Skip to content
CrucibleMark

Grok 4.7

Grok 4.7 is xAI’s Frontier flagship as of September 21, 2026, built on a larger base than Grok 4.6 and trained with a focus on multi-hour tasks. 500,000 tokens of context, text and image input, four reasoning levels from low to xhigh at unchanged pricing of 2 / 6 USD per million tokens. According to xAI, nearly doubled performance on long terminal tasks and a new safeguard stack against jailbreaks.

xAI Version 4.7 Commercial use restricted Dense 500 K Context 05/2026 $2 / $6 per 1M

  • Proprietary
  • Frontier
  • xAI
  • Text
  • Vision
  • Interactive

Sovereign Risk: MEDIUM The model is developed and hosted by a US-based company. Due to US jurisdiction, it is potentially subject to the CLOUD Act, which represents a moderate risk of data access by US authorities. Since the weights are proprietary and not distributed, there is no additional risk from disclosure of the weights themselves.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 100 40 64
HTTP Fetch & Extract 80 40 60
Tool Failure Handling (404) 80 60 73
Web Search & Tool Selection 100 60 82
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 40 70

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Not detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
1.87
First Request
MCP
1.25
Protocol Latency
Synthesis
15.67
Response Generation
Total
112.71
Sum of All Phases
Token
23237
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

Created

Deployment Verdict

Conditional deploy, because Grok 4.7 handles the tool layer correctly in most cases, but synthesis quality at Combined 71.50 and an invalid tool-call signal are not stable enough for trust-sensitive final outputs.

Tool Execution Profile

Tool execution is clearly the stronger side. With P1 90, the model demonstrates that it selects the right tool in most cases within an MCP-backed pipeline and structures calls in a formally usable way. Particularly strong is the Web Search and Tool Selection test, which checks whether the model chooses web_search over fetch without being prompted: here Grok 4.7 acts intelligently rather than purely schematically. This suggests genuine tool selection rather than rigid pattern matching.

Less clean is the URL Construction and Fetch test, which checks whether the model derives a target URL on its own and then retrieves it correctly. P1 80 is usable, but not precise enough for deterministic pipelines with tight success conditions. The global signal “Tool-Call valid: false” therefore remains relevant. There are no indications of a retry problem or format collapse — more likely isolated execution imprecision on specific calls.

Synthesis Fidelity

How well does it condense tool results? Only reliably to a limited degree. P2 53.33 is too low for a Frontier model in a production context, especially because the weakness runs across multiple tasks: EU License Research, HTTP Fetch & Extract, and Multilingual Search & Synthesis each stall at P2 40. The model often retrieves information correctly via tool, but then fails to condense it precisely enough into a reliable answer.

Does it stay within the tool result or fall back on training? In the honeypot EU License Research test, which checks whether current license restrictions genuinely come from web sources, no hallucination was detected. This is the central trust signal. Grok 4.7 fabricates nothing here, but it does not exploit the retrieved content with sufficient discipline.

Error Resilience

On the 404 test, which checks for transparent handling of failed tool calls, the model stays on the safe side. It does not hallucinate page content despite the error. P2 60 shows no elegant error handling, but an acceptable one for production: better to surface a gap than to invent substitute content. For tool pipelines, this matters more than linguistic polish.

Operational Profile

Total 112.71s per run. Call 1: 1.87s. MCP latency: 1.25s. Call 2: 15.67s. Slow relative to the synthesis performance shown. Price per model profile: 2.0 USD per 1M input and 6.0 USD per 1M output, higher for long context. Not cost-efficient relative to the final output quality demonstrated.

Conclusion & Recommendation

Suitable for agentic research and orchestration pipelines where tool selection, web access, and cautious error handling matter more than the initial final formulation. Not suitable for compliance, policy, or executive output stages where the response goes directly to humans or systems without downstream verification. Grok 4.7 makes most sense as a retrieval and planning intermediate step with a downstream validator or a second synthesis instance.

This analysis was generated automatically based on benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.