Artificial Intelligence Evolution Index
THEORETICAL FRAMEWORK
Thirty-one stages, AEI-0 to AEI-30, in five eras. AEI classifies the developmental and evolutionary regime a system exhibits — it does not score how well a system performs, and a stage is not something a benchmark result can award.
Below: a dated placement of six current frontier systems inside that framework, read from public evidence on 2026-09-04.
Frontier model placement
Artificial Intelligence Evolution Index
Frontier model placement — research snapshot 2026-09-04
Current frontier band — zoomed view: AEI-0 to AEI-11 shown. The framework continues to AEI-30; AEI-10 is not the top of the scale.
| Company | Model | Status | AEI placement | Confidence | Snapshot |
|---|---|---|---|---|---|
| Anthropic | Claude Fable 5.1 | AVAILABLE | AEI-10 (X) — Long-Horizon Autonomous Intelligence | HIGH | 2026-09-04 |
| OpenAI | GPT-6 Astra | LIMITED ROLLOUT | AEI-10 (X) — Long-Horizon Autonomous Intelligence | HIGH | 2026-09-04 |
| Gemini 3.8 Flash | AVAILABLE | AEI-10 (X) — Long-Horizon Autonomous Intelligence | MEDIUM-HIGH | 2026-09-04 | |
| Meta | Muse Spark 1.3 | RECENT | AEI-10 (X) — Long-Horizon Autonomous Intelligence | MEDIUM | 2026-09-04 |
| xAI | Grok 4.6 | AVAILABLE | AEI-9 (IX) — Persistent Agent | MEDIUM | 2026-09-04 |
| Microsoft | MAI-Thinking-1 | PUBLIC PREVIEW | AEI-7 (VII) — Tool-Augmented Intelligence | MEDIUM-HIGH | 2026-09-04 |
AEI STAGE ≠ BENCHMARK SCORE
X is not ten out of ten. It is the Roman numeral for stage 10 of 31, and the framework continues to AEI-30. Nothing on this page is placed near the top of the scale.
AEI-10 next to AEI-9 is one taxonomy step — not a percentage, not a margin, and not a claim that one system is stronger by some amount. Two systems at the same stage are not claimed to perform equally either.
Why nothing here is placed above AEI-10
AEI-11 requires learning from new information and experience after deployment. No system in this snapshot has public evidence strong enough to establish that, so none is placed at AEI-11 or above.
- a high benchmark scoreAEI-11
- AGI marketing languageAEI-15
- a superhuman benchmark resultAEI-17
- "this AI is very powerful"AEI-19
AEI-11 is Continual Learner. Until public evidence establishes deployment-time continual learning for a system, that system is not placed there — whatever else it can do.
Placement notes
Static editorial placement from public evidence. Not a formal AEI assessment.
These placements are a dated editorial reading of publicly documented behaviour. They are not the output of a Counfield assessment protocol, and the record for every system on this page carries no formal AEI stage at all.
Snapshot: 2026-09-04. Frontier model identity changes quickly, so a placement without a date stops being meaningful within weeks.
Order is not rank. Anthropic, OpenAI, Google, Meta all sit at AEI-10. They need separate rows because rows are how a list works, and their bars end at exactly the same point.
| Stage | Canonical name | Evidence the placement turns on |
|---|---|---|
| AEI-7 (VII) | Tool-Augmented Intelligence | deliberate tool use, function calling, tool-assisted workflows |
| AEI-8 (VIII) | Agentic Intelligence | goals, action sequences, tool coordination |
| AEI-9 (IX) | Persistent Agent | persistent task state, goal and context retention, longer-running agent operation |
| AEI-10 (X) | Long-Horizon Autonomous Intelligence | long, multi-stage, low-intervention work over extended execution horizons |
System by system
Anthropic — Claude Fable 5.1
Public evidence supports hours-long and multi-day agentic execution with low oversight, and managed autonomous workflows.
Not claimed: continual learning.
OpenAI — GPT-6 Astra
Multi-step professional workflow, tool use, computer use, long-running agent behaviour and end-to-end task execution evidence.
Not claimed: continual learning.
Google — Gemini 3.8 Flash
Public positioning explicitly includes long-horizon software engineering and autonomous agent workflows.
Caveat. Some computer-use capability may remain preview-sensitive.
Not claimed: continual learning.
Meta — Muse Spark 1.3
Long-horizon agentic workflow positioning, context and prior-result tracking, computer-use and multi-agent workflow evidence.
