Grok 4.7's $6 Output Can Become $13.20 • The Weighted Average

Teams evaluating xAI’s September 21 Grok 4.7 release should test their actual context and processing region before treating its unchanged $6-per-million output price as a budget. Combining the long-context rate with the US endpoint’s 10% premium yields $13.20, or 2.2x the base global output rate, without changing the model.

The upgrade keeps the price, not the bill

The launch is a meaningful capability update, not an announced base-price cut. xAI says Grok 4.7 uses a larger base model, longer reinforcement learning, and more difficult long-running tasks than its predecessor. It emphasizes checking work and handling longer context. Those claims justify an evaluation for teams whose current agents struggle through extended coding jobs. They do not establish how many tokens the new model will consume on a particular repository.

The September release notes list $2 uncached input, $0.50 cached input, and $6 output per million tokens below the long-context threshold. They list $4, $1, and $12 respectively above it. The pricing table places the threshold at 200,000 prompt tokens and says long-context rates apply to all tokens in the request. That is not a marginal surcharge on only the excess context. A budget built from average prompt size can miss which requests cross the tariff boundary.

The distinction between capacity and price is essential. The model overview advertises a 500,000-token context window. That is the maximum supported window, not the amount available at the lower rate. Long-running agents need separate controls for fitting within the model and for staying within their approved spending policy. A request can satisfy the first condition while violating the assumptions behind the second.

Now add location. The regional endpoint documentation applies a 10% token premium, including long-context rates. The original comparison is therefore $12 × 1.10 = $13.20 per million output tokens, followed by $13.20 ÷ $6 = 2.2x. Inputs come from the release’s tariff and the regional documentation. The calculation is a configuration comparison, not an estimated customer invoice or an assumption about how much output any job produces.

Context and region can lift Grok's output rate 2.2x

Standard API output, $/1M tokens. Long context starts at 200k prompt tokens; US adds 10%.

xAI release notes and regional endpoint documentation · Sep 22, 2026

The chart deliberately keeps the model fixed. It does not suggest that US processing or long context is wasteful; each may be necessary. It shows why procurement cannot approve a model name and infer a single unit cost. Record region, context tier, cache treatment, and service tier alongside the model. Otherwise the engineering configuration and the financial approval describe different purchases.

This extends our earlier analysis of Gemini’s scheduled price expiry, where a calendar boundary changed the tariff. Grok introduces a different budgeting boundary: the request itself. The lesson is not that the higher price is unacceptable. It is that a default route should be judged at the rates the intended workload will actually encounter, rather than at the most attractive row of an announcement.

Buy the configuration you can account for

Caching is the first practical lever, but it must work in the deployed harness. The Grok 4.7 overview recommends a stable prompt cache key, and the cache-routing guide explains how it keeps requests on the same server. An integration that omits this routing hint may pay full input price on a cold server. Measure cache hits rather than assuming that repeated conversation text automatically receives the discounted rate.

Compaction addresses a related but different problem. The context-compaction documentation describes replacing accumulated history with an opaque compacted item. It can reduce subsequent input and keep a conversation from expanding indefinitely. However, compaction itself consumes tokens, and the conversation must still fit the model’s window when compaction runs. Waiting for an over-limit failure is not an implementation of context management.

Treat that compacted state as an API artifact, not editable prose. xAI instructs clients to pass it back unchanged. Separately, Grok 4.7 always returns encrypted reasoning content on the Responses API, even without explicitly requesting it, and asks clients to preserve reasoning items across turns. Before changing the model identifier, check whether the wrapper retains these items. A migration that discards state can change behavior even if a simple single-turn test succeeds.

The billing instrumentation is unusually useful. xAI’s cost-tracking documentation exposes the exact per-request charge in cost_in_usd_ticks, including token and server-side tool costs after applicable discounts. The value is per request, not a cumulative conversation total. Sum it across the full task, including retries, and reconcile that total with accepted outcomes. Counting only the last response would understate the cost of a long agent run.

Speed has separate purchase paths. The model guide says Grok 4.7 Fast is confined to Cursor and Grok Build, not the public API. API users instead have Priority Processing, which confirms the tier actually served in the response. Do not treat the product names as interchangeable service guarantees. Check the current pricing table and returned tier before attributing a latency change to a paid upgrade.

There is a documentation discrepancy worth resolving before buying the fast route. The pricing page describes Fast as twice the standard token rates, but its detailed table lists $18 for long-context output, rather than twice the ordinary $12 long-context rate. Do not extrapolate the short-context multiplier into a contract. Confirm the tariff shown by the actual purchasing interface and retain that evidence with the evaluation. The discrepancy does not affect the standard API and US-region arithmetic above; it does prevent treating a broad product description as an exhaustive rate card.

Existing Copilot users can reduce integration scope. GitHub is rolling Grok 4.7 into Copilot with provider-list usage billing, and administrators can manage access through model policy. That offers a way to test the model within an established interface. It does not prove that a third-party route inherits the direct API’s regional guarantees or retention settings. Ask the intermediary about its own contract and data path.

