An eval (short for evaluation) is a repeatable test of an AI system: you give it a set of inputs, then grade what comes out against standards you chose. It is how you know a prompt, model or agent works for your task, rather than hoping it does.
Think of a driving test. The examiner does not ask whether you feel like a good driver; they take you through the same set of manoeuvres every candidate does and mark each one against a checklist. An eval is a driving test for an AI: same route, same checklist, every time you change something.
In a sentence: “Before switching our support bot to the new model, we ran our eval of real customer questions, and it got worse at refunds, so we waited.”
How is an eval different from a benchmark?
The plain point: a benchmark is a public test everyone uses; an eval is usually your own test, built from your own work.
| Term | How it differs |
|---|---|
| Benchmark | Differs: a shared, standard test that compares models with each other. It tells you little about your own task. |
| Eval | Covers: your own cases and your own idea of a good answer, run every time you change the prompt, model or setup. |
| Red teaming | Differs: people deliberately trying to make the AI misbehave. Its findings can become eval cases. |
| Vibe check | Differs: trying a few prompts and judging by feel. Quick, but not repeatable. |
Anthropic, which makes Claude, defines an eval as “a test for an AI system: give an AI an input, then apply grading logic to its output to measure success.”1 OpenAI’s guide says evals “test model outputs to ensure they meet style and content criteria that you specify.”2 Benchmarks, by contrast, are “standardized datasets or tasks”, as IBM puts it.3
Why do evals matter?
The plain point: AI answers vary, so a few good tries prove little; a fixed test shows whether a change really helped.
The same question can get a different answer each time. OpenAI says this is why ordinary software testing falls short: “Generative AI is variable. Models sometimes produce different output from the same input”.4 Without evals, Anthropic warns, teams “get stuck in reactive loops—catching issues only in production”, meaning they find problems only after real users hit them.5
Anthropic’s prompting docs put evals at the start of the work, not the end: building an AI app “starts with clearly defining your success criteria and then designing evaluations to measure performance against them.”6
How is an eval graded?
The plain point: by a fixed rule, by another AI with a marking guide, or by people.
Anthropic describes three kinds of grader: code-based checks (such as tests or matching exact text), model-based grading (another AI scoring against a rubric), and human review.5 Each has a catch. Fixed rules are fast but miss good answers worded differently; an AI grader can be swayed, and OpenAI names “position bias (response order), verbosity bias (preferring longer responses)” among the problems.7
You do not need hundreds of cases to begin. Anthropic says “20-50 simple tasks drawn from real failures is a great start”, and its closing advice is blunt: “Read the transcripts!”5
Write down a small set of real tasks and what a good answer looks like before you change anything. Run the same set after every change, and read some of the answers yourself rather than trusting the score alone.
Questions people ask
What does “eval” mean in AI?
It is short for evaluation: a repeatable test that feeds an AI set inputs and grades the outputs. Anthropic’s definition is “give an AI an input, then apply grading logic to its output to measure success.”5
What are capability and regression evals?
Two jobs an eval can do. Anthropic says a capability eval asks “What can this agent do well?” and a regression eval asks “Does the agent still handle all the tasks it used to?”, catching things that used to work and broke.5
Do I need special software to run evals?
Not to start. In this site’s use, a spreadsheet of test questions and what a good answer must contain is enough for a small project. Vendor tools exist but change: OpenAI’s Evals platform was due to become read-only for existing users on 31 Oct 2026, a date that has now passed, and is scheduled to shut down on 30 Nov 2026. OpenAI’s pages still list both dates as scheduled and, as of 3 Nov 2026, do not say either has happened or changed.8 How to test a prompt shows a hand-run method.
Where to go next
- Testing by hand: how to test whether a prompt change actually helped.
- The public kind of test: what is a benchmark in AI?
- The bigger picture: how AI is tested.
- Trying to break it on purpose: what is red teaming?
- Why answers vary: what is temperature in AI?
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is an eval?
