Updated · July 2026
How close are we to artificial general intelligence?When will AI learn almost any thinking task?
When will AI learn almost any thinking task?How close are we to artificial general intelligence?
Frontier models saturate most static benchmarks, yet reliable long-horizon autonomy, calibrated reasoning and a shared definition of general intelligence all remain open - so any single 'percent to AGI' is a guess dressed as a number.AI can already help with many kinds of thinking, but it still makes confident mistakes and struggles to carry a big job through on its own.
Recorded milestones
The transformer architecture
Benchmarks saturate, autonomy lags
Open tests · conceptual branches
- OPEN TEST01 ↗Learn many different kinds of workBroad cognitive competenceA general system should transfer what it knows to fresh problems, not need a new trick for every benchmark.Expert-level performance across diverse, contamination-resistant novel tasks
- OPEN TEST02 ↗Finish long jobs without hand-holdingLong-horizon autonomyThe measured span an agent can complete on its own is still counted in hours, not days.Reliably plan and finish multi-day digital projects with sparse supervision
- OPEN TEST03 ↗Stay reliable when things get weirdCalibrated reliabilityBeing capable is not enough if unusual inputs still produce confident, costly errors.Predictable, well-calibrated, corrigible behavior under distribution shift and adversarial pressure
The path shows Benchmarks saturate, autonomy lags. Open thresholds are separate tests in a conceptual map, not a forecast. No single percentage represents progress.
Where we standWhere we stand
The current stage, the evidence recorded, and the single threshold that gates the next stage.The stage reached, the evidence recorded, and the one test still blocking the next step.
What the evidence has shownWhat the evidence has shown
AI can already help with many kinds of thinking, but it still makes confident mistakes and struggles to carry a big job through on its own.Frontier models saturate most static benchmarks, yet reliable long-horizon autonomy, calibrated reasoning and a shared definition of general intelligence all remain open - so any single 'percent to AGI' is a guess dressed as a number.
The next testBlocking threshold
Learn many different kinds of workBroad cognitive competenceARC-AGI-2 and held-out expert evals resistant to training-data leakageARC-AGI-2 and held-out expert evals resistant to training-data leakage
What would count as proof?What would count as proof?
A clear test, the evidence behind it, and the limits of the claim.A clear test, the evidence behind it, and what it still leaves open.
Learn many different kinds of workBroad cognitive competence
A general system should transfer what it knows to fresh problems, not need a new trick for every benchmark.Expert-level performance across diverse, contamination-resistant novel tasks
Next test ARC-AGI-2 and held-out expert evals resistant to training-data leakage
Still to demonstrateWhat this does not establishA separate system can still have different conditions and results.The named measurement has no wider implication until the supporting system test is also shown.
View every evidence record
Open test
Learn many different kinds of workBroad cognitive competence
A general system should transfer what it knows to fresh problems, not need a new trick for every benchmark.Expert-level performance across diverse, contamination-resistant novel tasks
Next test: ARC-AGI-2 and held-out expert evals resistant to training-data leakage
Open test
Finish long jobs without hand-holdingLong-horizon autonomy
The measured span an agent can complete on its own is still counted in hours, not days.Reliably plan and finish multi-day digital projects with sparse supervision
Next test: METR 50%-task time horizon extending from hours to multi-day tasks
Open test
Stay reliable when things get weirdCalibrated reliability
Being capable is not enough if unusual inputs still produce confident, costly errors.Predictable, well-calibrated, corrigible behavior under distribution shift and adversarial pressure
Next test: Pre-registered external safety and calibration evaluations on frontier deployments
The record behind the verdictThe record behind the verdict
Major milestones and smaller context events share one source-linked record. Scheduled targets are marked separately from recorded events.Milestones, smaller events and setbacks share one sourced record. Scheduled targets are marked separately from recorded events.
Event key
The symbolic dream
This record spans “Turing asks "Can machines think?"” to “The perceptron learns from data”.
Winters and the wilderness
This record spans “The first AI winter sets in” to “The second AI winter”.
Machines that learn
This record spans “Deep Blue beats Kasparov” to “AlexNet ignites the deep-learning era”.
Foundation models and agents
This record spans “The transformer architecture” to “Forecast: "weakly general" AI” and includes scheduled targets.
Additional records
This record spans “AlphaGo defeats Lee Sedol” to “Projected: AI handles week-long tasks” and includes scheduled targets.
The data behind the verdictThe data behind the verdict
Read each measure with its units, conditions, uncertainty and source; targets remain distinct from observations.Read each chart with its conditions, limits and source. Future targets remain distinct from recorded results.
log scale · minutes by year
GPT-2 · 3 s · Feb 2019 · interval 0.6–8.4 s · verified by METR
First record
GPT-23 s · Feb 2019
observed · Verified by METR
Published interval: 0.6–8.4 s
Latest record
Claude Mythos Preview (early)17 h · Apr 2026
observed · Verified by METR
Published interval: 8.5–55 h
Methods & conditions
Under the measurement.
The boundary of a measurement changes what it means.
Autonomy gains depend on the time window
Method and conditions
TH1.1's full 2019-2025 fitted doubling time remains about 196 days, but its post-2023 estimate shortened from 165.3 to 130.8 days and its post-2024 estimate from 108.9 to 88.6 days. These rates are sensitive to task selection and should not be extrapolated as physical laws.
Source notes
Technical notes and sources
Each note keeps its authored method, conditions and citations available for review.
Autonomy gains depend on the time window
TH1.1's full 2019-2025 fitted doubling time remains about 196 days, but its post-2023 estimate shortened from 165.3 to 130.8 days and its post-2024 estimate from 108.9 to 88.6 days. These rates are sensitive to task selection and should not be extrapolated as physical laws.
Long-task baselines are mostly estimated
Only 5 of TH1.1's 31 tasks estimated at eight hours or longer have measured human baselines; the rest use expert estimates. The suite covers mainly software engineering, machine learning and cybersecurity, not general intellectual work.
What a time horizon measures
A time horizon denotes the human duration of task difficulty at which fitted agent success is 50%, not how long the agent literally operates. Agents are often several times faster than the associated human time.
MirrorCode gives unusually rich feedback
MirrorCode gives agents execute-only access to the reference program, documentation, shell tools and very large inference budgets-up to one billion tokens, about $550 per task in preliminary experiments. This unusually precise, testable specification helps explain its much longer apparent horizons.
The gotree baseline is uncertain
The gotree human-duration range-2-17 weeks-comes from four informal estimates, not completed human trials. A shorter 2,000-line baseline was still only 42% complete after 20 hours.
Interactive adaptation depends on the harness
ARC-AGI-3 tests continual interactive adaptation rather than static answer production. Early replay analysis found GPT-5.5 at 0.43% and Opus 4.7 at 0.18%, failing to form correct world models or transfer level strategies. In September 2026 ARC Prize verified GPT-6 Astra at 62.7% with its Standard harness and 99.9% with a Provider Adapter harness that preserves the model's reasoning state; ARC reported that Astra built compact symbolic world models of each game.
Why the 2025 IMO result is stronger
The 2025 IMO result is technically stronger than the 2024 silver result: it worked end-to-end in natural language within 4.5 hours, whereas the 2024 systems needed human translation into formal languages and as much as three days. The IMO validated the submitted proofs, not Google's model or experimental process.
Productivity estimates remain inconclusive
METR's attempted late-2025 productivity follow-up produced raw estimates of 18% speedup for returning developers and 4% for new recruits, but both confidence intervals included zero and METR judged selection and time-reporting biases too severe for a reliable magnitude.
Benchmark errors limit headline scores
Static benchmark quality is itself a constraint: Stanford's 2026 review reports invalid-question estimates ranging from 2% on MMLU Math to 42% on GSM8K, reinforcing that saturated scores should not directly set an AGI percentage.
Every test we can still hold it toThe tests — and who leads each one
Each row is an independent benchmark: its verified frontier, any named human mark, and how much ruler remains. Solid dots are steward-verified; dashed circles are lab claims. No row is averaged into another.Each row is one test. A solid dot is independently checked; a dashed circle is a company claim. We never add the rows into one big number.