The benchmark is an invitation, not a waiver

xAI’s strongest coding comparison needs a footnote in the buyer’s head. Its launch table reports 46.3% on CursorBench 4.0 for Grok 4.7 versus 40.4% for Grok 4.6. Subtraction gives a 5.9-percentage-point gain. But the columns label the new model xHigh and the predecessor High. This is vendor evidence under different reasoning settings, not an experiment isolating the model change at identical effort and cost.

The improvement can still matter. A model that checks difficult work more effectively may avoid failed attempts and expensive human repairs. Conversely, a model that deliberates longer can raise the bill without improving the tasks your team actually accepts. Neither possibility is settled by the launch table. The appropriate response is a matched internal evaluation that preserves review rules and records full task cost, not a declaration that the cheaper-looking benchmark column has won.

The US endpoint also has a narrower guarantee than its name may imply. xAI covers API request handling, inference, moderation, and retained request data in the United States. It explicitly excludes Files, Collections, server-side tools, and the network path from the customer’s systems. Those features may work through the endpoint while remaining outside its location guarantee. A contract covering an entire workflow cannot be satisfied merely by changing the inference base URL.

Retention is a separate setting again. The security FAQ says API requests and responses are retained for 30 days by default, without training on them absent explicit permission. Zero Data Retention is team-wide and disables capabilities that depend on storage, including stateful Responses, Files, Collections, and Batch. That is a functional migration boundary, not a decorative compliance toggle. Verify the intended interaction pattern before promising both stateless handling and server-maintained conversation history.

The strongest argument against switching is therefore not a rival leaderboard score. It is an integration whose essential behavior depends on assumptions the new route does not satisfy. A lower token price cannot compensate for unacceptable handling terms, lost conversation state, or a failure to finish required tasks. Keep the incumbent available while testing those boundaries; rollback is cheaper when it is designed before the first production failure.

Competition supplies another useful check. Today’s MiMo V2.6 brief finds a 6.9x gap between ordinary output tariffs, without claiming equivalent task economics. Today’s Strands analysis normalizes a vendor harness benchmark by passed tasks. Together they show why neither model price nor harness reputation should decide the whole purchase. The deployable unit is a model, runtime, tools, controls, and workload tested together.

Promote a measured route, not a launch-day favorite

Start with the tasks that motivated the upgrade. Preserve the same repository state, test commands, reviewer expectations, and tool permissions across candidates. Record the exact reasoning effort rather than relying on defaults. Separate short jobs from long-running jobs, because their context and retry patterns can expose different tariff behavior. These are proposed evaluation controls, not results of a benchmark performed for this article.

Give each task a complete accounting record. Capture total charged requests, cache usage, context tier, returned service tier, endpoint, elapsed time, and final acceptance. Include rejected attempts and any manual intervention required to repair the result. Human time should remain a separate measured quantity unless the organization supplies a real labor-cost input. There is no need to invent an hourly rate to identify a route that creates more review work.

Then inspect the boundary cases deliberately. Test near the long-context threshold using the current pricing table, verify that compaction preserves necessary evidence, and confirm encrypted state survives the harness. Exercise the no-storage mode if it is contractually required. For US processing, inventory every server-side tool and storage feature rather than assuming that inference placement covers them. Each test answers a concrete uncertainty disclosed by the documentation.

A subscription pilot still needs its own scope. Our Grok Build analysis distinguished weekly allowance from API unit price. The new model does not make those purchasing units identical. A developer evaluating an included allowance should observe continuity through the relevant usage cycle, while a platform owner purchasing API calls should reconcile metered spending. Do not compare an interface subscription with raw inference rates as if both promise the same capacity.

Evidence that would change the verdict is repeatable improvement in cost per accepted task, acceptable latency, and handling terms that cover the actual workflow. Evidence against it is a gain confined to the vendor’s reasoning configuration, longer traces without better outcomes, or a necessary regional or retention setting that breaks the integration. A revised tariff would update the arithmetic; it would not remove the need to test usefulness.

Coding-agent owners: trial Grok 4.7 where long tasks currently fail, keeping the incumbent and acceptance rules fixed. Promote only measured improvements, not the launch’s best benchmark row.
Platform and finance teams: budget the configured route. Long-context US output is $13.20 per million tokens, 2.2x base global output; reconcile actual request charges rather than multiplying all traffic by $6.
Security and procurement owners: verify retention and the full regional boundary. Files, tools, and intermediary platforms need their own review before sensitive workloads move.

The decision this quarter is to add a well-instrumented candidate, not crown a permanent default. Grok 4.7 may deliver more useful work at the same base tariff. Whether that becomes a lower operating bill depends on the configuration and the results. A model’s price is a property of the configured request, not its name.

The upgrade keeps the price, not the bill

Buy the configuration you can account for

The benchmark is an invitation, not a waiver

Promote a measured route, not a launch-day favorite

Sources