evals
Tests that measure how well an AI model or product performs the tasks its builders care about, usually graded by code, people or another model, and rerun after changes.
- Real customer questions, each graded against an approved answer
- A check that every SQL query a model writes actually runs
- A second model scoring chatbot replies against a written rubric
- A test of whether a model can help with a cyberattack
- Short for
- Evaluations
- A clipped plural; Google’s glossary gives it as short for LLM evaluations.
- Current meaning
- 2023
- OpenAI’s open-source Evals lets anyone test models on their own use cases.
- Opposite
- Vibe check
- Trying a few prompts by hand and judging the answers on impressions.
In plain words
In AI, evals are tests that measure how well a model or AI product does its job, run on the team’s own tasks and repeated after each change.
The word is short for evaluations.1 Each eval is a set of test cases: realistic inputs, each paired with a way to judge the answer. A team runs the whole set, records a score and runs it again after every change to a prompt, a model or the data. OpenAI has said it uses its own evals to find weaknesses in its models and to catch regressions, where a change makes something worse.2
Evals matter because a large language model can answer the same question differently from one run to the next. Informal testing, sometimes called a vibe check, gives impressions rather than measurements. The engineer Hamel Husain writes that unsuccessful AI products almost always share one root cause: a failure to build systems for evaluating them. By 2025 his course on evals, taught with Shreya Shankar, had trained more than 2,000 engineers and product managers.3
The word also covers safety testing. AI labs and government bodies run evals that probe new models for dangerous capabilities. Google DeepMind tested its Gemini 1.0 models in four areas, including persuasion and cyber-security, and found early warning signs but no strong dangerous capabilities.4 In 2024 the UK’s AI Safety Institute released its evals software, Inspect, as open source. A newer worry is that frontier models can often tell when they are being tested.
Evals are often confused with a benchmark. A benchmark is a fixed, public test that many developers run so their models can be compared. Evals also include tests a team builds for its own product or for a specific risk, often kept private. A public benchmark can be one eval among many. Some practitioners now complain that the word is used for too many different things.
How it works
An eval starts with test cases. Each case is an input the system will meet in real use, such as a customer email. Many cases also carry a reference answer or a list of criteria.
The system answers every case, and graders score the answers. Anthropic describes three kinds.5 Code-based graders check what a program can verify, such as an exact match. Human graders, often experts in the subject, judge what code cannot. Model-based graders give a second model a rubric and ask it to score each answer, a method known as LLM-as-a-judge. In a 2023 study, GPT-4 used this way agreed with human preferences over 80% of the time, as often as two people agree.6
A model’s output varies from run to run, so each case may be run several times and the results reported as pass rates. Anthropic separates two kinds of eval. Capability evals target what the system cannot yet do well, so they start with low pass rates. Regression evals check that it still does what it used to, and should pass almost every time.5 The whole set runs again after every change, as automated tests do in ordinary software.
Origin and first known use
A clipped plural of evaluations. Google’s machine learning glossary treats evals as mainly short for LLM evaluations, though eval can stand for any evaluation. The earliest AI use found dates from 2022, when the Alignment Research Center named its model-testing team ARC Evals. Anthropic released model-written evals the same year. OpenAI’s open-source Evals framework, released with GPT-4 in March 2023, carried the word to developers testing their own products. By 2025 a podcast for product managers was calling evals the hottest new skill for product builders.
- 2022 The Alignment Research Center starts ARC Evals, a team for independent evaluations of the most advanced AI models. It later tests GPT-4 and Claude. Earliest AI use found.Beth Barnes, “ARC Evals is spinning out from ARC”, METR blog, 19 September 2023 metr.org →: source for 2022
- 2023 OpenAI open-sources Evals with GPT-4, letting anyone write custom evals for their own use cases, even private ones. The current sense starts here.OpenAI, “Evals”, open-source repository on GitHub, released with GPT-4 in March 2023 github.com →: source for 2023
- 2023 Zheng and colleagues examine using a strong model such as GPT-4 to grade chatbot answers, a method they call LLM-as-a-judge.Zheng et al., “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena”, NeurIPS 2023 Datasets and Benchmarks Track arXiv 2306.05685 →: source for 2023
- 2024 Hamel Husain’s post “Your AI Product Needs Evals” sets out three levels of evaluation for AI products: unit tests, human and model review, and A/B tests.Hamel Husain, “Your AI Product Needs Evals”, hamel.dev, 29 March 2024 hamel.dev →: source for 2024
- 2024 The UK’s AI Safety Institute releases Inspect as open source. The software tests specific capabilities of a model and produces a score.Department for Science, Innovation and Technology and AI Safety Institute, press release on the Inspect platform, GOV.UK, 10 May 2024 gov.uk →: source for 2024
- 2024 OpenAI adds Evals to its developer platform in beta, so developers can run custom evaluations of model performance on their own tasks.OpenAI, “Model Distillation in the API”, 1 October 2024 openai.com →: source for 2024
- 2025 Needham and colleagues find that frontier models can often tell evaluation transcripts from real use, which could make evaluations less reliable.Needham et al., “Large Language Models Often Know When They Are Being Evaluated”, arXiv, 28 May 2025 arXiv 2505.23836 →: source for 2025
In the wild
-
“Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.”
