howclose.to
Intelligence & Machines · Updated July 2026Momentum · accelerating

How close are we to artificial general intelligence?When will AI learn almost any thinking task?

or, simply: When will AI learn almost any thinking task?or, precisely: 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.

We are here

MirrorCode tests whole-program reconstruction - MirrorCode asked agents to recreate complete programs from their behavior without source code. Its strongest model passed 56% across 25 targets and nearly reproduced the 16,000-line gotree toolkit, but used unusually precise executable feedback and large budgets; the claimed weeks of human work were estimates, not completed trials. Next up - Forecast: "weakly general" AI (expected 2028).

01 · Where we stand

State of playWhere agi stands right now

The current stage, the honest metric, and the single threshold that gates the next stage. Each threshold is a falsifiable claim with a named next test.How far up the ladder we've climbed, the honest verdict, and the one thing blocking the next step.

The five stagesMaturity ladder
Deployed
Out in the real worldDeployed at scale
Scaling
Making it cheap enough at scaleScaling toward competitive cost
Engineering← HERE
Building one that pays for itselfEngineering a system that pays back
Lab demo
Shown to work in a labDemonstrated in the laboratory
Theoretical
The idea is worked out on paperTheoretical basis established
The honest verdictVerified state

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.

No single honest number for this one - read where it stands by the milestones and the next test.No single honest scalar - progress is read by milestones and the named next test, not a headline figure.

Blocking the next stepBlocking threshold

Learn many different kinds of workBroad cognitive competence Next test: ARC-AGI-2 and held-out expert evals resistant to training-data leakage

The verdict, in five rungsHow far up the ladder↓ next — the thresholds & the gap
01 · The evidence

The thresholds that gate the next stageWhat has to happen next

Each threshold is a falsifiable claim with a named next test; the gap chart shows how far today's metric sits from the goal.Each row is one thing that has to be proven — and how far today's number is from the target.

Broad cognitive competenceLearn many different kinds of workIn progress
62%
Next testARC-AGI-2 and held-out expert evals resistant to training-data leakage
Long-horizon autonomyFinish long jobs without hand-holdingEarly
28%
Next testMETR 50%-task time horizon extending from hours to multi-day tasks
Calibrated reliabilityStay reliable when things get weirdEarly
18%
Next testPre-registered external safety and calibration evaluations on frontier deployments
THRESHOLDS - Thresholds for AGI.
02 · How we got here

The record behind the verdict

Major events set large; context events set small but never hidden. Everything below the TODAY rule is a schedule, not a result.

1950-19733 events1 shown

The symbolic dream

The symbolic dream moved the field from turing asks "can machines think?" to the perceptron learns from data. The results narrowed the next question without closing it.

1950
Turing asks "Can machines think?"
Alan Turing's paper proposes the Imitation Game (Turing Test) as an operational bar for machine intelligence.
1956
The Dartmouth workshop names the fieldTheory
A summer workshop at Dartmouth College coins the term "artificial intelligence" and launches AI as a formal research discipline.
1958
The perceptron learns from data
Frank Rosenblatt builds the perceptron, an early trainable artificial neural network that adjusts its own weights to classify inputs.
1974-19963 events1 shown

Winters and the wilderness

Winters and the wilderness moved the field from the first ai winter sets in to the second ai winter. The results narrowed the next question without closing it.

1974
The first AI winter sets in
After the critique of perceptrons and the pessimistic Lighthill report, funding and optimism collapse for roughly a decade.
1986
Backpropagation revives neural netsTheory
Rumelhart, Hinton, and Williams popularize backpropagation, making it practical to train multi-layer networks and reigniting connectionism.
1987
The second AI winter
The specialized LISP-machine market collapses and expert-system hype deflates, triggering a second multi-year funding drought.
1997-20162 events2 shown

Machines that learn

Machines that learn moved the field from deep blue beats kasparov to alexnet ignites the deep-learning era. The results narrowed the next question without closing it.

1997
Deep Blue beats KasparovExperiment
IBM's Deep Blue defeats reigning world chess champion Garry Kasparov, the first computer to beat a champion under tournament conditions.
2012
AlexNet ignites the deep-learning eraExperiment
A deep convolutional network wins the ImageNet competition by a huge margin, proving GPU-trained neural nets on big data beat hand-engineered vision.
2017-202917 events4 shown

Foundation models and agents

Foundation models and agents moved the field from the transformer architecture to forecast: "weakly general" ai. The results narrowed the next question without closing it.

