Skip to main content

​
Copy this into your coding agent

Paste into Cursor, Codex, Claude Code, or another coding agent in your project.
The quick scan is available as soon as a system exists; no bench is needed. It reads the components Bench drew from your traces and the prompts your coding agent registered (or a repository scan found) and shows quality and cost findings side by side in Feed, Findings, a system’s Results and the AI systems list. Open a finding to see its right sidebar: what Bench found, the evidence and the next step. Close it to return to the same list. Every finding has a Bench button. An unmeasured prompt or an untested tool is a gap in coverage, not a confirmed defect.

​
What the scan checks

The scan does not execute your application or produce a measured quality score. AI review suggestions cite their evidence and remain suggestions until tested. An AI review cannot prove that an unexecuted tool works. An additional Evidence review, when available, checks whether cited sources support the AI review’s claims. Its percentage is an estimate of evidence agreement, not your application’s quality score. Missing or oversized evidence stays unavailable; it does not become a passing check.

​
Potentially lower model cost

Each cost finding shows a model’s cost per 1,000 typical requests (1,500 input and 400 output tokens each) at its listed price, next to a lower-priced model in the same family, with the potential saving. Potentially 86% lower model cost means the candidate has a lower price for the same input and output volume. It does not mean your production bill has fallen. For example, standard text prices checked on 19 September 2026 are: Both prices are about 86% lower for GPT-5 Mini. At 1,500 input and 400 output tokens per request, that is about ~$8.23 versus ~$1.18 per 1,000 requests. This example assumes identical usage, standard text requests, and no caching, batch discounts, tools, hosting or negotiated rates. Different models can use different numbers of tokens, retries or tool calls. Sources: GPT-5.2 pricing and GPT-5 Mini pricing. The finding sidebar identifies both models and links to their price sources. If input and output prices fall by different percentages, Bench shows a range. Cost is never shown as unknown. Bench uses the model’s listed price when it has one; otherwise the rate for its family (Sonnet, Haiku, GPT-5 mini and so on); otherwise a typical mid-tier rate. Anything not taken from a listed price is labelled as an estimate with ~ and “est.”. Listed prices older than 30 days are still shown, labelled as estimates until the price list is refreshed.

​
Turn an opportunity into evidence

  1. Bench the original setup against system-specific criteria and cases.
  2. Test supported candidates against the same cases.
  3. Compare quality, failures and estimated model cost per call.
  4. Validate the proposed change on independent cases before rollout.
  5. Use production usage after deployment to verify the real impact.
A cheaper model is not automatically a better choice. The recommendation must meet the quality and regression checks. A quick-scan comparison also does not guarantee that the named alternative is enabled in the current candidate set. During evaluation, Bench calculates model cost from recorded input/output usage and its price catalog. Judge and test-generation spend are separate from the customer workflow’s projected cost. These are estimates, not imported invoices. See evaluation results and production checks.

​
Why there is no predicted quality percentage before testing

Repository structure and model price do not establish answer quality. A quality gain needs a baseline and a candidate evaluated on the same representative cases. A small preliminary test can support a provisional result, but it must show its sample size and cannot stand in for a full benchmark or independent validation. Learn about real app testing and simulation testing, including what each can verify.