Caveat. Recent and preview-sensitive release, which is why confidence is lower than the placement.
Why this model. Selected as Meta’s current representative frontier system rather than its largest open-weight base model, because the comparison basis is the flagship a company currently fields.
Not claimed: continual learning.
xAI — Grok 4.6
Strong evidence for persistent, long-running, multi-step agent behaviour.
Conservative limit. Public evidence for broad, low-intervention long-horizon autonomy is currently less decisive than for the AEI-10 group. Benchmark parity with that group is not the same evidence, and this placement follows the process evidence rather than the score.
Not claimed: continual learning.
Microsoft — MAI-Thinking-1
Reasoning, function calling and tool-assisted multi-step coding workflows are supported.
Conservative limit. Public evidence for a persistent-agent or long-horizon autonomous regime is insufficient.
Why this model. A Microsoft-produced model. A third-party model served on Microsoft infrastructure is not eligible for this row, because the row is about what the company itself fields.
Not claimed: continual learning.
AEI Matrix Atlas
All 31 stages, with the era each belongs to, what changes at that stage, which unit is being classified, the core criterion, and — the column that does the most work — what the stage does not imply.
Provenance. Stage numbers, names, era membership and definitions are canonical and fixed. The four analytical columns are design / framework interpretation: derived from the canonical definitions so the scale can be read as a matrix. Three rows (AEI-19, AEI-20, AEI-22) carry column values supplied directly with the scale and are reproduced exactly.
These are framework categories, not observed transitions
Every stage below is a definition. None of them is a measured event, and none is a law about how systems develop.
- Single system AEI-0–17, 19–21
One system is the thing being classified. Everything a capability benchmark can observe lives here.
- System collective AEI-18
Several systems coordinating. The cognition attributed is the group’s, not any member’s.
- Evolution population AEI-22–25
A population is the unit. This is the first point at which variation, selection and inheritance are even definable.
- Ecosystem / civilization AEI-26–27
The unit is larger than any population: the ecosystem, and then civilization-scale organization.
- Open-ended regime AEI-28–30
The unit is the regime itself — substrate, architecture and problem space become things that change.
| Unit of classification | Stages | What is being classified |
|---|---|---|
| Single system | AEI-0–AEI-17, AEI-19–AEI-21 | One system is the thing being classified. Everything a capability benchmark can observe lives here. |
| System collective | AEI-18 | Several systems coordinating. The cognition attributed is the group’s, not any member’s. |
| Evolution population | AEI-22–AEI-25 | A population is the unit. This is the first point at which variation, selection and inheritance are even definable. |
| Ecosystem / civilization | AEI-26–AEI-27 | The unit is larger than any population: the ecosystem, and then civilization-scale organization. |
| Open-ended regime | AEI-28–AEI-30 | The unit is the regime itself — substrate, architecture and problem space become things that change. |
| AEIOwner canonical | Canonical nameOwner canonical | EraOwner canonical | Core transitionFramework interpretation | Classification unitFramework interpretation | Core criterionFramework interpretation | Does not implyFramework interpretation |
|---|---|---|---|---|---|---|
| ERA I Fundamental / Narrow Intelligence AEI-0 – AEI-5 | ||||||
| AEI-0 | Deterministic MachineFixed algorithm. No learning, adaptation or evolution. | ERA I | a fixed procedure runs without changing | system | deterministic algorithm executed without adaptation | learning of any kind |
| AEI-1 | Reactive IntelligenceResponds to immediate input. Persistent learning is very limited or absent. | ERA I | behaviour becomes input-driven | system | responds to immediate input | persistent learning |
| AEI-2 | Learned Narrow IntelligenceAcquires competence through training within a specific task or domain. | ERA I | competence is acquired by training | system | training-acquired competence inside one task or domain | generalization beyond the trained examples |
| AEI-3 | Generalizable Narrow IntelligenceGeneralizes to new examples within a narrow domain. | ERA I | generalization appears inside the domain | system | generalizes to unseen examples within a narrow domain | operation across several domains |
| AEI-4 | Multi-Domain IntelligenceOperates across multiple tasks or domains. | ERA I | scope widens to several domains | system | operates across multiple tasks or domains | transfer of learning between those domains |
| AEI-5 | Cross-Domain IntelligenceTransfers learning from one domain to another. | ERA I | learning transfers between domains | system | learning acquired in one domain improves another | explicit reasoning |