Brand-new visual puzzlesARC-AGI-2
95%human panelhuman 66ARC-AGI-2GPT-6 Astra (Max)Engineer detail · representative record, conditions and limitations
GPT-6 Astra (Max) · 95 % of tasksOpenAI · 2026-09 · verified · ARC Prize Foundation
ARC Prize semi-private verified result. Claude Opus 5.5 (High) scored 93.3 and GPT-6 Sol (Max) 89.6. The competition threshold applies to the private evaluation under its efficiency limits.
Evaluation conditions
Abstract reasoning on novel grid tasks designed to resist memorisation.Puzzles the model has not seen before.
Steward: ARC Prize Foundation. Updated 2026-09-30.
Sources and methods
ARC Prize semi-private verified result. Claude Opus 5.5 (High) scored 93.3 and GPT-6 Sol (Max) 89.6. The competition threshold applies to the private evaluation under its efficiency limits.
Source record ↗human panel source ↗2026 prize threshold source ↗Limitations
Public and private evaluation scores are not interchangeable.
A high score on one static puzzle suite is not a measure of general intelligence.
Learning by playingARC-AGI-3
62.7%human referencehuman 100ARC-AGI-3GPT-6 Astra (Max)Engineer detail · representative record, conditions and limitations
GPT-6 Astra (Max) · 62.71 % scoreOpenAI · 2026-09 · verified · ARC Prize Foundation
Standard harness, the basis of earlier ARC-AGI-3 records here (GPT-5.6 Sol 7.78, Claude Opus 5 30.16). With ARC's Provider Adapter harness, which keeps the model's own reasoning state between turns, the same model scored 99.95 (High). The two harnesses are not comparable.
Evaluation conditions
Interactive adaptation in unfamiliar environments rather than static answer production.Figuring out a new game by trying things, without instructions.
Steward: ARC Prize Foundation. Updated 2026-09-30.
Sources and methods
Standard harness, the basis of earlier ARC-AGI-3 records here (GPT-5.6 Sol 7.78, Claude Opus 5 30.16). With ARC's Provider Adapter harness, which keeps the model's own reasoning state between turns, the same model scored 99.95 (High). The two harnesses are not comparable.
Source record ↗human reference source ↗Limitations
Scores depend heavily on the harness: GPT-6 Astra scored 62.7% with ARC's Standard harness and 99.9% with the Provider Adapter harness in the same verification.
ARC Prize says a high ARC-AGI-3 score is not by itself proof of general intelligence; the benchmark is still new.
Questions at the edge of human knowledgeHumanity's Last Exam
54.8%Humanity's Last ExamGPT-6 AstraEngineer detail · representative record, conditions and limitations
GPT-6 Astra · 54.8 % correctOpenAI · 2026-09 · verified · Scale SEAL
Published interval: 52.86–56.74 % correct
Scale SEAL leaderboard as read 30 Sep 2026 (last updated 17 Sep 2026). Next: Claude Fable 5.1 (xhigh) 46.50 ±2.00 and Gemini 3.1 Pro Preview 46.44 ±1.96. Lab-run HLE numbers use other harnesses and are not comparable.
Evaluation conditions
Expert-written questions intended to test frontier knowledge and reasoning.Very hard questions written by experts in many subjects.
Steward: Scale SEAL. Updated 2026-09-30.
Sources and methods
Scale SEAL leaderboard as read 30 Sep 2026 (last updated 17 Sep 2026). Next: Claude Fable 5.1 (xhigh) 46.50 ±2.00 and Gemini 3.1 Pro Preview 46.44 ±1.96. Lab-run HLE numbers use other harnesses and are not comparable.
Source record ↗Limitations
The score is specific to the benchmark setup; it is not a cross-domain capability percentage.
Confidence intervals mean nearby entries may not support a meaningful ranking.
Longer coding jobsSWE-bench Pro (public)
61.5%SWE-bench Pro (public)Muse Spark 1.1Engineer detail · representative record, conditions and limitations
Muse Spark 1.1 · 61.5 % solvedMeta · 2026-09 · verified · Scale SEAL
Published interval: 58.4–64.6 % solved
Scale SEAL public set with the mini-swe-agent harness, as read 30 Sep 2026; the page shows no update date. Scale ranks it tied first with GPT-5.4 (xHigh) at 59.10 ±3.56, so the order is not meaningful. Lab-run SWE-bench Pro numbers (e.g. 89.9 for Claude Opus 5.5) use other harnesses and are not comparable.