- A public leaderboard that ranks models
- A repeatable test that gives an AI set inputs and grades the outputs against your standards
- A setting that makes answers less random
Show the answer
Answer: A repeatable test that gives an AI set inputs and grades the outputs against your standards. An eval runs the same inputs each time and grades what comes out against criteria you chose.
2. How does an eval differ from a benchmark?
- An eval is usually built from your own tasks; a benchmark is a shared, standard test
- A benchmark is always private
- They are two names for exactly the same thing
Show the answer
Answer: An eval is usually built from your own tasks; a benchmark is a shared, standard test. Benchmarks compare models on standard tasks; evals check how well something works for your task.
3. How big does a first eval need to be?
- Thousands of cases before it is useful
- Exactly one test question
- A small set of real tasks, such as tasks drawn from real failures
Show the answer
Answer: A small set of real tasks, such as tasks drawn from real failures. Anthropic says a small set of simple tasks drawn from real failures is a good start.
Related terms: Benchmark · Coding agent · Embedding · Function calling · Human-in-the-loop · MCP (Model Context Protocol)
- 1Anthropic, Demystifying evals for AI agents, Anthropic Engineering, 9 Jan 2026, read at source 19 Oct 2026an eval is “a test for an AI system: give an AI an input, then apply grading logic to its output to measure success.” ↑
- 2OpenAI, Working with evals, OpenAI API documentation, read at source 19 Oct 2026“Evaluations (often called evals) test model outputs to ensure they meet style and content criteria that you specify.” ↑
- 3IBM, LLM evaluation: Why testing AI models matters, undated, read at source 1 Oct 2026 (as cited on benchmark meaning): LLM benchmarks “are standardized datasets or tasks that are used to analyze results and guide the evaluation process.”↑
- 4OpenAI, Evaluation best practices, read at source 18 Sep 2026 (as cited on how to test a prompt): “Generative AI is variable. Models sometimes produce different output from the same input, which makes traditional software testing methods insufficient for AI architectures.”↑
- 5Anthropic, Demystifying evals for AI agents, 9 Jan 2026, read at source 19 Oct 2026without evals it is easy to “get stuck in reactive loops—catching issues only in production.” ↑adescribes code-based, model-based and human graders, with the strengths and weaknesses of each. ↑b“20-50 simple tasks drawn from real failures is a great start” and “Read the transcripts!” ↑c“give an AI an input, then apply grading logic to its output to measure success.” ↑dcapability evals ask “What can this agent do well?”; regression evals ask “Does the agent still handle all the tasks it used to?” ↑e
- 6Anthropic, Define success criteria and build evaluations, read at source 18 Sep 2026 (as cited on how to test a prompt): “Building a successful LLM-based application starts with clearly defining your success criteria and then designing evaluations to measure performance against them. This cycle is central to prompt engineering.”↑
- 7OpenAI, Evaluation best practices, read at source 18 Sep 2026 (as cited on how to test a prompt), challenges of LLM-as-a-judge: “Position bias (response order), verbosity bias (preferring longer responses)”.↑
- 8OpenAI, Working with evals, read at source 19 Oct 2026 and again 3 Nov 2026on both dates the page says evals become read-only for existing users on “October 31, 2026” and the platform is scheduled to shut down on “November 30, 2026”; the page suggests Datasets for new work. On 3 Nov 2026 the page was still worded as scheduled (it did not say the read-only date had happened), and OpenAI’s deprecations page, read the same day, still listed “Existing evals become read-only” for 31 Oct 2026 and the Evals dashboard and API as “scheduled to shut down” on 30 Nov 2026, with no change noted. ↑
OpenAI, Evaluation best practices, read at source 18 Sep 2026 (as cited on how to test a prompt).
IBM, LLM evaluation: Why testing AI models matters, read at source 1 Oct 2026 (as cited on benchmark meaning).
The comparisons and examples on this page (the driving test, the support bot, the spreadsheet) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK