What it is
Grok 4.7 is the flagship model of SpaceXSI, the company formerly called xAI, which SpaceX acquired in February 2026 and renamed SpaceXAI in July and SpaceXSI in October; its docs and API still use x.ai domains. Released on 21 September 2026, it is recommended for chat, code and anything that does not need audio, image or video output. It reads text and images, writes text, and has a 500K-token context with a May 2026 knowledge cutoff.
The API price is $2 per million input tokens and $6 per million output tokens, with cached input at $0.50, for prompts under 200K tokens; at 200K or more, every token in the request is billed at double. Reasoning effort runs from low to xhigh, with high as the default, and on the Responses API the reasoning is always returned encrypted. A US-only endpoint costs 10% more.
It is available on the API, in Grok on X and the Grok apps, in Cursor and Grok Build, and through OpenRouter, Vercel and Cloudflare. A Fast variant at twice the price is limited to Cursor and Grok Build.
Our review
Grok 4.7's per-token price flatters it. At $2 input and $6 output per million tokens it undercuts Claude Sonnet 5.5 and GPT-6.1 Sol on output, both at $10, yet Artificial Analysis found it used 240 million output tokens to run its Intelligence Index, 112th of 115 models for verbosity against a median of 87 million. The result is about $3.74 per index task, 100th of 112 on cost.
Capability is respectable rather than leading. In its xhigh setting it scores 46 on the same index, 30th of 214 models, against a class median of 27. Output runs at a reasonable 95.9 tokens a second, but the first token takes 38.9 seconds because reasoning comes first, against a median of 3.85 seconds. An interactive feature built on it needs a lower effort setting or an interface that tolerates a long pause. SpaceXSI has not published a benchmark report for the model that we could find.
The pricing table holds a trap and an alternative. Prompts of 200K tokens or more are billed at double for every token, not just the excess, so a single 250K-token prompt costs $1 in input alone. The older Grok 4.3 and 4.20 models keep 1M-token contexts at $1.25 and $2.50, which suits long-document work better than 4.7's 500K window. Logprobs are silently ignored from 4.20 onwards, which rules the family out of calibrated classification.
The API data terms are clearer than the company's public record. SpaceXSI says it never trains on API inputs or outputs without permission, keeps requests for 30 days for abuse review and is SOC 2 Type 2; team-wide zero data retention is available but switches off the stateful Responses API, Batch, Files and hosted image output. Encrypted reasoning that must be passed back unchanged also means a developer debugging an agent can see which tool the model called, but not why.
Where it earns its place
A developer building social-listening dashboards for a PR agency
Classify and summarise brand mentions across X and the web
Call grok-4.7 with the X Search tool at low effort to cut latency and tokens. At 3,000 input and 500 output tokens a call, 100,000 calls a month is about $900; a prompt_cache_key bills repeated system prompts at $0.50 per million.
A backend engineer on a team that already pays for Cursor
Run a coding agent on multi-file refactors
Grok 4.7 is the default model in Grok Build and available on all Cursor plans, billed through the plan; the Fast variant costs twice as much. Long first-token waits matter less in background agent runs than in chat.
A compliance officer at a US health-tech startup
Approve an LLM vendor for patient-adjacent workflows
Use the US endpoint at 10% extra to keep inference in the United States, request a BAA through SpaceXSI's questionnaire, and decide whether zero data retention's loss of Batch and stateful features is acceptable.
Before you commit
- Setup
- Create a key at console.x.ai and call grok-4.7 on the Responses or Chat Completions API with OpenAI-compatible SDKs; pass returned encrypted reasoning items back unchanged.
- Integrations
- Function calling, Web Search, X Search and code execution tools; also on OpenRouter, Vercel, Cloudflare, Cursor and Grok Build.
- Data and privacy
- API: no training without explicit permission, 30-day retention for abuse review, SOC 2 Type 2, BAA on request; team-wide zero data retention disables stateful Responses, Batch, Files and hosted media.
- Cost at scale
- 1M support replies a month at 2,000 input and 1,000 output tokens, reasoning included, is about $10,000 before caching; prompts at or above 200K tokens double every token in the request.
- Limits
- 500K context and no text output cap; 38.9-second first token at xhigh on Artificial Analysis; logprobs ignored; rate limits not published on the model page.
What's good
- $2/$6 per 1M tokens, cached input $0.50
- Intelligence Index 46, 30th of 214 on Artificial Analysis
- Built-in Web Search and X Search tools
- No training on API data without permission; optional zero retention
- US-only endpoint for data residency
What's not
- 38.9 seconds to first token at xhigh
- 240M tokens to run the index; $3.74 per task
- Every token billed double on 200K+ prompts
- 500K context, half that of the cheaper Grok 4.3
- Reasoning always encrypted; logprobs ignored
Best for
Developers who need live X search inside an agent and can tolerate long thinking times, or teams already committed to SpaceXSI's platform.
Alternatives
Grok 4.7: common questions
- Is Grok 4.7 worth it?
- Cheap on the rate card and dear in practice: Grok 4.7 thinks for so long that its cost per task lands among the highest measured. We rate it 3.8 out of 5.
- How much does Grok 4.7 cost?
- API: $2 input / $6 output per 1M tokens under 200K-token prompts, $4 / $12 at or above; cached input $0.50; US endpoint 10% extra.
- What are the downsides of Grok 4.7?
- 38.9 seconds to first token at xhigh; 240M tokens to run the index; $3.74 per task; every token billed double on 200K+ prompts; 500K context, half that of the cheaper Grok 4.3; reasoning always encrypted; logprobs ignored.
- Who is Grok 4.7 best for?
- Developers who need live X search inside an agent and can tolerate long thinking times, or teams already committed to SpaceXSI's platform.
- How much does Grok 4.7 cost for a team or in production?
- 1M support replies a month at 2,000 input and 1,000 output tokens, reasoning included, is about $10,000 before caching; prompts at or above 200K tokens double every token in the request.
- Is Grok 4.7 safe for company data?
- API: no training without explicit permission, 30-day retention for abuse review, SOC 2 Type 2, BAA on request; team-wide zero data retention disables stateful Responses, Batch, Files and hosted media.
- What are the best Grok 4.7 alternatives?
- The closest alternatives we have reviewed are GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 3.8 Flash.
Answers checked Oct 23, 2026.