Evaluation conditions
Agent performance on long-horizon software-engineering tasks in public open-source repositories.Whether an AI agent can complete harder real coding work.
Steward: Scale SEAL. Updated 2026-09-30.
Sources and methods
Scale SEAL public set with the mini-swe-agent harness, as read 30 Sep 2026; the page shows no update date. Scale ranks it tied first with GPT-5.4 (xHigh) at 59.10 ±3.56, so the order is not meaningful. Lab-run SWE-bench Pro numbers (e.g. 89.9 for Claude Opus 5.5) use other harnesses and are not comparable.
Source record ↗Limitations
This is an agent-plus-harness evaluation, not a property of the base model alone.
The board does not compare this percentage directly with puzzle or knowledge benchmarks.
The earlier puzzle setARC-AGI-1
98.5%ARC-AGI-1saturatedGPT-6 Astra (High)Engineer detail · representative record, conditions and limitations
GPT-6 Astra (High) · 98.5 % of tasksOpenAI · 2026-09 · verified · ARC Prize Foundation
Tied by Claude Opus 5.5 (High) at 98.5. The benchmark is saturated.
Evaluation conditions
The original static ARC puzzle suite, superseded by ARC-AGI-2.The first version of the visual puzzle test.
ARC Prize explicitly described ARC-AGI-1 as saturating and raised the bar with ARC-AGI-2.
Sources and methods
Tied by Claude Opus 5.5 (High) at 98.5. The benchmark is saturated.
Source record ↗Limitations
ARC Prize's own transition to ARC-AGI-2 is the reason this ruler is shown as outgrown rather than as a fresh frontier.
The old bug-fixing testSWE-bench Verified
80.9%SWE-bench VerifiedwithdrawnGPT-5.2Engineer detail · representative record, conditions and limitations
GPT-5.2 · 80.9 % resolvedOpenAI · 2026-02 · self-reported
Shown as a lab-reported historical mark, not as a current verified frontier.
Evaluation conditions
Resolving issues from public code repositories; now judged contaminated and affected by flawed tests.The old test for fixing real code bugs.
OpenAI stopped reporting this score because contamination and flawed tests made it unreliable for frontier comparison.
Sources and methods
Shown as a lab-reported historical mark, not as a current verified frontier.
Source record ↗Limitations
OpenAI's audit found evidence that frontier models had seen some problems or solutions in training.
The replacement ruler, SWE-bench Pro, is still active above.
How long a job it can finish is on a log axis — see the data figure.
each ruler 0–100 in its own unit · never summed
View every result as a table
| Ruler | Model | Lab | Value | CI | Date | Verification | Conditions | Caveats | Record note | Source |
|---|---|---|---|---|---|---|---|---|---|---|
| Brand-new visual puzzles | GPT-6 Astra (Max) | OpenAI | 95 % of tasks | — | 2026-09 | verified · ARC Prize Foundation | Puzzles the model has not seen before. | Public and private evaluation scores are not interchangeable. A high score on one static puzzle suite is not a measure of general intelligence. | ARC Prize semi-private verified result. Claude Opus 5.5 (High) scored 93.3 and GPT-6 Sol (Max) 89.6. The competition threshold applies to the private evaluation under its efficiency limits. | Source ↗ |
| Learning by playing | GPT-6 Astra (Max) | OpenAI | 62.71 % score | — | 2026-09 | verified · ARC Prize Foundation | Figuring out a new game by trying things, without instructions. | Scores depend heavily on the harness: GPT-6 Astra scored 62.7% with ARC's Standard harness and 99.9% with the Provider Adapter harness in the same verification. ARC Prize says a high ARC-AGI-3 score is not by itself proof of general intelligence; the benchmark is still new. | Standard harness, the basis of earlier ARC-AGI-3 records here (GPT-5.6 Sol 7.78, Claude Opus 5 30.16). With ARC's Provider Adapter harness, which keeps the model's own reasoning state between turns, the same model scored 99.95 (High). The two harnesses are not comparable. | Source ↗ |
