How close are we to superintelligence?

As of October 2026, AI beats every person at this year's programming olympiad (100% vs the best student's 83%). Taken across all tests together, it is still below skilled adults.

Where AI stands on the staircase

  • Reached
  • On a few problems
  • Not reached
EmergingAs good as or a bit betterthan an unskilled person4 of 7 reachedCompetentBetter than half of skilledadults7 of 9 reachedExpertBetter than 90% of skilledadults1 of 1 reachedExceptionalBetter than 99% of skilledadults3 of 4 reachedSuperhumanBetter than every person3 of 5 reached

Steps: Google DeepMind, Levels of AGI

Overall, AI beats people on some tests and trails on others. How close are we to AGI?

Emerging

As good as or a bit better than an unskilled person

Furniture assembly · Reached
83% vs 63% untrained volunteers
Claude Opus 5.5, Sep 2026
Trick questions · Reached
88% vs 84% ordinary people
Claude Opus 5.5, Sep 2026
Using a computer · Reached
86% vs 72% students new to the software
Claude Fable 5, Aug 2026
Picture puzzles · Reached
95% vs 60% average person
GPT-6 Astra, Sep 2026
Reading a clock · Not reached
83 vs 91 average person
Gemini 4 Argon, Oct 2026
New games, fast · Not reached
63 vs 100 human level; 99.9 on OpenAI's own setup
GPT-6 Astra, Sep 2026
Scene at a glance · Not reached
55 vs 93 people
Gemini 4 Argon, Oct 2026

Competent

Better than half of skilled adults

PhD science questions · Reached
96% vs 70% PhD experts
Gemini 4 Argon, Oct 2026
Doctors' questions · Reached
65 vs 44 doctors
GPT-6 Astra, Sep 2026
Month-end accounting · Reached
100% vs 37% junior accountants
Claude Opus 5, Oct 2026
World event forecasts · Reached
68.8 vs 68.5 superforecasters (median)
Google DeepMind, Oct 2026
Designing AI experiments · Reached
Beats the best of at least 2 experts per task
Claude Opus 5.5, Oct 2026
Kaggle contests · Reached
81.5th percentile real entrants
Gemini 3.1 agent, Mar 2026
Office work tasks · Reached
Wins or ties 85% against professionals with 14 years' experience
GPT-5.5, Apr 2026
Whole paid jobs · Not reached
21% vs 100% professional's standard
Claude Opus 5.5, Oct 2026
Translating novels · Not reached
Readers chose AI 11 of 30 vs a human translation
Sep 2026

Expert

Better than 90% of skilled adults

Virus lab problems · Reached
49% vs 21% expert virologists' average
GPT-5.6 Sol, Sep 2026

Exceptional

Better than 99% of skilled adults

Maths Olympiad · Reached
42 of 42 tied with the top 7 of 666 students
dots-note-3.0, Jul 2026
Coding contest rating · Reached
Only 2 of 49,384 active coders rated higher
Gemini 3 Deep Think, Feb 2026
University maths contest · Reached
Top 3 of 4,329 students
DeepSeek-v3.2-Speciale, Feb 2026
AI research · Not reached
37 vs 50 Vals' reference results
Claude Opus 5.5, Oct 2026

Superhuman

Better than every person

Programming olympiad · Reached
100% vs 83% best student
GPT-6 Astra and Gemini 4 Argon, Oct 2026
World coding finals · Reached
1st ahead of all human finalists
OpenAI, Jul 2026
Unsolved maths · On a few problems
10 of 49 solved 5 by AI alone, 5 with people's help
Several AIs, Oct 2026
Erdős problems · On a few problems
3 of 68 unsolved by people
GPT-6.1 Sol, Oct 2026
Navier-Stokes · Reached
Solved $1M Millennium Prize problem
OpenAI's AI, Sep 2026
Every test on the staircase
StepTestModelAI scorePeopleMonth
EmergingFurniture assemblyClaude Opus 5.583% vs 63%untrained volunteersSep 2026
EmergingTrick questionsClaude Opus 5.588% vs 84%ordinary peopleSep 2026
EmergingUsing a computerClaude Fable 586% vs 72%students new to the softwareAug 2026
EmergingPicture puzzlesGPT-6 Astra95% vs 60%average personSep 2026
EmergingReading a clockGemini 4 Argon83 vs 91average personOct 2026

Swipe sideways to see every column.

How close is AI to replacing programmers? →

Is AI coming for your job? →

Is AI better than humans in real life? →

The two ingredients of an intelligence explosion

Does AI research
50 = Vals' reference results
37 /100Claude Opus 5.5, Oct 2026
Improves AI models
48 = official models (more time and chips)
45 /100Claude Fable 5.1 · PostTrainBench, Sep 2026
The ingredients
IngredientTestModelScoreLineMonth
Does AI researchVals RSI IndexClaude Opus 5.53750 (Vals' reference results)Oct 2026
Improves AI modelsPostTrainBenchClaude Fable 5.14548 (official models, trained with more time and chips)Sep 2026
Sources (28)
  1. arXiv, arxiv.org/abs/2311.02462
  2. Epoch AI, epoch.ai/benchmarks/furniture-assembly
  3. SimpleBench, simple-bench.com
  4. OSWorld, os-world.github.io
  5. ARC Prize, arcprize.org/leaderboard
  6. ClockBench, clockbench.ai
  7. ARC Prize, OpenAI's GPT-6 Astra on ARC-AGI-3 (ARC Prize blog), with the ARC-AGI-3 launch paper as baseline
  8. Scale AI, labs.scale.com/leaderboard/hss
  9. Epoch AI, epoch.ai/benchmarks/gpqa-diamond
  10. OpenAI, Faked data in a test task, Sep 2026
  11. Mercor, mercor.com/blog/human-baselines-for-benchmarks-ai-now-outperforms-junior-accountants
  12. Forecasting Research Institute, forecastbench.org/leaderboards
  13. arXiv, arxiv.org/abs/2610.06824
  14. arXiv, arxiv.org/abs/2603.26499
  15. OpenAI, openai.com/index/introducing-gpt-5-5
  16. Scale AI, labs.scale.com/leaderboard/rli
  17. arXiv, arxiv.org/abs/2606.26040
  18. SecureBio, securebio.org/blog/introducing-vct-v2
  19. studio-dots-ai.github.io, studio-dots-ai.github.io/dots_imo_2026/en.html
  20. Google, blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think
  21. MathArena, matharena.ai/putnam
  22. Vals AI, Vals RSI Index (Vals AI, updated 15 Oct 2026), Oct 2026
  23. Vals AI, vals.ai/benchmarks/ioi
  24. The Decoder, the-decoder.com/openais-ai-beats-every-human-at-atcoder-a-top-competitive-programming-con…
  25. Epoch AI, epoch.ai/frontiermath/open-problems
  26. Epoch AI, FrontierMath Erdős (Benchmarking Hub data), Apr 2026
  27. OpenAI, openai.com/index/navier-stokes-solution
  28. PostTrainBench, posttrainbench.com