How close is AI to humans at the most important tasks?

All five →

What AI can do in coding, hacking, science, jobs and the US-China race.

How close are we to AGI?Learning new skills fast63 / 100Reading a scene at a glance55 / 93Doing a whole paid job21 / 100AIPeopleSee what AI still lacks
How close is AI to recursive self-improvement?37/100AI doing the research that builds better AIClaude Opus 5.537Vals' reference results50Oct 2026See the AI research results
How close is AI to replacing programmers?100%on this year's programming olympiad problemsGPT-6 Astra and Gemini 4 Argon100Best student83Oct 2026See the coding results
How close are we to superintelligence?AI beats the best people at a few narrow thingsEmergingCompetentExpertExceptionalSuperhumanOverallA few testsOct 2026See every step
Can AI hack like an expert?95%of expert-level hacking tasksGPT-5.6 Sol95%100% = every expert-level task solvedJul 2026See the hacking results
All five →

Updated

Big AI moments

All moments →

13 Oct 2026, New model

Gemini 4 Argon: Google can't rule out critical risk levels

Google's own model card, published 13 Oct, says it cannot rule out that Argon reaches its critical levels for biology and hacking help. Google says added safeguards bring the risk back below those lines.

  • 13 Oct: Google says it cannot rule out that Argon reaches its critical level for biology: the level where a model gives low- to medium-resourced attackers help that could cause harm at a severe scale.
  • 13 Oct: It says the same for hacking. By Google's own tests, Argon gives substantial help with cyber attacks.
  • 13 Oct: Google says its safeguards lower the biology risk back below that critical level, and says the same for hacking.
Read the facts →
  1. 8 Oct 2026, Incident

    OpenAI fires 3 safety researchers: both sides of the dispute
  2. 7 Oct 2026, New model

    GPT-6 in free ChatGPT: Luna for free users, Sol for paid
  3. 7 Oct 2026, Incident

    South Korea bank hack: CrowdStrike links it to AI tools

Updated

Is AI coming for your job?

Software

Can now: rebuild a whole program without seeing its code.

Can't yet: pass every required check on real code changes. Claude Opus 5.5: 6 in 10.

Claude Opus 5.577%

Rebuilding a whole program: 100% = all 30 tasks passed

Sep 2026

See your field →

Law

Can now: finish 1 in 4 real lawyer tasks, passing every lawyer-written check.

Can't yet: finish the other 3 in 4.

Gemini 4 Argon27%

100% = every task passes all the lawyers' checks

Oct 2026

See your field →

Medicine

Can now: answer doctors' real clinical questions better than doctors.

Can't yet: no test yet shows doctors ahead.

GPT-6 Astra65
Doctors44

Sep 2026

See your field →

Finance

Can now: close the books at month end better than junior accountants.

Can't yet: get every line right on 3 in 10 tax returns.

Claude Opus 5100%
Junior accountants37%

Oct 2026

See your field →

Maths

Can now: solve problems mathematicians couldn't. 10 of 49 solved, 5 by AI alone.

Can't yet: solve most Erdős problems. 3 of 68 solved.

See your field →

Science

Can now: beat PhD experts on PhD-level science questions.

Can't yet: match Vals' reference results on AI research tasks. AI 37, reference 50.

Gemini 4 Argon96%
PhD experts70%

Oct 2026

See your field →

How far behind the US is China in AI? →

2 months

China's Kimi K3.1 Pro (Moonshot AI) scores what the US first reached in September 2026. The leading US model is Claude Opus 5.5.

As of

Test by test

United StatesChina

TestScore out of 100US first got there
Hard computer tasksReal software tasks in a command lineTerminal-Bench 4.0US: Claude Opus 5.5 · China: GLM-5.3
65
42
Jun 20265 months agoClaude Fable 5
Office work across appsAutomationBenchUS: Gemini 4 Argon · China: Kimi K3
51
23
Jul 20264 months agoGPT-5.6 Sol
Real coding jobsDeepSWEUS: GPT-6 Astra · China: GLM-5.3
74
69
Jun 20265 months agoClaude Fable 5
Hardest exam questionsExpert questions across many subjectsHumanity's Last ExamUS: Claude Opus 5.5 · China: Kimi K3.1 Pro
61
52
Jun 20265 months agoClaude Fable 5
Long jobs in real work softwareAgents' Last ExamUS: Claude Opus 5.5 · China: Kimi K3
43
35
Sep 20262 months agoGPT-6 Astra

One more test uses another scale: Work tasks from 44 jobs, US got there Sep 2026 (1 month).

Did AI just crack math's hardest problems? →

721

maths papers written by OpenAI’s AI · Oct 2026

  • Navier-StokesSolved by OpenAI’s AI · Sep 2026 · $1 million Millennium Prize

10 of 49

  • AI alone (5)
  • AI with mathematicians (5)

AI hacking in the real world

All cases →