| How long a job it can finish | GPT-2 | OpenAI | 0.05 minutes (log scale) | 0.01–0.14 minutes (log scale) | 2019-02 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | GPT-3 (davinci-002) | OpenAI | 0.14 minutes (log scale) | 0.09–0.22 minutes (log scale) | 2020-05 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | GPT-3.5 Turbo Instruct | OpenAI | 0.6 minutes (log scale) | 0.26–1.12 minutes (log scale) | 2022-03 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | GPT-4 | OpenAI | 3.99 minutes (log scale) | 1.93–8 minutes (log scale) | 2023-03 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | GPT-4 (1106) | OpenAI | 4.04 minutes (log scale) | 1.87–8.44 minutes (log scale) | 2023-11 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | Claude 3 Opus | Anthropic | 3.95 minutes (log scale) | 1.71–8.76 minutes (log scale) | 2024-03 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | GPT-4 Turbo | OpenAI | 3.73 minutes (log scale) | 1.98–6.74 minutes (log scale) | 2024-04 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | GPT-4o | OpenAI | 6.99 minutes (log scale) | 4–12.91 minutes (log scale) | 2024-05 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | Claude 3.5 Sonnet (Jun 2024) | Anthropic | 11.4 minutes (log scale) | 5.49–22.38 minutes (log scale) | 2024-06 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | o1-preview | OpenAI | 20.33 minutes (log scale) | 11.72–33.38 minutes (log scale) | 2024-09 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | Claude 3.5 Sonnet (Oct 2024) | Anthropic | 20.52 minutes (log scale) | 10.14–40.82 minutes (log scale) | 2024-10 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | o1 | OpenAI | 38.83 minutes (log scale) | 21.16–64.95 minutes (log scale) | 2024-12 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | Claude 3.7 Sonnet | Anthropic | 60.39 minutes (log scale) | 33.01–104.23 minutes (log scale) | 2025-02 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | o3 | OpenAI | 119.73 minutes (log scale) | 74.62–190.94 minutes (log scale) | 2025-04 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | Claude Opus 4 | Anthropic | 100.37 minutes (log scale) | 59.99–163.45 minutes (log scale) | 2025-05 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | Claude Opus 4.1 | Anthropic | 100.47 minutes (log scale) | 59.27–159.45 minutes (log scale) | 2025-08 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | GPT-5 | OpenAI | 203.01 minutes (log scale) | 112.64–405.55 minutes (log scale) | 2025-08 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | Gemini 3 Pro | Google DeepMind | 224.33 minutes (log scale) | 139.57–379.24 minutes (log scale) | 2025-11 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | GPT-5.1 Codex Max | OpenAI | 223.71 minutes (log scale) | 134.31–396.21 minutes (log scale) | 2025-11 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | Claude Opus 4.5 | Anthropic | 292.99 minutes (log scale) | 161.72–623.7 minutes (log scale) | 2025-11 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | GPT-5.2 | OpenAI | 352.25 minutes (log scale) | 198.07–815.18 minutes (log scale) | 2025-12 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | Claude Opus 4.6 | Anthropic | 718.81 minutes (log scale) | 316.69–3633.79 minutes (log scale) | 2026-02 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| How long a job it can finish | GPT-5.3 Codex | OpenAI | 349.53 minutes (log scale) | 194.91–816.35 minutes (log scale) | 2026-02 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | Gemini 3.1 Pro | Google DeepMind | 384.15 minutes (log scale) | 233.5–694.75 minutes (log scale) | 2026-02 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | GPT-5.4 | OpenAI | 341.74 minutes (log scale) | 186.58–768.78 minutes (log scale) | 2026-03 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | — | Source ↗ |
| How long a job it can finish | Claude Mythos Preview (early) | Anthropic | 1044.78 minutes (log scale) | 508.88–3304.26 minutes (log scale) | 2026-04 | verified · METR | The length of a job it can finish alone about half the time. | METR reports wide uncertainty for current models and warns that horizons vary sharply by domain. METR revised its own Horizon estimates down by up to 20% in March 2026; these are the corrected TH1.1 figures, re-pulled from METR's raw results file. A horizon is human task-duration at 50% success, not wall-clock autonomous runtime. The hatched range marks where this simplified display should not invite extrapolation, not a hard physical ceiling. | State of the art on METR's Horizon suite at release. | Source ↗ |