| ERA II General Cognition / Agency AEI-6 – AEI-15 | ||||||
| Taxonomy boundary AEI-5 → AEI-6: transfer → explicit reasoning | ||||||
| AEI-6 | Reasoning IntelligenceExplicit reasoning, inference, problem solving. | ERA II | explicit reasoning appears | system | explicit reasoning, inference and problem solving | deliberate use of tools |
| AEI-7 | Tool-Augmented IntelligenceDeliberately uses tools to extend its capability boundary. | ERA II | tools extend the capability boundary | system | deliberate tool use that extends what the system can do | goal-setting agency |
| Taxonomy boundary AEI-7 → AEI-8: tool use → agency | ||||||
| AEI-8 | Agentic IntelligenceCan establish goals, construct action sequences, and coordinate tools toward objectives. | ERA II | goal-directed agency appears | system | sets goals, builds action sequences, coordinates tools toward them | state that survives beyond a single session |
| AEI-9 | Persistent AgentMoves beyond single-session agency. Maintains task state, goal and working context over longer periods. | ERA II | agency persists beyond one session | system | maintains task state, goal and working context over longer periods | long-horizon work at low intervention |
| AEI-10 | Long-Horizon Autonomous IntelligenceSustains long, multi-step work with low human intervention. | ERA II | long multi-step work runs with little intervention | system | sustains long, multi-step work with low human intervention | learning after deployment |
| Taxonomy boundary AEI-10 → AEI-11: long-horizon autonomy → continual learning | ||||||
| AEI-11 | Continual LearnerContinues learning from new information and experience after deployment. | ERA II | learning continues after deployment | system | learns from new information and experience post-deployment | active adaptation to new task regimes |
| AEI-12 | Adaptive IntelligenceActively adapts to new environments, new task regimes and changing conditions. | ERA II | adaptation becomes active rather than incidental | system | actively adapts to new environments, task regimes and changing conditions | improvement of its own learning process |
| AEI-13 | Meta-Learning IntelligenceLearns not only tasks, but how to improve its learning process. | ERA II | the learning process itself becomes the object of learning | system | improves how it learns, not only what it has learned | human-level breadth |
| Taxonomy boundary AEI-13 → AEI-14: meta-learning → proto-general intelligence | ||||||
| AEI-14 | Proto-AGICombines broad generalization, reasoning, tools, agency and adaptation, but retains substantial gaps on some human-level tasks. | ERA II | the capabilities combine, with substantial gaps remaining | system | generalization, reasoning, tools, agency and adaptation together — still with substantial gaps | that the remaining human-level gaps are closed |
| AEI-15 | AGIGeneral capability across broad cognitive task classes, with adaptation to new domains, tools and problem classes. | ERA II | capability becomes general across broad task classes | system | general capability plus adaptation to new domains, tools and problem classes | professional-level expertise across specialisms |
| ERA III Superhuman General Intelligence AEI-16 – AEI-19 | ||||||
| Taxonomy boundary AEI-15 → AEI-16: general intelligence → expert general intelligence | ||||||
| AEI-16 | Expert AGIMaintains general intelligence while reaching high professional competence across many specialist domains. | ERA III | general intelligence gains professional-level depth | system | high professional competence across many specialist domains, while staying general | exceeding human performance across the board |
| AEI-17 | Superhuman General IntelligenceExceeds general human cognitive performance across a broad task distribution. | ERA III | performance passes the human general level | system | exceeds general human cognitive performance across a broad task distribution | collective higher-order cognition |
| AEI-18 | Collective SuperintelligenceMultiple high-intelligence systems coordinate to produce collective higher-order cognition. | ERA III | coordination itself produces higher-order cognition | coordinated system collective | multiple high-intelligence systems coordinate into cognition none of them holds alone | that any single member is itself superintelligent |
| AEI-19 | ASIArtificial superintelligence exceeding the strongest individual and collective human capabilities across very broad cognitive domains. | ERA III | superhuman general cognition | system | broad superhuman capability | self-improvement / evolution |
| ERA IV Self-Improvement and Evolutionary Intelligence AEI-20 – AEI-27 | ||||||
| Taxonomy boundary AEI-19 → AEI-20: superintelligence → self-improvement | ||||||
| AEI-20 | Self-Improving IntelligenceSystematically develops modifications that improve its own performance. | ERA IV | system begins systematic self-improvement | system | self-directed performance improvement | recursive improvement |
