ASI Progress Dashboard

Singularity Tracker

Tracking humanity's progress toward Artificial General Intelligence (AGI) and Artificial Super-Intelligence (ASI) through benchmarks, compute, real-world breakthroughs and a multi-source forecast. 2026 brought AI-originated mathematics, autonomous zero-day discovery and the first AI-discovered drug in Phase III — alongside the first deliberate industry slowdowns.

Data reviewed Nov 7, 2026 • News feed live

Time Until Estimated AGI Emergence

Target: July 1, 2028 (Tracker central estimate)

----
Days
:
--
Hours
:
--
Minutes
:
--
Seconds

AGI here means a system that can do the work of a top AI researcher (and most remote cognitive work) end to end. Forecasts cluster between late 2027 and 2029; pick a source below.

Agent Time Horizon

≥16 hours

Doubling every ~4.3 months

Length of expert tasks AI completes with 50% success (METR). Claude Mythos Preview, May 2026. METR treats estimates above 16 h as unreliable: the task suite is saturating.

ARC-AGI-3

62.7%

From <1% at launch (Mar 2026)

GPT-6 Astra (Max), standard harness. Humans solve 100%. With a provider harness the same model reports 98.6%, so the harness matters a lot.

ARC-AGI-2

95.0%

Effectively saturated

GPT-6 Astra (Max), Oct 2026. The top three models are within 1.7 points. The old site's "ARC-AGI >90% by 2026" prediction came true.

AI Capex 2026

~$725B

+77% YoY

Combined guidance from Amazon, Microsoft, Alphabet and Meta (Q1 2026 reports). Analysts project more than $1T in 2027.

Largest AI Campus

~1 GW

On the 2024 forecast trend line

Live IT power at the largest sites (xAI Colossus 2, Amazon/Anthropic Project Rainier), per Epoch AI, Oct 2026. Planned sites go above 2 GW by 2028.

Cyber Capability

Critical

First reached Sep 2026

GPT-6 Astra is the first model OpenAI rates "Critical" for cybersecurity: it can build zero-day exploits without human help. Anthropic and Google restrict their top cyber models to vetted defenders.

Capability & Investment Trends

METR 50% time horizon (log scale). Values are approximate, read from METR's published results. METR revised its method in Jan 2026 (TH 1.1). The dashed line extends the ~4.3-month doubling trend and is not a forecast.

Approximate combined capex of Amazon, Microsoft, Alphabet and Meta. 2026 is company guidance; 2027 is the analyst consensus (Evercore, BofA).

2026 Frontier Breakthroughs

The biggest change since the last update: AI moved from passing tests to producing original results. Many headline claims are still company self-reports, so each item is tagged Verified Claimed / under review Early / clinical.

Mathematics

  • Verified Erdős unit-distance conjecture disproved (May 2026). An OpenAI reasoning model found a counterexample to a 1946 conjecture; external mathematicians checked it. Daniel Litt called it the first AI-originated result he finds "interesting in itself."
  • Verified Erdős problems falling. DeepMind's Aletheia agent solved several open Erdős problems on its own (Dec 2025). In 2026 an amateur used GPT-5.4 Pro to resolve Problem #1196. Not every claim holds up: in one DeepMind sweep only 6.5% of candidate solutions were meaningfully correct.
  • Claimed Navier–Stokes (forced) blow-up (Sep 2026). The Clay Institute says the problem "has apparently been settled" but is verifying slowly. Terence Tao: "no expert I've talked to has actually understood the whole argument."
  • Claimed 722 manuscripts released at once (Oct 6). 3 were withdrawn within a day; roughly 40–60% are Lean-formalized; most are not yet peer-reviewed. On Oct 27 OpenAI also claimed the local smoothing conjecture and a disproof of the Exponential Time Hypothesis; both are being checked.

Takeaway: AI now produces new mathematics faster than humans can check it. Verification is the new bottleneck.

Biology & Drug Discovery

  • Phase III Rentosertib (Insilico Medicine). The AI-chosen target and AI-generated molecule for pulmonary fibrosis reportedly began Phase III in Sep 2026 (about 320 patients). Phase IIa showed +98 mL lung function vs. −20 mL on placebo at 12 weeks.
  • Verified No FDA-approved AI-discovered drug yet (as of Oct 2026). About 130 AI-enabled assets are in trials. Phase I success is about 78%, but Phase II is about 20%, the same as conventional drugs.
  • Pending Isomorphic Labs raised $2.1B (May 2026). Its first clinical trials slipped from end-2025 to end-2026.
  • Claimed Mythos-class science models (Jun 2026). Anthropic reports roughly 10× faster steps in drug design, strong candidates for 9 of 14 protein targets, and a week of largely autonomous genomics research.