| Questions at the edge of human knowledge | GPT-6 Astra | OpenAI | 54.8 % correct | 52.86–56.74 % correct | 2026-09 | verified · Scale SEAL | Very hard questions written by experts in many subjects. | The score is specific to the benchmark setup; it is not a cross-domain capability percentage. Confidence intervals mean nearby entries may not support a meaningful ranking. | Scale SEAL leaderboard as read 30 Sep 2026 (last updated 17 Sep 2026). Next: Claude Fable 5.1 (xhigh) 46.50 ±2.00 and Gemini 3.1 Pro Preview 46.44 ±1.96. Lab-run HLE numbers use other harnesses and are not comparable. | Source ↗ |
| Longer coding jobs | Muse Spark 1.1 | Meta | 61.5 % solved | 58.4–64.6 % solved | 2026-09 | verified · Scale SEAL | Whether an AI agent can complete harder real coding work. | This is an agent-plus-harness evaluation, not a property of the base model alone. The board does not compare this percentage directly with puzzle or knowledge benchmarks. | Scale SEAL public set with the mini-swe-agent harness, as read 30 Sep 2026; the page shows no update date. Scale ranks it tied first with GPT-5.4 (xHigh) at 59.10 ±3.56, so the order is not meaningful. Lab-run SWE-bench Pro numbers (e.g. 89.9 for Claude Opus 5.5) use other harnesses and are not comparable. | Source ↗ |
| The earlier puzzle set | GPT-6 Astra (High) | OpenAI | 98.5 % of tasks | — | 2026-09 | verified · ARC Prize Foundation | The first version of the visual puzzle test. | ARC Prize's own transition to ARC-AGI-2 is the reason this ruler is shown as outgrown rather than as a fresh frontier. | Tied by Claude Opus 5.5 (High) at 98.5. The benchmark is saturated. | Source ↗ |
| The old bug-fixing test | GPT-5.2 | OpenAI | 80.9 % resolved | — | 2026-02 | self-reported | The old test for fixing real code bugs. | OpenAI's audit found evidence that frontier models had seen some problems or solutions in training. The replacement ruler, SWE-bench Pro, is still active above. | Shown as a lab-reported historical mark, not as a current verified frontier. | Source ↗ |
If the remaining tests pass.If the remaining tests pass.
Downstream capabilities remain conditional on results that are not yet in.These are possible next capabilities, conditional on results that are not yet in.
- OPEN TEST01 ↗Learn many different kinds of workBroad cognitive competenceA general system should transfer what it knows to fresh problems, not need a new trick for every benchmark.Expert-level performance across diverse, contamination-resistant novel tasks
- OPEN TEST02 ↗Finish long jobs without hand-holdingLong-horizon autonomyThe measured span an agent can complete on its own is still counted in hours, not days.Reliably plan and finish multi-day digital projects with sparse supervision
- OPEN TEST03 ↗Stay reliable when things get weirdCalibrated reliabilityBeing capable is not enough if unusual inputs still produce confident, costly errors.Predictable, well-calibrated, corrigible behavior under distribution shift and adversarial pressure
Possible, if the open tests pass.
Faster robotics research
This is a possible benefit only if faster robotics research works safely, reliably and at useful scale.Potential benefit only if the stated use (Faster robotics research) is demonstrated with evidence for safety, reliability, affordability and scale.
Related problem ↗Faster longevity research
This is a possible benefit only if faster longevity research works safely, reliably and at useful scale.Potential benefit only if the stated use (Faster longevity research) is demonstrated with evidence for safety, reliability, affordability and scale.
Related problem ↗Accelerated science everywhere
This is a possible benefit only if accelerated science everywhere works safely, reliably and at useful scale.Potential benefit only if the stated use (Accelerated science everywhere) is demonstrated with evidence for safety, reliability, affordability and scale.
AI plasma control
This is a possible benefit only if ai plasma control works safely, reliably and at useful scale.Potential benefit only if the stated use (AI plasma control) is demonstrated with evidence for safety, reliability, affordability and scale.
