Reality check
Is it actually super intelligence?
Renaming the technology doesn’t change what it can do. These are the strongest public measures of how close AI is to beating people at almost everything, each with its source and date.
Four steps, tracked separately
We don’t publish a single “distance to super intelligence” number, because no test can measure it yet. Instead we track four steps. Today: 1 of 4 met, 2 partly met. As of October 13.
- Step 1: Does it match people on well-defined tests?MetARC-AGI-2 at 95%, ARC-AGI-3 at 99.9% against a human 100%, PhD-level science questions at 96.4% against 65 to 70% for PhD experts.
- Step 2: Can it finish long, real projects on its own?PartlyMETR's longest reliable task is at least 16 hours, but that page stopped updating on Sept 8. Rebuilding whole programs from scratch: 77% of tasks.
- Step 3: Does it solve problems no human has?Early signsAn unreleased OpenAI model disproved a 1946 Erdos conjecture in May. On Epoch's set of open Erdos problems, AI has solved 2 of 68.
- Step 4: Is it far beyond the best people in almost every field?Nobody can measure it yetNo test or survey measures this, and no forecasting market asks about superintelligence itself. This is the original meaning of the word.
Learning brand-new video-game-like environments as efficiently as a human (ARC-AGI-3)
AI agents are dropped into unfamiliar interactive games with no instructions and must work out the rules and goal; the score compares their action efficiency with humans.
99.9%
99.9% with the provider-adapter harness; 62.7% with the standard harness
Skilled human: 100%: humans solve every environment; the per-level human baseline is the median human action count. GPT-6 Astra used fewer actions than that baseline on 96% of levels.
Leader: GPT-6 Astra (OpenAI), ARC Prize semi-private set, verified by ARC Prize
ARC-AGI-3 launched Mar 25, 2026 with frontier AI at 0.51%; GPT-5.5 scored 0.43% on May 1, 2026; GPT-6 Astra reached 62.7% (standard) and 99.9% (provider adapter) on Sep 3, 2026. GPT-6.1 Sol scored 96.4% (provider adapter) on Sep 29, 2026.
ARC Prize, OpenAI's GPT-6 Astra on ARC-AGI-3, Sep 3 ↗Longest autonomous task AI can complete (METR time horizon)
How long a task, measured by how long a skilled human expert takes, an AI agent can finish about half the time on its own.
17.4 hours
At least 16 hours (point estimate 17.4 hours, 95% CI 8.5 to 55 hours)
Skilled human: Tasks are timed against skilled human experts. METR says measurements above 16 hours are unreliable with its current 228-task suite.
Leader: Claude Mythos Preview (early version), Anthropic, restricted release; evaluated March 2026
Up from 3.4 hours (GPT-5, Aug 2025) to 12.0 hours (Claude Opus 4.6, Feb 2026) to 17.4 hours (Mythos Preview, Apr 2026), roughly a doubling every 3 to 4 months. METR froze the page on Sep 8, 2026, so GPT-6 Astra, Opus 5.5 and Gemini 4 Argon have no METR number.
METR, Task-Completion Time Horizons of Frontier AI Models (page frozen Sep 8, 2026); data file benchmark_results_1_1.yaml, May 8 ↗AI-produced results in research mathematics
Whether AI systems have produced original, checkable results on problems mathematicians care about.
1946 Erdos unit-distance conjecture disproved by an unreleased OpenAI model (announced May 20, 2026); OpenAI later reported 10 more advances from Astra, including three more Erdos problems (Aug 1, 2026)
Skilled human: Human mathematicians improved on the unit-distance result within weeks.
Leader: OpenAI (unreleased internal model, then Astra); Google DeepMind resolved nine lesser Erdos problems
Before 2026 AI had no historically significant research proof. Quanta calls the unit-distance result the first historically significant proof to come from an AI model.
Quanta Magazine, Why the Legendary Erdos Problems Are Falling to AI, Aug 3 ↗Epoch Capabilities Index (single general-capability score)
A statistical combination of dozens of benchmarks into one scale so models can be compared even after individual tests saturate.
167.3 ECI points
167 (167.3, 95% CI 164.1 to 171.7)
Skilled human: None: the scale has no human reference point.
Leader: Claude Opus 5.5 (Anthropic), narrowly ahead of GPT-6 Astra (166.5) and GPT-6.1 Sol (166.1); intervals overlap
Up about 17 points in 12 months, from 150.3 (GPT-5 Pro, Oct 7, 2025). Scores are refit as data arrives: Astra fell 2.9 points in the 10 days after launch.
Epoch AI, AI benchmarks & capabilities (Epoch Capabilities Index), Oct 1 ↗PhD-level science multiple choice (GPQA Diamond)
198 graduate-level biology, chemistry and physics questions designed to be hard to look up.
96.4%
Skilled human: PhD domain experts scored 65% (74% after discounting clear mistakes) in the original paper; skilled non-experts with web access scored 34%. OpenAI later measured 69.7% for PhD experts on the Diamond subset.
