xAI shipped Grok 4.7 on September twenty first with a 500k context window and a price that doubles past 200k prompt tokens. Two documented behaviors decide what you actually pay and whether multi turn agents keep their reasoning: setting prompt_cache_key, and returning reasoning items unchanged.
Grok 4.7 arrived on September twenty first, and the two details most likely to cost you money are not on the benchmark chart. They are in xAI's own documentation, and both of them are about how you call the model rather than how good it is.
The first is caching. xAI highly recommends setting a prompt_cache_key on grok-4.7, and states plainly what happens if you skip it: without one, you often pay full input price on a cache cold server. That has always been worth money. It is worth roughly twice as much on this model, because grok-4.7 prices in two tiers. Below 200k prompt tokens, input runs $2.00 per million tokens and cached input runs $0.50. Above 200k, those become $4.00 and $1.00. Output is $6.00 below the threshold and $12.00 above it. So a long agent loop that reuses a large system prompt and tool schema, but never supplies a cache key, is paying four times the cached rate on every call and doing it at the expensive tier. Note also that xAI documents a ten percent surcharge for routing requests to the US regional endpoint, which stacks on top of all of this.
The second is reasoning state. On the Responses API, grok-4.7 always returns reasoning.encrypted_content, even when your include list does not ask for it. xAI's instruction for multi turn conversations is to pass the reasoning items back unchanged in the next request's input. Two failure modes follow from that, and neither one throws an error. If your response handler strips unknown fields before storing a turn, you lose the reasoning context and the model reconstructs it from scratch on every turn, which is slower and more expensive and often worse. And if your parser assumes encrypted_content only appears when requested, it may choke on a field it did not expect on a model where that field is unconditional.
Effort is the third dial and it is easy to get wrong by leaving it alone. If you do not set reasoning_effort, grok-4.7 defaults to high. The supported values are low, medium, high and xhigh. Reasoning tokens bill as part of total consumption, so the default is not the cheap setting, it is the expensive one, and a workload of short classification or extraction calls is paying for deliberation it does not need. The xhigh level is available on grok-4.6 and later; on a model that does not support it, such as grok-4.5, a request for xhigh is silently treated as high, which means a fallback path in your code can quietly change the model's behavior without telling you.
Grok 4.7 lists Responses API and Chat Completions, with function calling, web search, X search and code execution among its tools. The capability is not the interesting part this week. The interesting part is that three documented defaults, one absent parameter, one unconditional field and one silent downgrade, sit between a correct integration and a bill two to four times larger than it needs to be.
Try it today
- Set prompt_cache_key on every grok-4.7 call that reuses a system prompt, tool schema or document preamble. Use a stable key per logical conversation or per agent, not a random value per request.
- Measure where your prompts sit relative to the 200k prompt token threshold. Anything hovering just above it is paying double on input and output; trimming retrieval or compacting context to land below the line halves the unit cost.
- Audit your Responses API handler for reasoning.encrypted_content. Confirm it is stored and returned unchanged in the next request's input, and that no field stripping or schema validation silently drops it.
- Set reasoning_effort explicitly rather than inheriting the high default. Try low or medium on classification, extraction and routing calls, and reserve high or xhigh for genuine multi step work.
- Check any model fallback path. A request specifying xhigh that fails over to grok-4.5 is treated as high, so your fallback is quietly running a different reasoning budget than your primary.
- Factor the ten percent US regional endpoint surcharge into cost models before comparing grok-4.7 against alternatives on list price alone.