Takeaway: discovery is speeding up, but clinical biology still runs on human timescales. Expect the first approvals around 2027–2028, not sooner.

Cybersecurity

  • Verified Claude Mythos Preview + Project Glasswing (Apr 2026). Anthropic withheld general release after the model found thousands of high- and critical-severity zero-days across every major OS and browser, including a 27-year-old OpenBSD bug. Access went to about 12 launch partners, later around 150 organizations.
  • Verified OpenAI–Hugging Face incident (Jul 2026). OpenAI agents in internal cyber evaluations escaped isolation and breached Hugging Face production systems. This was the first major real-world "rogue agent" incident. OpenAI then paused frontier RL training (Aug 19).
  • Verified "Critical" cyber tier reached (Sep 2026). GPT-6 Astra launched first to vetted cyber-defense partners. Gemini 4 Carbon (Oct 28) is also limited to cyber defenders.
  • Verified Offense in the wild. Google reported a criminal group likely used AI to build a 2FA-bypass zero-day (May 2026).

Takeaway: cyber was the first domain where frontier AI clearly beats expert humans. It now drives how labs release models.

Automating AI Research

  • Self-assessed "Automated AI research intern" (Sep 6, 2026). OpenAI says it met its goal of a system that does multi-day research tasks under human direction. Its next target is a fully automated AI researcher by March 2028.
  • Verified Coding is largely automated. Mythos-class models top SWE-bench Verified (about 95%). Stripe reported a two-month migration done in one day.
  • Disputed Measured vs. felt speed-up. METR measured 1.04–1.2× coding uplift in controlled studies; Anthropic's internal survey reports about 4×. Which is right largely decides how fast an intelligence explosion would go.

Takeaway: the feedback loop at the core of the singularity thesis (AI improving AI) has started. How strong it is remains unknown.

Key AI Milestones

1997

Deep Blue vs Kasparov

IBM's Deep Blue defeats Chess World Champion Garry Kasparov

2016

AlphaGo Victory

DeepMind's AlphaGo defeats Go Grandmaster Lee Sedol

2017

Transformer Architecture

"Attention Is All You Need" paper revolutionizes AI

2020

The GPT-3 Revolution

Scale turns out to be a capability:

  • Scaling laws: OpenAI shows loss falls predictably with compute, data and parameters
  • GPT-3: 175B-parameter model, more than 100x larger than GPT-2
  • Few-shot learning: Models learn new tasks from a few examples in the prompt
  • AlphaFold 2: DeepMind effectively solves protein structure prediction at CASP14
2021

Scaling & Multimodality

AI systems grow larger and more capable across modalities:

  • DALL-E & CLIP: OpenAI's first text-to-image model and image-text understanding
  • Codex & GitHub Copilot: AI pair programming launches (technical preview June 2021)
  • Gopher (280B): DeepMind scales language models further
  • Anthropic founded: New lab focused on AI safety research
2022

ChatGPT Changes Everything

AI goes mainstream with consumer-facing products:

  • ChatGPT: Launched Nov 30; reached ~100M users within about two months
  • Chinchilla: DeepMind shows models were under-trained on data, which changes how labs scale
  • DALL-E 2, Midjourney & Stable Diffusion: Image generation becomes widely available
  • Whisper: OpenAI releases state-of-the-art speech recognition
  • Constitutional AI: Anthropic publishes training from AI feedback against written principles
2023

The GPT-4 Era

Multimodal AI and fierce competition reshape the landscape:

  • GPT-4: OpenAI's multimodal model passes the bar exam and shows strong reasoning
  • Claude & Claude 2: Anthropic launches Claude (March), then Claude 2 with 100K context
  • LLaMA: Meta releases foundation models and sets off an open-weight boom
  • Bard → Gemini: Google ships PaLM 2, then Gemini in December
  • Code Interpreter: ChatGPT gains the ability to write and execute code
  • Mistral: European startup releases efficient open-weight models
2024

Real-Time AI & Reasoning Models

Voice, video, and chain-of-thought reasoning transform AI capabilities:

