Grok 4.7 makes completed work the useful price comparison

xAI introduced Grok 4.7 on September 21 with an emphasis on coding and knowledge work. Its announced standard prices started at $2 per million input tokens and $6 per million output tokens, matching Grok 4.6. A faster variant offered twice the output speed at twice the price. The company also described longer reinforcement learning on difficult tasks and stronger self-checking.

That combination changes the purchasing question. An unchanged price per token can still produce a different bill when a model spends longer investigating a problem, reads more files, or repeats more tests. Conversely, a longer run can be economical if it eliminates a second attempt and a substantial manual repair.

Measure the delivered change

For a software team, the useful unit is an accepted change with its review complete. A comparison should record model charges, tool charges, elapsed time, and the minutes a developer spends correcting the result. A patch that passes a narrow test but changes an unrelated interface is unfinished work, even when the agent confidently stops.

A small evaluation can use three distinct job classes: a contained bug fix, a change that crosses several modules, and a documentation task requiring evidence from the repository. Keep the initial files, instructions, and acceptance criteria fixed. Record why each failed run failed instead of combining all failures into one percentage.

The faster variant deserves a separate decision. If a developer is blocked while waiting for output, reduced latency may be valuable. If jobs run overnight and finish comfortably before review begins, paying more for speed may provide little operational benefit. Output speed also does not remove time spent running tests or waiting for external tools.

Budget for the long tail

Before increasing autonomy, set a stopping rule for repeated unsuccessful attempts. A model that persists can make progress on a difficult migration, but persistence without a budget can also become an expensive loop. Preserve intermediate findings so a stopped job does not lose everything it learned.

Procurement teams should retain both the median task cost and the most expensive unsuccessful cases. The former describes normal use; the latter exposes budget surprises. Grok 4.7’s launch pricing makes it a candidate for comparison, while the accepted result, review burden, and failure behavior determine whether it is a useful replacement.