Short answer
Yes. xAI published an official Grok 4 Model Card, last updated August 20, 2025. It describes safety-evaluation areas and safeguards for original Grok 4; the passages summarized here do not establish how effective those safeguards were. It is not documentation for Grok 4.6, Grok 4.7, or Grok 4.20. For production planning, note that xAI’s later retirement notice says requests using the original grok-4-0709 API slug are routed to Grok 4.3 with low reasoning effort—not the original model. (Grok 4 Model Card)
What the official Grok 4 model card says
The card describes Grok 4 as a reasoning model with tool-use capabilities and organizes its safety work around abuse potential, concerning propensities, and dual-use capabilities. It says it evaluates both the GROK 4 API and GROK 4 WEB, including robustness to jailbreaks and prompt injections, and describes safeguards such as refusing requests that may enable foreseeable serious harm. The card’s stated examples include CBRN or cyber weapons, self-harm, and child sexual abuse material. These passages describe evaluation scope and mitigation intentions; they do not, by themselves, establish that the evaluations demonstrated effective protection. (Grok 4 Model Card)
A concrete, separately scoped finding comes from Microsoft’s evaluation for Grok-4 offered through Azure Direct: Microsoft reports higher risks of potentially harmful content and successful jailbreak attacks than for other models it evaluated and offered through Azure Direct. Microsoft requires system safety messages and Azure AI Content Safety for that deployment, while warning that these safeguards may not mitigate all risks. This is Microsoft’s deployment-specific evaluation—not a result to generalize to every Grok 4 provider or setup. (Microsoft Foundry)
A useful capability result comes from xAI’s launch announcement: on the full Humanity’s Last Exam set with Python and internet tools, xAI reports 38.6% for Grok 4 and 44.4% for Grok 4 Heavy. The configurations matter: the 44.4% result is for Heavy, not standard Grok 4, and both figures are reported with tools. The launch post also describes native tool use, including code-interpreter and web-search use. (xAI Grok 4 announcement)
There is a version-specific caveat about inputs and limits: third-party summaries disagree on whether original Grok 4’s API accepted images, and the primary card and launch-announcement passages cited here do not settle that question or establish the API’s context and output limits. Verify these requirements against documentation for the precise model and provider; do not infer them from a later Grok version or an uncited summary.
The card’s last-updated date is August 20, 2025, so it explains xAI’s documented evaluation approach for the original model but does not establish current API availability. (Grok 4 Model Card)
Don’t mix up the versions
“Grok 4” is sometimes used loosely for the model family. The official card concerns Grok 4; Grok 4 Heavy is a separate configuration, and Grok 4.6, 4.7, and 4.20 are later versions. The launch announcement reports separate standard and Heavy benchmark results, but the card excerpts available here do not clarify whether every evaluation also covers Heavy. Keep the benchmark configuration attached to each score. Do not use a later version’s benchmarks, limits, or API details as specifications for original Grok 4. (xAI Grok 4 Model Card; xAI Grok 4 announcement)
Can you still call original Grok 4 in production?
Not as the original model through xAI’s API. xAI’s retirement notice says that after May 15, 2026, requests to grok-4-0709 are served by Grok 4.3 with low reasoning effort. The old slug may continue to resolve, so a successful request does not mean the original model is still running. xAI’s documentation lists Grok 4.3 pricing at $1.25 per million input tokens and $2.50 per million output tokens; confirm the current terms and limits before estimating production spend. (xAI May 15, 2026 model retirement notice; xAI Grok 4.3 documentation)
Illustrative cost for Grok 4.3: 10,000 input tokens and 1,000 output tokens total $0.015 at those standard rates: (0.01 × $1.25) + (0.001 × $2.50). This small example assumes standard-rate eligibility; it is not a quote for every request or deployment. (xAI Grok 4.3 documentation)
For a new or updated integration, choose the intended current model explicitly instead of relying on the retired slug’s redirect. The xAI Grok 4.3 page lists a 1,000,000-token context window, text and image input, text output, function calling, structured outputs, and configurable reasoning. These are Grok 4.3 specifications, not original Grok 4 specifications. Check the current provider documentation for supported features, limits, and pricing before deployment. (xAI Grok 4.3 documentation)
How to decide whether a Grok model fits
- Name the exact version and provider. The original Grok 4 card does not document later versions, and a third-party model listing may not reflect xAI’s API behavior or pricing.
- Confirm which model your request will reach. xAI says
grok-4-0709requests are routed to Grok 4.3 at low reasoning effort—not the original Grok 4. Make that change explicit in your configuration and tests. (xAI May 15, 2026 model retirement notice) - Check workload-critical features. Verify required input types, context size, tool behavior, output format, and safety constraints in documentation for the exact model and endpoint. The original Grok 4 modality discrepancy remains unresolved in the sources cited here.
- Benchmark representative tasks. Compare answer quality, latency, failure modes, and total cost using your own prompts and expected traffic. Published benchmark results do not establish which currently available model will perform best on your workload.
Bottom line: use xAI’s Grok 4 Model Card to understand its stated safety-evaluation scope and safeguards, without treating those descriptions as evidence that the safeguards are effective. Microsoft’s Azure Direct evaluation reports higher harmful-content and jailbreak risks for its Grok-4 deployment than for other models it evaluated there; its required safeguards may not mitigate every risk. For production selection, treat grok-4-0709 as retired and validate an explicitly chosen current model against your requirements and current provider terms.