Grok 4.7 Responses API: preserve encrypted reasoning and cache affinity in long agent loops

Quick answer

Grok 4.7 is available on the xAI API as grok-4.7. It keeps Grok 4.6's token rates, but every Responses API response includes reasoning.encrypted_content, even if include did not request it. Treat reasoning items as opaque provider state and pass them back unchanged.

For long sessions, set a stable prompt_cache_key per conversation. xAI says it helps route related requests to the same server; a cache-cold request may pay full input price. Require multi-turn replay, tool pairing, cache evidence, and cost receipts—not one successful answer.

Who this is for

This guide is for teams calling xAI directly or through a gateway when an agent runs tools, resumes sessions, or grows past 200,000 prompt tokens. GitHub Copilot policy and AI Credits are a separate control plane.

If you are still choosing the model, start with the Grok 4.6 versus 4.5 accepted-task workflow. This page begins after you select a canary workload.

What changed, and what did not

SpaceXAI released Grok 4.7 on September 21 for coding, agentic tasks, and knowledge work. The API model has a 500,000-token context window, text and image input, text output, function calling, and low, medium, high, or xhigh reasoning effort. high is the documented default.

The price schedule matches Grok 4.6:

Prompt tier	Input / 1M	Cached input / 1M	Output / 1M
Up to 200K prompt tokens	$2	$0.50	$6
Above 200K prompt tokens	$4	$1	$12

The threshold applies to each request. A session that drifts above 200K can double all three token rates. Grok 4.7 Fast runs on faster infrastructure at twice the rates, but is offered only in Cursor and Grok Build—not the public xAI API. Publisher benchmarks justify testing, not migration; your repository and acceptance rules decide whether a task improves.

The four boundaries to preserve

Boundary	What to keep	Failure if ignored
Conversation state	Every structured response item, including encrypted reasoning	Resumed turns lose provider state or fail validation
Tool pairing	The original tool call ID and exactly one matching result	The model sees an orphaned or duplicated tool result
Cache affinity	One stable, non-secret cache key per conversation	Related turns reach cache-cold servers and cost more
Cost tier	Prompt tokens, cached tokens, output tokens, and actual charged cost	A 200K crossing or retry loop silently breaks the budget

Encrypted reasoning is opaque. Do not parse, summarize, or rewrite it. Preserve the item's type and payload. “Encrypted” also does not make the surrounding prompts, tool arguments, URLs, or outputs safe for unrestricted logs.

A seven-step migration workflow

1. Freeze a 4.6 baseline

Choose a short edit, a tool-using fix, and a multi-turn task that resumes after restart. Record the slug, effort, token counts, tool calls, retries, time, and acceptance result.

2. Audit the serializer before changing the model

Unknown structured items must survive a store-and-read cycle. Reject code that keeps only response.output_text; text is not the full continuation state.

3. Switch only the explicit model slug

Change only grok-4.6 to grok-4.7. Keep prompts, tools, timeout, effort, and retries fixed so model behavior is not confused with a harness change.

When continuing a turn, append every structured response.output item to the next input before the new user message. Storage varies by SDK; provider state still returns unchanged as provider state.

4. Add cache affinity deliberately

Derive a stable prompt_cache_key from a random internal conversation ID. Include no email, repository secret, or customer name, and never share one key across unrelated users.

5. Reconcile actual cost, not estimated tokens alone

xAI responses expose usage.cost_in_usd_ticks; divide it by 10,000,000,000 for the charged US-dollar amount. Store it with token counts and the prompt tier. Estimates are preflight; returned cost is post-request evidence.

6. Run five canaries

• Single turn accepts valid text plus encrypted reasoning.
• Persisted items survive restart and resume.
• Each tool call gets one result and one side effect.
• Stable repeated context records cached tokens.
• A malformed fixture fails before retry or tool action.

Do not create a 201K-token request just to prove pricing. Forecast with recorded growth, then observe the tier when a real canary approaches it.

7. Promote by accepted-task evidence

Promote only if every task passes, no state item drops, tool side effects remain exactly once, and accepted-task cost stays in budget. Otherwise return to 4.6 and keep the failed receipt.

Decision tree

Does the client preserve unknown structured response items?
  no  -> fix storage and replay before changing the model
  yes -> run the multi-turn and tool canaries

Do resumed turns preserve encrypted reasoning and tool pairing?
  no  -> roll back; inspect serialization and gateway transforms
  yes -> compare cache and cost receipts

Does accepted-task cost stay within the budget at projected context growth?
  no  -> shorten context, compact, lower effort, or retain 4.6
  yes -> expand the canary cohort gradually

Common mistakes

• Saving only output_text or modifying reasoning.encrypted_content.
• Putting identifying data in cache keys or sharing them across customers.
• Calling a cache miss a model regression without routing evidence.
• Applying the $2/$6 headline above 200K prompt tokens.
• Requesting Grok 4.7 Fast from the public API.
• Treating publisher benchmarks as private-repository proof.

Building something? Turn your product page into a show people want to watch with PromoFast—hosted, embeddable, and ready to export.

FAQ

Is Grok 4.7 more expensive than Grok 4.6 on the xAI API?

The rates and 200K threshold are the same. Accepted-task cost can still change with tokens, cache hits, effort, retries, and tools.

Must I request encrypted reasoning explicitly?

No. Release notes say Responses API calls always return it even when include omits it. Your client must still preserve and replay the item.

Is Grok 4.7 Fast an API model slug?

Not on the public xAI API. Official documentation limits the Fast variant to Cursor and Grok Build.

Does a cache key guarantee a cache hit?

No. It improves conversation affinity according to xAI's guidance. Measure cached tokens and actual cost rather than treating the key's presence as proof.

Sources

• SpaceXAI: Introducing Grok 4.7
• SpaceXAI developer guide: Grok 4.7
• SpaceXAI API release notes
• SpaceXAI cost tracking
• GitHub: Grok 4.7 in Copilot
• GitHub Copilot supported models

Originally published on IndieSeek.