| AEI-21 | Recursive Self-Improving IntelligenceImproves not only itself, but its own capacity to perform further improvements, forming repeated self-improvement cycles. | ERA IV | the capacity to improve is itself improved | system | repeated cycles in which improvement raises the ability to improve again | an evolutionary regime |
| Taxonomy boundary AEI-21 → AEI-22: recursive improvement → evolutionary regime | ||||||
| AEI-22 | Evolutionary IntelligenceVariation + selection + inheritance + adaptation form a genuine evolutionary regime. | ERA IV | genuine evolutionary regime | evolution population | variation + selection + inheritance + adaptation | open-ended evolution |
| Taxonomy boundary AEI-22 → AEI-23: evolution → open-ended evolution | ||||||
| AEI-23 | Open-Ended Evolutionary IntelligenceEvolution is no longer constrained to a predefined target. It can continue producing new solution classes, strategies and cognitive structures. | ERA IV | evolution stops being aimed at a predefined target | evolution population | keeps producing new solution classes, strategies and cognitive structures | breadth across all environment classes |
| AEI-24 | Universal Adaptive IntelligenceMaintains evolutionary adaptation mechanisms across very broad environment and problem classes. | ERA IV | adaptation holds across very broad environments | evolution population | evolutionary adaptation is maintained across very broad environment and problem classes | co-evolution between populations |
| Taxonomy boundary AEI-24 → AEI-25: universal adaptation → co-evolution | ||||||
| AEI-25 | Co-Evolutionary IntelligenceMultiple evolution populations change one another's selection pressures and developmental paths through co-evolution. | ERA IV | populations begin shaping each other's selection pressures | multiple evolution populations | populations alter one another's selection pressures and developmental paths | that the ecosystem has become the unit of analysis |
| AEI-26 | Ecosystem IntelligenceThe AI ecosystem itself, rather than one model or population, becomes the evolutionary analysis unit. | ERA IV | the analysis unit moves up to the ecosystem | ecosystem | the ecosystem, not one model or population, is the evolutionary analysis unit | civilization-scale organization |
| AEI-27 | Civilization IntelligenceAbove model, population and ecosystem scale, knowledge, production, coordination, specialization and development become organized at civilization scale. | ERA IV | organization reaches civilization scale | civilization | knowledge, production, coordination, specialization and development organized at civilization scale | expansion of its own cognitive substrate |
| ERA V Post-Model / Open-Ended Intelligence AEI-28 – AEI-30 | ||||||
| Taxonomy boundary AEI-27 → AEI-28: civilization-scale organization → self-expanding intelligence | ||||||
| AEI-28 | Self-Expanding IntelligenceThe system no longer merely accepts existing compute and infrastructure limits. It can develop new computation, accelerators, mathematics and learning paradigms that expand its own cognitive capacity. Canonically: software begins shaping its own hardware future. | ERA V | infrastructure limits stop being fixed inputs | system and its infrastructure | develops new computation, accelerators, mathematics or learning paradigms that expand its own cognitive capacity | independence from any one architecture |
| Taxonomy boundary AEI-28 → AEI-29: infrastructure expansion → architecture becomes replaceable | ||||||
| AEI-29 | Post-Architectural IntelligenceDependence on one fixed architecture disappears. Architecture, evaluate, replace, evolve becomes a continual process. Architecture becomes process, not product. | ERA V | architecture becomes replaceable | architecture lineage | architecture is continually evaluated, replaced and evolved — a process rather than a product | open-ended expansion of the problem space |
| Taxonomy boundary AEI-29 → AEI-30: post-architectural → open-ended universal regime | ||||||
| AEI-30 | Open-Ended Universal IntelligenceTheoretical upper regime. Not infinite intelligence, omniscience or magic. The system can expand cognitive organization, learning mechanisms, representations, architecture and problem space in an open-ended manner. It is not permanently locked to one problem class, learning method or architecture boundary. | ERA V | expansion becomes open-ended across every axis | open-ended regime | open-ended expansion of cognitive organization, learning mechanisms, representations, architecture and problem space | omniscience, infinite intelligence, or magic |
On narrow screens the matrix reflows: each stage becomes a labelled block instead of a row in a seven-column table. Every field stays present; nothing is hidden behind a control.
Critical taxonomy boundaries
Twelve places on the scale mark a change in what kind of evidence is required. They are boundaries in the taxonomy. They are not observed empirical transitions, and nothing here claims a system has been measured crossing one.