Leader: Gemini 4 Argon (Google DeepMind), run by Epoch on Oct 13, 2026; GPT-6 Astra 95.8%
Up from 87.0% (Grok 4, Jul 2025). Saturated: the top nine models are within about 2 points of each other and all above the PhD-expert baseline.
Epoch AI benchmark data, gpqa_diamond.csv; benchmark page, Oct 13 ↗Largest AI data centers (electric power)
Epoch's satellite- and permit-based estimates of the electrical power available to AI computing hardware at the biggest sites.
946 MW IT power at the largest operating site
946 MW operating (Colossus 2); 2,263 MW planned for Microsoft Fairwater Wisconsin by 2028; 14.1 GW across 94 tracked sites
Skilled human: Not applicable. For scale Epoch notes San Diego draws about 800 MW.
Leader: Colossus 2 (SpaceXAI, Memphis) operating; Microsoft Fairwater Wisconsin planned
In Nov 2025 Epoch projected the first sites crossing 1 GW of total facility power in 2026 (Anthropic-Amazon New Carlisle in January, Colossus 2 in February). Epoch's IT-power and facility-power figures are different measures.
Epoch AI, Largest AI data centers ranked by power capacity, Oct 12 ↗All 14 measures
Novel visual reasoning puzzles (ARC-AGI-2)
Abstract grid puzzles that are easy for untrained people but were built to resist memorization.
95%
95.0%
Skilled human: Every task was solved by at least two of the 2 to 10 untrained people who attempted it; individual people completed 66% of test pairs and 75% by task (ARC-AGI-2 paper).
Leader: GPT-6 Astra (max reasoning), OpenAI
From 18.3% (GPT-5 Pro, Oct 2025) to 37.6% (Claude Opus 4.5, Dec 2025) to 95.0% (GPT-6 Astra, Sep 2026). Near saturation: GPT-6.1 Sol 94.2%, Claude Opus 5.5 93.3%.
ARC Prize, Results, Sep 2 ↗Hard math problems written by mathematicians (FrontierMath Tiers 1 to 3)
295 unpublished problems from undergraduate level up to exploratory problems for advanced graduate students.
93.7%
Skilled human: Not reported as a single human score; problems are written so a typical one takes a specialist hours of effort.
Leader: GPT-6.1 Sol and GPT-6 Astra (OpenAI), tied; run by Epoch AI
The best score on the earlier whole-benchmark set (v1) was 32.4% (GPT-5, Aug 2025). Epoch corrected errors in 42% of problems on Jun 12, 2026, so v1 and v2 are not strictly comparable.
Epoch AI benchmark data, frontiermath_tiers_1_3_v2.csv; hub page for GPT-6 Astra, Sep 29 ↗Research-level math problems (FrontierMath Tier 4)
43 exceptionally difficult problems that Epoch says can take a researcher multiple hours to multiple days.
100%
100% (GPT-6.1 Sol); 97.6% (GPT-6 Astra)
Skilled human: Epoch: a typical problem takes a researcher in the relevant field multiple hours, the hardest multiple days.
Leader: GPT-6.1 Sol (OpenAI), run by Epoch AI on Sep 29, 2026
Best score on the v1 Tier 4 set was 14.6% (GPT-5 Pro, Oct 2025). Now at the ceiling with only 43 problems (one problem is 2.3 points), so the benchmark no longer discriminates.
Epoch AI benchmark data, frontiermath_tier_4_v2.csv; hub page for GPT-6 Astra, Sep 29 ↗Open Erdos problems solved (FrontierMath Erdos)
68 unsolved problems chosen by mathematician Thomas Bloom, where AI must write a machine-checkable Lean proof or disproof.
2.9% of 68 open problems
2 of 68 (2.9%)
Skilled human: These problems were open as of August 2026: no human has solved them.
Leader: GPT-6 Astra, with Claude Opus 5.5 and Claude Sonnet 5.5 also at 2 of 68
No model solved any when the benchmark launched on Sep 1, 2026; three models have since solved two each.
Epoch AI, Announcing FrontierMath Erdos; benchmarks hub, Sep 3 ↗Expert-written exam questions across fields (Humanity's Last Exam)
2,500 hard questions written by subject experts across many fields, intended to be too hard for earlier AI.
54.8%
54.8% (plus or minus 1.9), no tools
Skilled human: No single human score is reported; questions come from expert contributors and reviewers.
Leader: GPT-6 Astra (OpenAI) on Scale AI's leaderboard
Up from 31.6% (GPT-5 Pro, Oct 2025). Other leaderboards (Artificial Analysis, LLM Stats, BenchLM) report higher numbers for Claude models under different settings, so compare only within one board.
Scale AI Labs, Humanity's Last Exam leaderboard, Sep 3 ↗Fixing real software bugs (SWE-bench Verified)
484 real GitHub issues from open-source Python projects; the AI must produce a fix that passes the tests.
83.5%
Skilled human: No human baseline.
