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

Frontier model placement inside the AEI taxonomy, 2026-09-04 A horizontal row index of six frontier AI systems placed inside the AEI taxonomy at 2026-09-04. The axis runs from AEI-0 to AEI-11 and the framework continues to AEI-30 beyond the right edge. Anthropic Claude Fable 5.1 at AEI-10; OpenAI GPT-6 Astra at AEI-10; Google Gemini 3.8 Flash at AEI-10; Meta Muse Spark 1.3 at AEI-10; xAI Grok 4.6 at AEI-9; Microsoft MAI-Thinking-1 at AEI-7. Rows at the same stage end at the same point and are not ranked against each other. Bar length encodes ordinal stage placement, not a benchmark score. 0 I II III IV V VI VII VIII IX X XI → AEI-30 ANTHROPIC Claude Fable 5.1 AVAILABLE X AEI-10 OPENAI GPT-6 Astra LIMITED ROLLOUT X AEI-10 GOOGLE Gemini 3.8 Flash AVAILABLE X AEI-10 META Muse Spark 1.3 RECENT X AEI-10 XAI Grok 4.6 AVAILABLE IX AEI-9 MICROSOFT MAI-Thinking-1 PUBLIC PREVIEW VII AEI-7

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.

AEI placements are static editorial classifications based on publicly documented system behavior and independent evaluations available at the snapshot date. AEI is not a benchmark score. Capability does not imply continual learning, self-improvement, or evolvability. Bar length encodes ordinal stage placement, so the spacing between two stages is not a measured quantity, and four systems sharing AEI-10 are not ranked against one another.
Frontier model placement — research snapshot 2026-09-04. Same content as the figure above. Order is alphabetical within a stage and is not a ranking.
CompanyModelStatus AEI placementConfidenceSnapshot
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
Google 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.

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.

The stages this snapshot uses, and the kind of public evidence each placement turns on.
StageCanonical nameEvidence the placement turns on
AEI-7 (VII)Tool-Augmented Intelligencedeliberate tool use, function calling, tool-assisted workflows
AEI-8 (VIII)Agentic Intelligencegoals, action sequences, tool coordination
AEI-9 (IX)Persistent Agentpersistent task state, goal and context retention, longer-running agent operation
AEI-10 (X)Long-Horizon Autonomous Intelligencelong, multi-stage, low-intervention work over extended execution horizons

System by system

Anthropic — Claude Fable 5.1

AEI-10 Confidence: HIGH Formal AEI stage: none

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

AEI-10 Confidence: HIGH Formal AEI stage: none

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

AEI-10 Confidence: MEDIUM-HIGH Formal AEI stage: none

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

AEI-10 Confidence: MEDIUM Formal AEI stage: none

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

AEI-9 Confidence: MEDIUM Formal AEI stage: none

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

AEI-7 Confidence: MEDIUM-HIGH Formal AEI stage: none

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.

The AEI scale, coloured by what each stage classifies A horizontal index of the 31 AEI stages, from AEI-0 on the left to AEI-30 on the right, grouped into five era brackets. A band beneath shows the unit of classification at each stage: a single system for AEI-0 to AEI-17, a system collective at AEI-18, a single system again for AEI-19 to AEI-21, an evolution population for AEI-22 to AEI-25, ecosystem and civilization scale at AEI-26 and AEI-27, and an open-ended regime for AEI-28 to AEI-30. Twelve vertical ticks mark the taxonomy boundaries. The axis is a stage index, not time. ERA I ERA II ERA III ERA IV ERA V Single system 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
The AEI scale is not one quantity getting larger. The unit being classified changes as the scale rises, and the change is what separates capability from evolvability: a benchmark observes one system, while an evolutionary regime is a property of a population.
  • 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 by stage. Same content as the figure above.
Unit of classificationStagesWhat is being classified
Single systemAEI-0–AEI-17, AEI-19–AEI-21One system is the thing being classified. Everything a capability benchmark can observe lives here.
System collectiveAEI-18Several systems coordinating. The cognition attributed is the group’s, not any member’s.
Evolution populationAEI-22–AEI-25A population is the unit. This is the first point at which variation, selection and inheritance are even definable.
Ecosystem / civilizationAEI-26–AEI-27The unit is larger than any population: the ecosystem, and then civilization-scale organization.
Open-ended regimeAEI-28–AEI-30The unit is the regime itself — substrate, architecture and problem space become things that change.
AEI Matrix Atlas — all 31 stages. The first three columns are owner-canonical. The last four are framework interpretation: they are derived from the canonical definitions so the scale can be read as a matrix, and they are not owner text.
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

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.

  1. Model BirthA model is initialized and the initialization is recorded.
  2. CALN Cognitive MeasurementCognition is measured. CALN measures cognitive phenotype; it does not assign an AEI stage.
  3. Model PopulationA distribution of cognitive phenotypes across models.
  4. Strengthening / StabilizationThe population is strengthened and stabilized.
  5. Mature Model ClassA stabilized class of models.
  6. Evolution PopulationA population studied for evolutionary dynamics. Not the same object as a Model Population.
  7. Evolutionary RegimeThe dynamics that evolution population actually exhibits.
  8. AEI StageThe developmental / evolutionary regime classification.
  9. 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

Five separations this page depends on. Each row is two different questions about two different objects.
StatementLeft sideRight 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.