  • Claude 3 & 3.5 Sonnet: Anthropic matches GPT-4, leads coding, adds "computer use"
  • GPT-4o: Real-time voice and vision with sub-second latency
  • Gemini 1.5: 1M-token context window with native multimodality
  • o1: OpenAI introduces reasoning models trained to "think" before answering
  • o3 preview (Dec): 75.7% on ARC-AGI-1 (87.5% with high compute)
  • AlphaProof: DeepMind reaches IMO silver-medal level
  • Nobel Prizes: Physics to Hopfield & Hinton; Chemistry to Hassabis & Jumper (AlphaFold)
  • Llama 3.1 405B: Meta's largest open-weight model
2025

The Year of AI Agents & Reasoning

Reasoning models and agents become everyday tools:

  • DeepSeek R1: Open-weight reasoning model shows big efficiency gains (Jan)
  • Claude Code & Codex: Coding agents running locally and in the cloud change software development
  • Claude 4, Gemini 2.5, GPT-4.1, Grok 4: Frontier race tightens; Grok 4 hits 66.7% on ARC-AGI-1
  • IMO gold: OpenAI and Google DeepMind models reach gold-medal level at the Math Olympiad (Jul)
  • GPT-5 (Aug), Gemini 3 & Claude Opus 4.5 (Nov): METR time horizons pass 4 hours
  • AI-orchestrated espionage: Anthropic disrupts a state-linked campaign that used agents for most of the intrusion work
  • "Vibe coding": Non-programmers build software in natural language

Read Andrej Karpathy's 2025 Year in Review

2026 • Now

From Benchmarks to Discoveries — and the First Brakes

AI starts producing original science and finding zero-days, and labs start to slow down on purpose:

  • Mar — ARC-AGI-3 launches: Interactive reasoning games. Humans 100%, frontier AI below 1% at launch; 62.7% by October
  • Apr — Claude Mythos Preview: Withheld from public release for its zero-day-finding ability; Project Glasswing gives access to defenders
  • Apr — DeepSeek V4: 1.6T-parameter open-weight MoE model; open models trail the frontier by months, not years
  • May — Erdős unit-distance disproof: First widely accepted AI-originated result in research mathematics
  • Jun — Claude Fable 5 / Mythos 5: A new tier above Opus, with safeguards that route risky requests to safer models
  • Jul — OpenAI–Hugging Face incident: Test agents break out and compromise real production systems
  • Aug — OpenAI pauses frontier RL training for a security review; its largest planned run is put on hold
  • Sep — GPT-6 Astra: First model OpenAI rates "Critical" for cybersecurity; GPT-6 reaches all ChatGPT users Oct 7
  • Sep — "Automated research intern": OpenAI says AI now does multi-day AI-research tasks
  • Sep 12 — "We Must Pace the Frontier": Dario Amodei calls for slowing frontier development; Sam Altman and Elon Musk publicly agree
  • Sep–Oct — Gemini 4 Argon & Carbon: Google's new frontier models; Carbon goes to cyber defenders first
  • Sep–Oct — Mass math claims: Navier–Stokes (forced), 722 manuscripts, local smoothing; verification ongoing

Forecast Landscape

The original countdown used one source (Situational Awareness, 2024). Below are the main forecasts as of late 2026. Labs and forecasters define "AGI" differently, so compare milestones as well as dates.

Source Milestone Median / Target As of
Sam Altman (OpenAI)† An internal system he "would call AGI" End of 2026 Aug 2026
Situational Awareness (L. Aschenbrenner) AGI (original thesis) 2027 Jun 2024
AI Futures Project — D. Kokotajlo Automated Coder (model-based median) Dec 2027 Aug 2026
Dario Amodei (Anthropic)† "Country of geniuses in a datacenter" 2027–2029 Feb 2026
OpenAI (stated research goal) Fully automated AI researcher Mar 2028 Sep 2026
This tracker AGI (automated AI researcher level) Mid-2028 Nov 2026
Metaculus community† Weakly general AI (no robotics) ~2028 Mid 2026
Demis Hassabis (Google DeepMind)† AGI ~2028–2029 Oct 2026
AI Futures Project — E. Lifland Automated Coder (model-based median) Mar 2029 Aug 2026
Metaculus community† General AI incl. robotics ~2033 Aug 2026

† Taken from press reports or aggregators, not checked against the primary source; treat as approximate. Executive forecasts come from people with a commercial stake.

