What is AGI — and how would we know?
A plain-language guide to the question behind this site: what artificial general intelligence would be, why nobody can measure it with one number, and which evidence is worth looking at instead.
- Step 1 of 6
What AGI means
Today’s AI systems are remarkably capable at many individual things: writing, programming, summarising, analysing. AGI — artificial general intelligence — describes a different threshold: a system that can do most intellectual work as well as a person, across fields, including work it has never seen.
- Solves the tasks it is pointed atWork out for itself what needs doing
- Is strong in some fields, weak in othersBe dependable across almost every field
- Needs people to check and correct itWork on its own for days and notice its own mistakes
There is no agreed definition and no test that would prove AGI. This comparison is a way to think about it, not a measurement.
- Step 2 of 6
Why there is no single number
There is no single agreed definition. Some describe AGI by economic value — a system that can do most paid knowledge work. Others describe levels of performance and breadth. Others focus on independence: acting on its own over long periods. These views overlap, but they are not the same, and none of them comes with an accepted test.
So when someone says “AGI is 80% done”, ask: by which definition, measured how? An honest observer shows evidence, not a percentage.
- Step 3 of 6
How we would recognise it
No single result would settle it. What would count is evidence piling up from different directions and pointing the same way. agi.jetzt separates three kinds:
- Public benchmarksStandardised tests run by researchers and labs: reasoning, mathematics, programming, tool use, long tasks. Useful, but each measures one narrow thing, and unrelated scores cannot be averaged into “intelligence”.Each result is recorded with its source, the model it was reported for and its date. Values are shown as published — never averaged into a score.
- Our own tests — the PrüfstandWe give models a written brief and publish what they built. So far this is a pilot: one kind of task (a single HTML artifact from a brief), screenshots, and static checks of the file — no check yet renders the artifact or tries its function.A narrow slice — not “the AGI test”. The pilot ran earlier models; current top-tier models are not in it yet.
- What reaches real productsA capability matters differently once it is switched on in products people use. We record which models a checked source names as usable in which product, and which capabilities are on record.Each statement has its source. Where there is none, the site says “no statement” — never “not available”.
- Step 4 of 6
Ten fields instead of one number
To make “how far are we?” answerable at all, we split it into ten capability fields. The fields are a framework: our way of ordering the question, not an agreed scientific scale.
How well a field is evidenced differs from field to field: “Reasoning” has 133 reported benchmark values on record, “Learning and memory” has 2.
Well evidenced only means: results with a source are on record for this field. It does not mean:
- Capability. A lot of evidence does not say that models are good in this field.
- Intelligence. The amount of evidence does not measure how intelligent a model is.
- “Percent to AGI”. Even with evidence on record for all ten fields, AGI has not been reached.
Our own tests and public benchmarks stay separate: what agi.jetzt tested itself in the Prüfstand is never combined with results that others published. So far there are own pilot tests for one field only: “Building software and artifacts”.
See the full field table ↓ - Step 5 of 6
What agi.jetzt shows today — and what it does not
A scored AGI methodology does not exist on this site. agi.jetzt publishes observable signals with a source only.
- shownModel lifecycleAnnouncement, release, phase-out, retirement — confirmed dates.
- shownAvailability at workWhich product offers which model, with its history.
- shownCapabilities and integrationsConfirmed product features and connections.
- shownThe current frontierThe top-tier models of OpenAI, Anthropic, Google and SpaceXAI as the providers themselves position them — with source and check date.
- shownPublic benchmark resultsResults published by others, per model and benchmark, each with its source.
- shownPrüfstand pilot runsPublished runs from July to October 2026 with screenshots and static checks. No render or function checks, no judge evaluations.
- not shownAn overall AGI valueNo scored methodology exists, so none is published.
- Step 6 of 6
How a fact gets onto this site
- SourceA change is found in an official source: a provider’s documentation, changelog or announcement.
- Cross-checkIt is compared with the source and with what is already on record. Contradictions are held back.