Related problem ↗AI-designed vaccines
This is a possible benefit only if ai-designed vaccines works safely, reliably and at useful scale.Potential benefit only if the stated use (AI-designed vaccines) is demonstrated with evidence for safety, reliability, affordability and scale.
Related problem ↗Better neural decoders
This is a possible benefit only if better neural decoders works safely, reliably and at useful scale.Potential benefit only if the stated use (Better neural decoders) is demonstrated with evidence for safety, reliability, affordability and scale.
Related problem ↗
Meet the builders.Meet the builders.
The organisations, their approaches, and the evidence behind them.The teams doing the work, the money around them, and what the record can support.
USA · company
OpenAI
Frontier multimodal and reasoning models, agent products and large-scale training infrastructure.
Capital is context, not a demonstration.
Company source ↗View every builder record
OpenAI
USA · company · Frontier multimodal and reasoning models, agent products and large-scale training infrastructure.
Funding: $122B committed capital · Mar 2026 · $852B post-money valuation
Named investors: Amazon · NVIDIA · SoftBank · Microsoft
Committed capital is not cumulative funding, revenue or money already spent on compute.
Company source ↗Anthropic
USA · company · Frontier Claude models with interpretability, alignment and model-safety research.
Funding: Series F · $13B · Sep 2025 · $183B post-money valuation
Named investors: ICONIQ · Fidelity · Lightspeed · BlackRock affiliates · Qatar Investment Authority
The reported $5B August 2025 revenue figure is an annualized run-rate, not audited full-year revenue.
Company source ↗xAI
USA · company · Grok model family and vertically integrated large-scale training infrastructure.
Funding: Series E · $20B · Jan 2026
Named investors: Valor Equity Partners · Fidelity · Qatar Investment Authority · NVIDIA · Cisco
No cumulative total is encoded because the release does not reconcile equity, debt and infrastructure financing.
Company source ↗Mistral AI
France · company · European frontier and open-weight models, enterprise deployment and sovereign AI tooling.
Funding: Series C · €1.7B · Sep 2025 · €11.7B post-money valuation
Named investors: ASML · Bpifrance · NVIDIA
ASML invested €1.3B; valuation is stated separately from capital raised.
Company source ↗Google DeepMind
UK / USA · lab · Gemini frontier models, reinforcement learning, AI-for-science and robotics research.
Funding: Funding not recorded in this record.
Funded inside Alphabet; corporate capex is not a standalone lab round.
Company source ↗Meta AI
USA · lab · Open-weight Llama models and research in multimodal, world-model and embodied AI.
Funding: Funding not recorded in this record.
Funded inside Meta; company-wide infrastructure spending is not lab funding.
Company source ↗Follow the evidence.Follow the evidence.
54 sources — every figure on this page traces to one.54 sources — every figure on this page traces to one.
- 01
Scope: Not recorded. Limit: Not recorded.
- 02
ARC Prize - ARC-AGI benchmark
Read source ↗arcprize.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 03
OpenAI - Why Language Models Hallucinate (2025)
Read source ↗openai.comScope: Not recorded. Limit: Not recorded.
- 04
Metaculus - Date of the first general AI system
Read source ↗metaculus.comScope: Not recorded. Limit: Not recorded.
- 05
Timeline · Turing asks "Can machines think?"
Read source ↗en.wikipedia.orgUsed on this pageAll 2 uses across 1 destination
Scope: Not recorded. Limit: Not recorded.
- 06
Timeline · The Dartmouth workshop names the field
Read source ↗en.wikipedia.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 07
Timeline · The perceptron learns from data
Read source ↗en.wikipedia.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 08
Scope: Not recorded. Limit: Not recorded.
- 09
Timeline · Backpropagation revives neural nets
Read source ↗nature.comUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 10
Timeline · Deep Blue beats Kasparov
Read source ↗en.wikipedia.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 11
Timeline · AlexNet ignites the deep-learning era
Read source ↗en.wikipedia.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 12
Timeline · AlphaGo defeats Lee Sedol
Read source ↗en.wikipedia.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 13
Scope: Not recorded. Limit: Not recorded.
- 14
Timeline · GPT-2 shows scale generates fluent text
Read source ↗en.wikipedia.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 15
Timeline · GPT-3 and few-shot learning
Read source ↗arxiv.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 16
Scope: Not recorded. Limit: Not recorded.