Leader: Claude Opus 4.7 (Anthropic), the best of the models Epoch has run (evaluated Apr 20, 2026)
Up from 73.6% (GPT-5, Aug 2025). Epoch's newest run is a model released Jun 16, 2026, so no model released after mid-June 2026 is included.
Epoch AI, SWE-bench Verified benchmark page and data, Apr 20 ↗Rebuilding whole software programs from scratch (MirrorCode)
The AI must reimplement 15 real programs in two languages without seeing source code, with exact output matching, using up to 10 billion tokens and 7 days per attempt.
77.4% of 30 tasks
77.4% of tasks fully solved (plus or minus 8.3)
Skilled human: METR said an earlier task (a 16,000-line bioinformatics toolkit) would take a human engineer weeks.
Leader: Claude Opus 5.5 (Anthropic)
From 8.9% (Gemini 3.1 Pro, Feb 2026) to 31.1% (Claude Opus 4.7, Apr 2026) to 77.4% (Claude Opus 5.5, Sep 2026).
Epoch AI, MirrorCode (built with METR), Sep 22 ↗Largest AI training run (estimated total computation)
Estimated total floating-point operations used to train the model.
10²⁷ FLOP
about 1e27 FLOP (Epoch CI roughly 5e26 to 2e27)
Skilled human: Not applicable.
Leader: GPT-6 Astra (OpenAI), trained on at least 100,000 NVIDIA GB200 chips in Abilene, Texas, per Epoch's estimate
About 2x the largest training run Epoch lists from a year earlier: Grok 4 at about 5e26 FLOP (Jul 9, 2025).
Epoch AI, GPT-6 Astra model page, Sep 3 ↗When do people expect the real thing?
AI 2027 scenario : 2027
AI Futures Project : 2027
Metaculus: weakly general AI : 2027
Dario Amodei : 2028
Sam Altman : 2028
Demis Hassabis : 2030
Eli Lifland : 2030
Metaculus: general AI : 2031
AI Impacts Expert Survey : 2042
Each forecaster names a different milestone, from a “superhuman coder” to machines that beat people at every task. The list below gives each one in their own words, with the source.
- Metaculus community (about 1.8k forecasters): 'weakly general AI' questionCommunity median date for the first weakly general AI system to be built, tested and publicly announced is 25 Aug 2027. The question has not resolved, though on Oct 9, 2026 Metaculus said GPT-6 Astra beat Montezuma's Revenge, one of its criteria, and it is still checking the rest. Metaculus, When will the first weakly general AI system be devised, tested, and publicly announced?
- AI Futures Project (Kokotajlo, Lifland, Halstead)Reality is running at 70 to 90% of the pace in their AI 2027 scenario. At 75% pace an 'Automated Coder' (AI the leading lab would rather use than human programmers) arrives around mid-2027; at 60% pace, early 2028. AI Futures Project, Q2.5 2026 Timelines Update: Uplift and Revenue
- AI 2027 scenario (as scored by the independent AI 2027 Tracker)The scenario puts a 'superhuman coder' in March 2027; the tracker rates it 'not yet testable' and lists 13 of 66 predictions confirmed as of Oct 12, 2026. AI 2027 Tracker (independent, not affiliated with AI Futures Project)
- Sam Altman, OpenAI CEOSaid the world 'may be only a couple of years away from early versions of true superintelligence', adding 'we could be wrong' (India AI Impact Summit, Feb 19, 2026). Yahoo Tech, Sam Altman backs democratic AI (published Feb 23, 2026)
- Dario Amodei, Anthropic CEOHis hunch is a 'country of geniuses in a datacenter' in one to two years, and he is at 90% on it within ten years (Dwarkesh Patel interview). Dwarkesh Podcast, Dario Amodei: We are near the end of the exponential
- Eli Lifland, AI Futures Project (AI 2027 co-author)His all-things-considered median for a 'superhuman coder' AI is 2030, later than his model's own output of December 2028. AI Futures Project, Q2.5 2026 Timelines Update: Uplift and Revenue
- Demis Hassabis, Google DeepMind CEOExpects AGI around 2030, plus or minus a year, with 2029 possible (Google I/O 2026 interview with Axios). Axios, Google DeepMind CEO Demis Hassabis says we're close to AGI
- Metaculus community (about 1.9k forecasters): 'general AI' questionCommunity median date for the first general AI system to be built, tested and publicly announced is 1 Feb 2031. Metaculus, When will the first general AI system be devised, tested, and publicly announced?
- AI Impacts Expert Survey on Progress in AI (1,580 researchers, surveyed Dec 2024)Mean aggregate forecast gives a 50% chance of machines that can do every task better and more cheaply than humans (high-level machine intelligence) by 2042, down from 2047 in the 2023 survey; 10% chance by 2027. Full automation of all labor: 50% chance by 2098. AI Impacts, Advanced AI according to 1,580 researchers (ESPAI 2024)
- Mark Zuckerberg, Meta CEOPublished an essay on 'personal superintelligence' in August 2026 but, per TechCrunch, gives no arrival date. TechCrunch, Mark Zuckerberg's AI manifesto is exactly why people don't like AI