than the public xAI API. Developer guide, Cursor rates
Does long-context pricing apply to the whole request?
xAI and Cursor explicitly apply the relevant tier to the full request once their respective thresholds are crossed: at least 200,000 prompt tokens for xAI and above 256,000 input tokens for Cursor. GitHub lists a higher tier above 200,000 input tokens; check its billing rules rather than assuming another provider’s allocation method. xAI pricing, Cursor pricing, GitHub pricing
Can I use Grok 4.7 in the Grok app or on X?
A model-specific consumer rollout was not confirmed in the documentation checked on 21 September 2026. Developer access through the API, Cursor or Build does not establish access in the consumer apps or through an X subscription. Consumer documentation
Is Grok 4.7 available to UK businesses?
The checked provider documentation did not identify a UK-specific exclusion, but plan requirements and staged rollouts still apply. Availability does not establish UK data processing: the direct API’s regional page documents global and US endpoints, without a UK or EU option. Regional endpoints
Does Grok 4.7 know about current events?
Its training cut-off is May 2026. For newer information, supply current source material or use an authorised retrieval tool, then verify the sources. Search availability and charges depend on the product. Model catalogue
Why can I see Grok 4.7 in a selector but not use it?
A visible model name does not establish access to every feature. Check the plan and any organisation model policy. GitHub’s rollout is gradual, and Cursor feature access also depends on the plan. Linking a SuperGrok subscription for Bot usage does not upgrade a Cursor plan. GitHub rollout, Cursor Bot plans