-
“We’re open-sourcing OpenAI Evals, our software framework for creating and running benchmarks for evaluating models like GPT‑4, while inspecting their performance sample by sample.”
OpenAI, GPT-4 research announcement openai.com →: OpenAI, GPT-4 research announcement, 14 Mar 2023Archived →: copy of OpenAI, GPT-4 research announcement, 14 Mar 2023 -
“…we introduce a programme of new “dangerous capability” evaluations and pilot them on Gemini 1.0 models.”
Phuong et al., Google DeepMind, arXiv preprint arXiv 2403.13793 →: Phuong et al., Google DeepMind, arXiv preprint, 20 Mar 2024Archived →: copy of Phuong et al., Google DeepMind, arXiv preprint, 20 Mar 2024 -
“I’ve found that unsuccessful products almost always share a common root cause: a failure to create robust evaluation systems.”
-
“AI Safety Institute releases new AI safety evaluations platform”
-
“…every time we add a new eval and we’re always evaluating for all the old things, we have hundreds of these evaluations…”
Dario Amodei, Anthropic CEO, on the Lex Fridman Podcast lexfridman.com →: Dario Amodei, Anthropic CEO, on the Lex Fridman Podcast, 11 Nov 2024Archived →: copy of Dario Amodei, Anthropic CEO, on the Lex Fridman Podcast, 11 Nov 2024 -
“If AI models can detect when they are being evaluated, the effectiveness of evaluations might be compromised.”
-
“An evaluation (“eval”) is a test for an AI system: give an AI an input, then apply grading logic to its output to measure success.”
-
“Now, the issue with the word “evals” is that people use it for all sorts of things…”
Aishwarya Naresh Reganti and Kiriti Badam, O’Reilly Radar oreilly.com →: Aishwarya Naresh Reganti and Kiriti Badam, O’Reilly Radar, 28 Jan 2026Archived →: copy of Aishwarya Naresh Reganti and Kiriti Badam, O’Reilly Radar, 28 Jan 2026
Not to be confused with
-
Benchmark
A benchmark is a fixed, public test that lets many models be compared. Evals are tests chosen for a particular model, product or risk, often private and rerun after each change. A benchmark can be one of them.
-
Red teaming
Red teaming is an open-ended hunt for failures, in which people or models try to make a system misbehave. Evals are fixed sets of cases, scored the same way each time so results can be compared.
-
eval (programming)
In programming, eval is a function in languages such as JavaScript and Python that runs a string of text as code. It has nothing to do with testing AI models.
Questions
What are evals in AI?
Evals, short for evaluations, are tests that measure how well an AI model or product does its job. Each eval is a set of test cases with a way to grade the answers. The grader can be a code check, a human reviewer or a second model. Teams rerun the same evals after every change to a prompt, a model or the data, to see whether the system got better or worse.
What is the difference between evals and benchmarks?
A benchmark is a fixed, public test that many developers run so their models can be compared. Evals are broader. They include tests a team builds for its own product or for a specific risk, often kept private and rerun after each change. A public benchmark can be one of a team’s evals. OpenAI’s open-source Evals framework supports both existing benchmarks and custom, private evals.
How are AI evals graded?
Evals typically combine three kinds of grader. Code-based graders check what a program can verify, such as an exact match. Human graders, often experts in the subject, judge what code cannot. Model-based graders, an approach called LLM-as-a-judge, ask a second model to score answers against a rubric. In a 2023 study, GPT-4 used as a judge agreed with human preferences over 80% of the time.
What are dangerous capability evals?
They are safety tests that check whether a model could help cause serious harm. Google DeepMind tested its Gemini 1.0 models for persuasion and deception, cyber-security, self-proliferation and self-reasoning. It found early warning signs but no strong dangerous capabilities. The UK’s AI Safety Institute has released Inspect, open-source software for running such tests. A 2025 study found that frontier models can often tell when they are being evaluated.
Why do AI products need evals?
A model can answer the same input differently from run to run, and a change that fixes one problem can quietly break another. Evals catch these regressions by rerunning the same cases after every change. The engineer Hamel Husain writes that unsuccessful AI products almost always share one root cause: a failure to build systems for evaluating them. OpenAI has said it uses its own evals to prevent regressions.
Watch
Cite this entry
“Evals.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Evals. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Evals.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/