2017
The transformer architectureTheory
"Attention Is All You Need" introduces the transformer, the attention-based architecture that underpins virtually every modern large language model.
2019
GPT-2 shows scale generates fluent text
OpenAI's GPT-2 produces coherent long-form text and is released in staged fashion over stated misuse concerns, foreshadowing the scaling era.
2020
GPT-3 and few-shot learning
The 175-billion-parameter GPT-3 performs many tasks from a few examples in its prompt, demonstrating emergent capability from raw scale.
2022
ChatGPT reaches the publicDeployment
OpenAI releases ChatGPT, and a conversational, instruction-tuned LLM reaches mainstream users, becoming one of the fastest-growing apps ever.
2023
GPT-4 clears broad exam benchmarks
GPT-4 scored 298/400, about the 90th percentile, on a simulated Uniform Bar Exam and improved sharply over GPT-3.5 across many academic tests. This was broad benchmark performance, not AGI, and the report still described the model as unreliable.
2023
SWE-bench tests repository-scale coding
SWE-bench introduced 2,294 real GitHub issues from 12 Python repositories. The best system tested in the original paper, Claude 2, resolved only 1.96%, exposing how far code generation was from dependable software work.
2024
Formal systems reach IMO silver standard
AlphaProof and AlphaGeometry 2 solved four of six 2024 International Mathematical Olympiad problems for 28/42, one point below gold. Humans first translated the problems into formal languages, and some solutions took up to three days.
2024
Reasoning models and agents arrive
OpenAI's o1 spends inference-time "thinking" before answering, opening the o-series reasoning line that powers more autonomous agentic systems.
2025
AI task-time-horizon doubles ~every 7 monthsExperiment
METR finds the length of software tasks AI can finish at 50% reliability has grown exponentially, doubling roughly every seven months since 2019.
2025
Benchmarks saturate while autonomy lags
Frontier models top out static benchmarks like MMLU and ARC-AGI-1, yet still struggle with long-horizon autonomy, exposing a gap between test scores and real agency.
2025
Real-world coding trial finds a slowdown
In a randomized trial, 16 experienced open-source developers completed 246 issues 19% more slowly when allowed to use early-2025 AI tools. METR cautions that this was one historical, setting-specific result, not a verdict on all developers or later systems.
2025
Natural-language system reaches IMO gold standard
An advanced Gemini Deep Think system solved five of six 2025 IMO problems for 35/42. IMO graders certified the proofs, which were produced in natural language within the 4.5-hour contest limit; this was a closed competition task, not autonomous research.
2025
ARC-AGI-2 improves, but grand prize stays open
The ARC Prize competition winner scored 24.03% on the private evaluation at about $0.20 per task. The best verified commercial model reached 37.6% and a costlier refinement system 54%; the 85% grand-prize threshold remained unmet.
2026
METR expands its autonomy ruler
Time Horizon 1.1 expanded from 170 to 228 mainly software, machine-learning and cybersecurity tasks. Claude Opus 4.5's estimated 50%-success horizon was 320 minutes, with wide uncertainty and only five measured human baselines among 31 tasks estimated at eight hours or longer.
2026
ARC-AGI-3 exposes an interactive gap
ARC-AGI-3 introduced instruction-free interactive environments that require exploration and adaptation. Humans scored 100% while the reported frontier baseline scored 0.51%, a separate test that cannot be compared directly with ARC-AGI-1 or ARC-AGI-2 percentages.
2026
MirrorCode tests whole-program reconstructionWe are here
MirrorCode asked agents to recreate complete programs from their behavior without source code. Its strongest model passed 56% across 25 targets and nearly reproduced the 16,000-line gotree toolkit, but used unusually precise executable feedback and large budgets; the claimed weeks of human work were estimates, not completed trials.
2028
Forecast: "weakly general" AIPolicyTarget
The Metaculus community's aggregate forecast places the arrival of a publicly known "weakly general" AI in the late 2020s, around 2028.
2016-20292 events1 shown

Events outside the declared eras

Events outside the declared eras moved the field from alphago defeats lee sedol to projected: ai handles week-long tasks. The results narrowed the next question without closing it.

2016
AlphaGo defeats Lee SedolExperiment
DeepMind's AlphaGo beats top Go professional Lee Sedol 4-1, mastering a game long considered a landmark for intuition and search.
2029
Projected: AI handles week-long tasksExperimentTarget
Extrapolating METR's seven-month doubling implies agents could autonomously complete software tasks that take humans days or weeks by the end of the decade.
— end of record · 27 shown, 0 hidden —
27 events · below the TODAY rule = scheduled, not done
03 · The data behind the verdict

Why the meters read the way they do

The learning curves and comparisons that justify each threshold's percentage. Every series is measured, with the source event linked in the timeline above.