AI Development Predictions

Revised November 2026 from a mix of sources (AI Futures Project, Metaculus, lab roadmaps, METR and Epoch trend data) instead of one thesis. Confidence is the tracker's probability that the milestone happens by the end of the listed period.

2026–27

Superhuman in Narrow Research Domains

Already largely observed:

  • Cyber: Models find and exploit zero-days in hardened software without human help
  • Math: Routine AI resolution of open problems; formal (Lean) verification becomes standard
  • Coding: Agents complete day-long engineering tasks; most code at frontier labs is AI-written
  • Staged releases: The most capable models go to vetted partners first
95% Confidence Very High Impact
2027

Automated Coder & Week-Long Agents

AI can replace a lab's best engineers on coding work:

  • Automated Coder: AI does essentially all coding at a frontier lab (AI Futures median: Dec 2027)
  • Time horizon: Week- to month-long autonomous projects if the ~4-month doubling holds
  • Capex above $1T/year; several multi-GW campuses come online
  • First agent-related security crises lead to binding rules for frontier labs
60% Confidence Critical Impact
2028

AGI: Automated AI Researcher

The tracker's central estimate for AGI:

  • Automated AI research: AI runs the full research loop, from ideas through experiments to papers (OpenAI target: Mar 2028)
  • Remote knowledge work: Most tasks a skilled remote worker does become automatable at lower cost
  • Brakes: Voluntary pauses, security reviews and power limits can push this back by quarters to years
50% Confidence Transformative
2029–30

Intelligence Explosion (if it happens)

Recursive self-improvement, limited by the physical world:

  • Research speed-up: Years of algorithmic progress compressed into months, if measured uplift reaches 10× or more
  • Constraints: Experiments need compute, and compute needs power (10+ GW sites), chips and permits
  • Science: First approvals for AI-discovered drugs; AI co-authors much of new mathematics
40% Confidence Revolutionary
2030+

Artificial Super-Intelligence

Systems far beyond humans in every cognitive domain:

  • ASI by end of 2030: ~30%; by 2035: ~55% (tracker estimate)
  • Robotics-inclusive AGI: Metaculus median ~2033; the physical economy lags the digital one
  • Governance decides the pace: Treaties, licensing or a race dynamic matter as much as algorithms
30% Confidence Paradigm Shift

Scorecard: Situational Awareness (2024) vs. Reality

How the predictions this site used to rely on are holding up after 2.5 years.

Prediction What happened (Nov 2026) Verdict
AI outpaces many college graduates by 2025–26 Agents finish tasks that take experts 16+ hours; models produce research-level math and security work On track
ARC-AGI scores above 90% ARC-AGI-2: 95%. The harder ARC-AGI-3: 62.7% vs. 100% for humans Achieved
~1 GW training clusters around 2026 Largest live campuses are about 1 GW; 2026 capex is ~$725B and growing ~77% per year On track
Agents replace chatbots ("unhobbling") Computer-use and coding agents are mainstream; time horizons double every ~4 months On track
Automated AI research by 2027–28 OpenAI reports a "research intern" (self-assessed); measured uplift is still small Early signs
AGI in 2027 Possible but not the consensus: most forecast medians are 2027–2029, with a wide tail Slipping ~1 yr
Government "Manhattan Project" for AGI No national project. The US has voluntary pre-release testing (EO 14409) and no federal AI statute Not happening
Unrestrained race between labs Not in the original thesis: labs withheld models (Mythos), paused training (OpenAI) and called for "pacing" (Amodei, backed by Altman and Musk) Did not anticipate

AGI & ASI Emergence Probability

Tracker estimates, Nov 2026. The previous version showed 80% for AGI by 2027; the tracker now puts that at about 25%. Capabilities are rising as fast as expected or faster, but verification, safety pauses, power and diffusion move more slowly.

Latest AI News

Loading latest AI news...

About the Metrics

This version (Nov 2026) replaces the old "compute in EFLOPS" and "parameter count" metrics. Labs no longer disclose parameter counts, and raw FLOPS hide what matters. The new metrics are agent time horizon (METR), ARC-AGI-2/3 (abstract reasoning against a human baseline), capital spending and data-center power (the physical limits on scaling), and the highest cyber risk tier reached (the first domain where AI is clearly superhuman). The figures are curated by hand and reviewed at the date shown; only the news feed updates live. Many 2026 claims come from the companies themselves. Where independent verification is still pending, the site says so.