- 17
Timeline · GPT-4 clears broad exam benchmarks
Read source ↗cdn.openai.comUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 18
Timeline · SWE-bench tests repository-scale coding
Read source ↗arxiv.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 19
Timeline · Formal systems reach IMO silver standard
Read source ↗deepmind.googleUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 20
Timeline · Reasoning models and agents arrive
Read source ↗openai.comUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 21
Timeline · AI task-time-horizon doubles ~every 7 months
Read source ↗arxiv.orgScope: Not recorded. Limit: Not recorded.
- 22
Timeline · Real-world coding trial finds a slowdown
Read source ↗metr.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 23
Timeline · Natural-language system reaches IMO gold standard
Read source ↗deepmind.googleUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 24
Timeline · ARC-AGI-2 improves, but grand prize stays open
Read source ↗arcprize.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 25
Scope: Not recorded. Limit: Not recorded.
- 26
Timeline · ARC-AGI-3 exposes an interactive gap
Read source ↗arcprize.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 27
Timeline · MirrorCode tests whole-program reconstruction
Read source ↗arxiv.orgUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 28
Scope: Not recorded. Limit: Not recorded.
- 29
Series · Task length an AI finishes alone
Read source ↗metr.orgScope: Not recorded. Limit: Not recorded.
- 30
Engineer note · What a time horizon measures
Read source ↗metr.orgScope: Not recorded. Limit: Not recorded.
- 31
Engineer note · MirrorCode gives unusually rich feedback
Read source ↗epoch.aiScope: Not recorded. Limit: Not recorded.
- 32
Engineer note · Interactive adaptation remains weak
Read source ↗arcprize.orgScope: Not recorded. Limit: Not recorded.
- 33
Engineer note · Productivity estimates remain inconclusive
Read source ↗metr.orgScope: Not recorded. Limit: Not recorded.
- 34
Engineer note · Benchmark errors limit headline scores
Read source ↗hai.stanford.eduScope: Not recorded. Limit: Not recorded.
- 35
Scope: Not recorded. Limit: Not recorded.
- 36
Scope: Not recorded. Limit: Not recorded.
- 37
Scope: Not recorded. Limit: Not recorded.
- 38
companies · Mistral AI
Read source ↗presse.bpifrance.frUsed on this pageAll 2 uses across 1 destination
Scope: Not recorded. Limit: Not recorded.
- 39
companies · Google DeepMind
Read source ↗deepmind.googleUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 40
Scope: Not recorded. Limit: Not recorded.
- 41
Scope: Not recorded. Limit: Not recorded.
- 42
Scope: Not recorded. Limit: Not recorded.
- 43
Scope: Not recorded. Limit: Not recorded.
- 44
Scope: Not recorded. Limit: Not recorded.
- 45
market · €10B EU supercomputing and AI-factory programme
Read source ↗digital-strategy.ec.europa.euUsed on this pageAll 2 uses across 2 destinations
Scope: Not recorded. Limit: Not recorded.
- 46
Scope: Not recorded. Limit: Not recorded.
- 47
OpenAI official site
Read source ↗openai.comScope: Not recorded. Limit: Not recorded.
- 48
OpenAI logo review
Read source ↗openai.comScope: Not recorded. Limit: Not recorded.
- 49
Anthropic official site
Read source ↗anthropic.comScope: Not recorded. Limit: Not recorded.
- 50
xAI official site
Read source ↗x.aiScope: Not recorded. Limit: Not recorded.
- 51
Mistral AI official site
Read source ↗mistral.aiScope: Not recorded. Limit: Not recorded.
- 52
Mistral AI logo review
Read source ↗mistral.aiScope: Not recorded. Limit: Not recorded.
- 53
Google DeepMind official site
Read source ↗deepmind.googleScope: Not recorded. Limit: Not recorded.
- 54
Timeline · GPT-6 Astra scores 62.7% on ARC-AGI-3
Read source ↗arcprize.orgUsed on this pageAll 1 use across 1 destination
Scope: Not recorded. Limit: Not recorded.
Sources describe evidence. Inclusion is not endorsement.