The length of job an AI model can finish on its own has doubled about every four months since 2023.Task length an AI finishes alone · minutes · log scale
The length of job an AI model can finish on its own has doubled about every four months since 2023.6 s1 min10 min1.7 h17 hTask length an AI finishes aloneYear2019202220242026GPT-2Claude Mythos Preview (earl…17 h
NOTE - The length of job an AI model can finish on its own has doubled about every four months since 2023.
Technical notes

Read the evidence more closely

Definitions, system boundaries and experimental caveats behind the headline record.

01Autonomy 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.

02Long-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.

03What 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.

04MirrorCode 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.

05The 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.

06Interactive adaptation remains weak

ARC-AGI-3 tests continual interactive adaptation rather than static answer production. In later replay analysis, GPT-5.5 scored 0.43% and Opus 4.7 scored 0.18%; common failures included learning a local action effect without forming a correct world model and failing to transfer a successful level strategy.

07Why 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.

08Productivity 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.

09Benchmark 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.

04 · The rulers

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.

7 rulers tracked · 5 still standing · 2 outgrown or withdrawn

Trace a lab
steward-verified frontier lab claim / historical report human or named target exhausted or unreliable range
Still standingEvery dot has a source, date and verification status.
Outgrown or withdrawnKept on the record because benchmark deaths are evidence too.

READ THIS BOARD — a dot near the right edge means one ruler is nearly used up, not that AGI is nearly done. The rows that remain stubborn are often the more useful evidence.

05 · What it unlocks

If the remaining tests pass

Downstream capabilities, drawn dashed because they depend on results not yet in.

AGIFaster robotics researchAI improves robot learning, planning and hardware designFaster longevity researchAI helps design experiments, molecules and biological modelsAccelerated science everywhereresearch cycles shrink across fields from materials to medicineAI plasma controlMachine learning helps steer and stabilize the plasma inside a fusion reactor.AI-designed vaccinesAI can design antigens and compress the path from a new pathogen to an authorized shot.Better neural decodersAI turns noisy brain signals into fluent text and speech far faster than hand-built decoders.
06 · The players & the money

Who is building it-and what the money saysCapital, institutions and the global race

The teams doing the work, where they are based, and whether the money points to real delivery or only a plan.Company finance, public programmes, institutional leadership and market evidence-kept separate from valuations, forecasts and announced capacity.

The short versionCapital and institutional readout

The frontier is concentrated in a few heavily financed US labs, with Google and Meta funding research from corporate balance sheets and Mistral anchoring Europe's independent model effort. Capital and compute have scaled much faster than evidence of robust general intelligence: no lab has demonstrated the page's cross-domain, long-horizon threshold.

Players

Who is building itCompanies, laboratories and programmes

OpenAI

USA

Frontier multimodal and reasoning models, agent products and large-scale training infrastructure.

Anthropic

USA

Frontier Claude models with interpretability, alignment and model-safety research.

xAI

USA

Grok model family and vertically integrated large-scale training infrastructure.

Companies, laboratories and programmes working on this problem
PlayerCountryWhat they are doingFundingNamed investorsSource
OpenAIcompanyUSA

Frontier multimodal and reasoning models, agent products and large-scale training infrastructure.

Committed capital is not cumulative funding, revenue or money already spent on compute.

$122B committed capital · Mar 2026 · $852B post-money valuationAmazon · NVIDIA · SoftBank · MicrosoftSource · openai.com
AnthropiccompanyUSA

Frontier Claude models with interpretability, alignment and model-safety research.

The reported $5B August 2025 revenue figure is an annualized run-rate, not audited full-year revenue.

Series F · $13B · Sep 2025 · $183B post-money valuationICONIQ · Fidelity · Lightspeed · BlackRock affiliates · Qatar Investment AuthoritySource · anthropic.com
xAIcompanyUSA

Grok model family and vertically integrated large-scale training infrastructure.

No cumulative total is encoded because the release does not reconcile equity, debt and infrastructure financing.

Series E · $20B · Jan 2026Valor Equity Partners · Fidelity · Qatar Investment Authority · NVIDIA · CiscoSource · x.ai
Mistral AIcompanyFrance

European frontier and open-weight models, enterprise deployment and sovereign AI tooling.

ASML invested €1.3B; valuation is stated separately from capital raised.

Series C · €1.7B · Sep 2025 · €11.7B post-money valuationASML · Bpifrance · NVIDIASource · presse.bpifrance.fr
Google DeepMindlabUK / USA

Gemini frontier models, reinforcement learning, AI-for-science and robotics research.

Funded inside Alphabet; corporate capex is not a standalone lab round.

Not disclosedNot disclosedSource · deepmind.google
Meta AIlabUSA

Open-weight Llama models and research in multimodal, world-model and embodied AI.

Funded inside Meta; company-wide infrastructure spending is not lab funding.

Not disclosedNot disclosedSource · ai.meta.com
07 · Sources

Where every number comes from

4 sources — every figure on this page traces to one.