- RecordOnly then does it change what the site shows — with the date it happened and the date we recorded it.
- CorrectIf something turns out wrong, a correction is published. Earlier entries are marked, not silently rewritten.
What you never see here: proposals not yet decided, and anything that rests on a single claim without a source. Statements we hold but have not yet checked at the source stay visible — marked “not verified”.
The ten fields in detail
For each field the table shows what agi.jetzt has on record right now: own tests, capabilities in products and public benchmark results — each kept separate, across all recorded models, older ones included. How many of the current frontier models have evidence in a field is shown on AGI Progress.
| Capability field | Our own tests Prüfstand | In products confirmed capabilities | Public benchmarks external | Evidence on record |
|---|---|---|---|---|
ReasoningThinking through problems nobody has worked out for it before. | none on record | none on record | 100 model × benchmark pairs · 5 families · 133 reported values | Public benchmarksresults published by others, each with its source |
Learning and memoryTaking in something new and using it later — without being retrained. | none on record | on record in 1 productMemory | 1 model × benchmark pair · 1 family · 2 reported values | Public benchmarksresults published by others, each with its sourceSeen in productsprovider-confirmed, not independently tested |
AutonomyPursuing a goal on its own instead of waiting for every step. | none on record | on record in 4 productsAgent mode | 53 model × benchmark pairs · 7 families · 58 reported values | Public benchmarksresults published by others, each with its sourceSeen in productsprovider-confirmed, not independently tested |
Long tasksStaying on one job for hours or days without losing the thread. | none on record | on record in 5 productsDeep research | 19 model × benchmark pairs · 4 families · 19 reported values | Public benchmarksresults published by others, each with its sourceSeen in productsprovider-confirmed, not independently tested |
Tools and computersUsing search, code, files and programs the way a person at a computer does. | none on record | on record in 2 productsData analysis, Web search | 48 model × benchmark pairs · 6 families · 53 reported values | Public benchmarksresults published by others, each with its sourceSeen in productsprovider-confirmed, not independently tested |
Building software and artifactsTurning a description into something that actually works. | 6 runs with a decided check · 6 models4 of them (4 models) through an agent harness — model + harness | on record in 3 productsCoding agent | 53 model × benchmark pairs · 8 families · 60 reported values | Own pilot testspilot, July–October 2026: earlier models, one kind of task, static checksPublic benchmarksresults published by others, each with its sourceSeen in productsprovider-confirmed, not independently tested |
MultimodalityUnderstanding and producing text, images, sound and video together. | none on record | none on record | 55 model × benchmark pairs · 8 families · 68 reported values | Public benchmarksresults published by others, each with its source |
ReliabilityGetting the same thing right every time, even unsupervised. | none on record | none on record | 8 model × benchmark pairs · 2 families · 8 reported values | Public benchmarksresults published by others, each with its source |
AdaptabilityCoping with the unfamiliar when the situation changes. | none on record | none on record | 24 model × benchmark pairs · 1 family · 29 reported values | Public benchmarksresults published by others, each with its source |
GeneralityAll of this in almost any field — not only where it was trained. | none on record | none on record | 16 model × benchmark pairs · 1 family · 40 reported values | Public benchmarksresults published by others, each with its source |
Own pilot testsown pilot runs with a decided checkPublic benchmarkspublished results with a sourceSeen in productson record in a productOpennothing on record
The ten fields are agi.jetzt’s way of structuring the question — not an agreed scientific scale. A mark means evidence is on record here; it does not mean the field is solved, and it says nothing about how capable a model is. “Open” means agi.jetzt has no evidence on record yet, not that none exists in the world.
Product capabilities describe the product around a model (its tools, memory, agents) — they are not a measurement of the model itself.
Our own tests reach each model through a recorded runtime — an agent CLI such as Codex CLI or Claude Code, or a local inference server. A result describes the model in that setup, not the bare model; each run names its runtime.