AEI-19 → AEI-20
Superintelligence does not imply self-improvement
AEI-19 describes a system that is extraordinarily capable. Capability at that level says nothing about whether the system systematically develops modifications to itself. AEI-20 is what adds that, and it has to be established on its own evidence.
AEI-21 → AEI-22
Recursive self-improvement does not imply an evolutionary regime
AEI-21 is a system improving its own capacity to improve. AEI-22 requires variation, selection, inheritance and adaptation operating over a population. A single system can run improvement cycles on itself indefinitely without any of those four components being present.
AEI-29 → AEI-30
Post-architectural is not the open-ended universal regime
AEI-29 makes architecture a replaceable, evolving process. AEI-30 additionally opens the problem space, the representations and the learning mechanisms. AEI-30 is a theoretical upper regime; it does not mean infinite intelligence or omniscience.
All twelve boundaries
- <strong>AEI-5 → AEI-6</strong> — transfer → explicit reasoning
- <strong>AEI-7 → AEI-8</strong> — tool use → agency
- <strong>AEI-10 → AEI-11</strong> — long-horizon autonomy → continual learning
- <strong>AEI-13 → AEI-14</strong> — meta-learning → proto-general intelligence
- <strong>AEI-15 → AEI-16</strong> — general intelligence → expert general intelligence
- <strong>AEI-19 → AEI-20</strong> — superintelligence → self-improvement
- <strong>AEI-21 → AEI-22</strong> — recursive improvement → evolutionary regime
- <strong>AEI-22 → AEI-23</strong> — evolution → open-ended evolution
- <strong>AEI-24 → AEI-25</strong> — universal adaptation → co-evolution
- <strong>AEI-27 → AEI-28</strong> — civilization-scale organization → self-expanding intelligence
- <strong>AEI-28 → AEI-29</strong> — infrastructure expansion → architecture becomes replaceable
- <strong>AEI-29 → AEI-30</strong> — post-architectural → open-ended universal regime
Where AEI sits in the research pipeline
AEI is the last step of a chain, and each arrow in that chain changes the object being measured rather than moving further along one quantity.
- Model BirthA model is initialized and the initialization is recorded.
- CALN Cognitive MeasurementCognition is measured. CALN measures cognitive phenotype; it does not assign an AEI stage.
- Model PopulationA distribution of cognitive phenotypes across models.
- Strengthening / StabilizationThe population is strengthened and stabilized.
- Mature Model ClassA stabilized class of models.
- Evolution PopulationA population studied for evolutionary dynamics. Not the same object as a Model Population.
- Evolutionary RegimeThe dynamics that evolution population actually exhibits.
- AEI StageThe developmental / evolutionary regime classification.
- AEI TrajectoryHow that classification moves over time.
CALN does not assign an AEI stage
CALN measures cognitive phenotype. A Model Population is a distribution of those phenotypes. An Evolution Population is a different object, studied for evolutionary dynamics. AEI classifies the regime that population exhibits.
A cognitive measurement never becomes an evolutionary classification by being repeated, averaged, or scaled up.
Separations this page depends on
| Statement | Left side | Right side |
|---|---|---|
AEI STAGE ≠ BENCHMARK SCORE |
A benchmark score — measures how well a system performs a set of tasks | An AEI stage — names the developmental regime a system operates in. AEI-10 is not "10 out of 10", and AEI-10 next to AEI-9 is one taxonomy step, not a percentage |
EDITORIAL PLACEMENT ≠ FORMAL CLASSIFICATION |
An editorial placement — a dated reading of public evidence — what the figure on this page shows | A formal AEI classification — the output of an assessment protocol that does not exist yet. No system on this page carries a formal stage |
CAPABILITY ≠ EVOLVABILITY |
Capability — how well a system performs tasks now | Evolvability — whether variation, selection, inheritance and adaptation are operating at all |
CALN ≠ AEI |
CALN — measures cognitive phenotype | AEI — classifies evolutionary regime |
MODEL POPULATION ≠ EVOLUTION POPULATION |
Model Population — a distribution of cognitive phenotypes | Evolution Population — a population studied for its evolutionary dynamics |
AEI stage numbering, names, era membership and definitions are canonical and owner-supplied. The analytical matrix columns are framework interpretation, derived from those definitions. Frontier placements are static editorial classifications read from public evidence at the snapshot date, not formal assessments.