How to use Grok 4.7: xAI API, Bedrock, and Copilot pricing and access

Distinguish the stages: xAI API and Bedrock announced, Copilot rolling out

xAI’s September 21, 2026 announcement says Grok 4.7 was available that day through Grok API, Cursor, Grok Build, third-party coding tools, model routers, and cloud platforms. AWS separately announced Bedrock availability on September 28. These confirm channel support, rather than simultaneous access in every region, plan, and account.

GitHub’s announcement that day says now rolling out and rollout will be gradual. Pro, Pro+, Max, Business, and Enterprise are included. If the model is absent, wait for rollout and check the client and administrator policies before buying seats, rotating accounts, or blaming the network.

API boundaries: a 500K context is not a recommendation to fill it

The model page lists text/image input, text output, 500,000-token context, function calling, structured outputs, and reasoning. reasoning_effort supports low, medium, high, and xhigh, defaulting to high. Record the level explicitly: latency, output tokens, and cost can differ for the same prompt.

xAI currently marks Batch API Not supported. The 500K window is a maximum, not a suggested request size. Prompts above 200K tokens use higher xAI API prices; long conversations also increase latency, cache mismatches, and tool-loop costs. Bedrock follows AWS’s card and your account console. Current facts require appropriate search tools; the model’s knowledge cutoff is May 2026.

Price by channel and context, rather than remembering only $2 and $6

Standard xAI API prices per million tokens are $2 input, $0.50 cached input, and $6 output. Prompts above 200,000 tokens use higher-context prices. Record input, cached input, output, prompt length, and reasoning effort in direct-API estimates instead of relying on total tokens.

AWS prices Standard cross-region profiles separately: US Geo costs $2.20 input, $6.60 output, and $0.55 cache read per million tokens; Global costs $2.00, $6.00, and $0.50. Priority is 1.75 times the applicable Standard rate and Flex 0.5 times. Global may route to supported commercial regions. Check residency requirements, US Geo suitability, and your account region before selecting the cheaper Global option.

xAI describes a fast variant with roughly twice the output speed at twice the price, without promising the same fast entry point in Bedrock or Copilot. Its above-200K pricing rule is not a Bedrock billing formula. Copilot follows GitHub provider list pricing and organization usage rules.

Calling Bedrock: check region, profile, permissions, and API

AWS says bedrock-runtime supports this model only through cross-region inference profiles. Use us.xai.grok-4.7 for US Geo or global.xai.grok-4.7 for Global, not direct xAI’s grok-4.7 or bare xai.grok-4.7. Read Grok 4.6 channel differences on Bedrock, then use Grok 4.7’s current region table; older endpoints and regions do not automatically carry over.

For minimal validation, confirm model/profile visibility in the target AWS region’s console, then check IAM bedrock:InvokeModel authorization for the profile and default project. AWS’s example uses a bedrock-runtime OpenAI-compatible URL, Bedrock API key, and us.xai.grok-4.7 for one non-sensitive Responses or Chat Completions request. Converse is supported. Keep keys in secure configuration, never logs, pages, or published examples.

The card lists default projects, streaming, implicit caching, reasoning, structured outputs, and Guardrails. Application inference profiles apply to Invoke and Converse, not Responses or Chat Completions. Server-side tool use, Intelligent prompt routing, and Count tokens are unsupported. Retain the old route if a required capability is missing; generic Bedrock support is not model-specific support.

Enabling Copilot: plan, client, and model policy

GitHub lists VS Code, Visual Studio, Copilot CLI, GitHub Copilot cloud agent, Copilot app, JetBrains, Xcode, and Eclipse. Confirm your Pro, Pro+, Max, Business, or Enterprise plan, update the client, and check its model picker. Visibility differences between accounts on the same plan during rollout are not necessarily faults.

Administrators should check Grok 4.7 in Copilot settings/model policy. New models normally enable with the default on, but a disabled global default or this model blocks members. GitHub places it in usage-based billing at provider list pricing. Check organization budgets and billing fields; xAI token prices alone do not establish your Copilot bill.

Validate small samples in one channel before switching

First choose direct xAI, Bedrock US Geo, Bedrock Global, or Copilot, fixing the model and reasoning level. Second choose representative tasks without sensitive data or external writes. Third measure direct xAI below and above 200K prompt tokens separately; for Bedrock record profile, Standard/Priority/Flex, region, cache reads, output, latency, and bill without borrowing formulas. Fourth validate image input, structured outputs, and required tools. Fifth perform human or automated acceptance rather than relying on vendor scores. Sixth increase traffic only after permissions, budget, error rates, and rollback conditions are satisfied.

If asynchronous batch submission is essential while xAI marks Batch unsupported, retain the older model or use a controlled online queue. Confirm each Bedrock API/feature on the card. For API Key, Base URL, or model setup, see AI API configuration and troubleshooting; for connection, TLS, or407 issues, see proxy connection troubleshooting.

Read vendor benchmarks as candidate evidence, not business promises

xAI reports CursorBench, DeepSWE, Terminal-Bench, professional-knowledge, and safety results against Grok4.6, GPT-5.6 Sol, Fable5.1, and others. Vendor environments, harnesses, reasoning levels, and cost definitions may differ from your tasks. This article does not relabel those results as independent verification.

Compare a fixed task set on success, correct tool use, manual rework, latency, input/output tokens, cache hits, and retries. Stop expanding traffic and return to a verified model if gains are benchmark-specific, real costs exceed budget, or outputs need extensive rework.

Failure, stopping, and rollback conditions

If the Copilot picker lacks the model despite an allowed policy, wait for rollout. For Bedrock model not found, check region, profile, default project/IAM, and ID; do not guess aliases or switch to Global against residency rules. For 429,5xx, timeout, or writes with unknown outcomes, save request-id, time, profile, reasoning effort, token statistics, and task state. Establish prior side effects before bounded retries.

Stop switching and preserve the old model, configuration, and rollback switch when regions/permissions are unconfirmed, Batch is mandatory, tools are incompatible, errors exceed thresholds, or bills differ from estimates. A fixed exit can stabilize the network environment; it cannot grant channel eligibility, enable policy, raise allowances, or change official prices.