Inference cost

What it is

Every review is one or more model calls, billed by tokens in and tokens out. The cost of a single review is driven by how much context the tool gathered, how many steps it took, and which model it used.

Why pricing in this category is so hard to compare

Vendors sell seats, credits, reviews, or lines of code — four abstractions over the same meter, with different margins and different failure modes. Seats look predictable until a heavy month; credits look precise until the exchange rate changes; per-review pricing looks fair until you learn that every push re-triggers a review.

The practical consequence: two tools with identical headline prices can differ severalfold in real monthly cost depending on how much context they gather and how often they re-run.

How to estimate before you commit

Take last month’s real numbers: pull requests merged, average pushes per pull request, average diff size. Ask the vendor what a review of that size costs in their unit, and whether pushes re-trigger. Multiply. Then ask what happens when you exceed the bundled allowance, because that is where the surprise lives.

Common mistakes

• Comparing seat prices without normalising for review volume.
• Ignoring re-review behaviour, which can double or triple spend on active pull requests.
• Assuming a cheaper model is cheaper overall — a weaker model that needs more context or more retries can cost more per useful finding.