Grok moderation guidelines: API controls, limits and costs
Grok’s October 27, 2026 update added request-level moderation guidelines. The current guide describes extra restrictions, not a way to disable provider safeguards.
Request controls and compatibility
| Setting | Documented behavior |
|---|---|
| APIs | Responses and Chat Completions |
| guidelines entry | type: policy. category: nsfw, violence or profanity. behavior: decline |
| Omitted category | Keeps default moderation |
| Conversation | Resend guidelines every request |
| OpenAI Python SDK | Use extra_body |
| xAI Python SDK | chat.create retains guidelines for subsequent sample calls |
Multi-agent models and deferred completions are unsupported. Requests with neither server-side tools nor file attachments also reject guidelines when stop is set or n exceeds one.
Unknown categories or behaviors return 422. Duplicate categories, non-policy types and unsupported combinations return 400 before execution.
The nsfw restriction still permits educational or clinical discussion. Violence restrictions allow non-graphic history and news. These are provider-described behaviors, not guaranteed outcomes.
Budget for the model request
The pricing page, checked November 2, 2026, lists these global Grok 4.7 rates in USD per million tokens:
| Prompt size | Input | Cached input | Output |
|---|---|---|---|
| Below 200,000 tokens | $2.00 | $0.50 | $6.00 |
| At least 200,000 tokens | $4.00 | $1.00 | $12.00 |
The higher band applies to all tokens in that request, not just the portion above the threshold. US regional inference adds 10% to global token rates.
For example, 1,000 uncached input tokens and 500 output tokens at the lower global rates total $0.005. This is arithmetic, not a measured request.
Measure accepted requests and actual cost
The cost-tracking guide documents usage.cost_in_usd_ticks: divide by 10,000,000,000 for USD. It includes that request’s token and server-tool charges after applicable discounts. Sum separate turns yourself. The figure is not a cumulative conversation total.
A useful evaluation should keep the model and prompt fixed while comparing ordinary requests against one chosen guideline configuration. Record rejected configurations separately from generated answers. A request accepted by the API is not proof that its answer meets your application’s policy.
Use representative benign cases alongside boundary cases, and review whether legitimate content was unnecessarily declined. Track response quality, latency and billed usage together. ByteForward has not run these tests or measured the controls’ effectiveness.
For another integration concern, see our Grok watermarking checks, which distinguish provenance from application validation.
AI-generated conceptual illustration by ByteForward with the official Grok logo. Not a product screenshot. No endorsement implied.
