Can AI do your job yet?
As of October 2026, AI beats professionals in all 5 head-to-head work tests from the last three months, in science, medicine and finance. The tests measure finished work, not jobs lost.
head-to-head work tests where AI beats professionals. On real lawyer tasks, Gemini 4 Argon gets 1 in 4 right.
Software
Builds whole programs on its own. Passes every required check on about 6 in 10 real code changes.
Can do
Can't do yet
Maths
Has solved problems that professional mathematicians tried and failed to solve. Most famous Erdős problems still stand.
Can do
5 by AI alone
Can't do yet
Science
Beats PhD experts on hard science questions and on fixing failed lab work. Still behind Vals' reference results on AI research tasks.
Can do
Can't do yet
Medicine
Beat doctors at answering clinicians' real questions.
Can do
Can't do yet
No test yet shows doctors ahead
Law
Gets 1 in 4 real lawyer tasks right when it must pass every check.
Can do
No test with a lawyer score yet
Can't do yet
Finance and accounting
Ahead of junior accountants at month-end close, and of Wall Street analysts at forecasting revenue.
Can do
Can't do yet
How close is AI to recursive self-improvement? →
In the jobs most exposed to AI, employment of 22-25-year-olds is 19% behind similar young workers in other jobs.
At Anthropic, junior technical hires fell from 27% to 9%. AI agents now work 6 hours for every hour employees work.
No test yet compares AI with designers, marketers, video editors or architects.
| Field | Task | Model | AI score | People | People score | Who's ahead | Tested |
|---|---|---|---|---|---|---|---|
| Software | Rebuild a whole program without seeing its code | Claude Opus 5.5 | 77% | — | — | — | Sep 2026 |
| Software | Write code changes that pass every required check | Claude Opus 5.5 | 60% | — | — | — | Sep 2026 |
| Software | Make experienced developers faster | — | — | — | — | Not proven yet | 2025 and 2026 |
| Maths | Solve problems mathematicians could not | Several AIs | 10 of 49 | — | — | 10 of 49 solved · 5 by AI alone | Oct 2026 |
| Maths | Solve famous Erdős problems | GPT-6.1 Sol | 3 of 68 | — | — | 3 of 68 solved | Oct 2026 |
Show all 13
| Field | Task | Model | AI score | People | People score | Who's ahead | Tested |
|---|---|---|---|---|---|---|---|
| Science | Answer PhD-level science questions | Gemini 4 Argon | 96% | PhD experts | 70% | AI ahead by 26 points | Oct 2026 |
| Science | Work out why a virus experiment failed | GPT-5.6 Sol | 49% | Virologists | 21% | AI ahead by 28 points | Sep 2026 |
| Science | Match Vals' reference results on AI research tasks | Claude Opus 5.5 | 37 | Vals' reference results | 50 | — | Oct 2026 |
| Medicine | Answer doctors' real clinical questions | GPT-6 Astra | 65 | Doctors | 44 | AI ahead by 21 points | Sep 2026 |
| Law | Finish a real lawyer task and pass every check | Gemini 4 Argon | 27% | — | — | — | Oct 2026 |
| Finance and accounting | Close the books at month end | Claude Opus 5 | 100% | Junior accountants | 37% | AI ahead by 63 points | Oct 2026 |
| Finance and accounting | Forecast company revenue | GPT-6 Astra | — | — | — | AI ahead | Oct 2026 |
| Finance and accounting | Gets every line right on 7 in 10 tax returns | Meta Muse Wonder | 71% | — | — | — | Oct 2026 |
Sources (17)
- Epoch AI, epoch.ai/MirrorCode
- Mercor, mercor.com/apex/partner-benchmarks/frontiercode-leaderboard
- METR, We are Changing our Developer Productivity Experiment Design
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
- Epoch AI, epoch.ai/frontiermath/open-problems
- Epoch AI, FrontierMath Erdős (Benchmarking Hub data), Apr 2026
- Epoch AI, epoch.ai/benchmarks/gpqa-diamond
- SecureBio, securebio.org/blog/introducing-vct-v2
- Vals AI, Vals RSI Index (Vals AI, updated 15 Oct 2026), Oct 2026
- OpenAI, Faked data in a test task, Sep 2026
- arXiv, arxiv.org/abs/2604.27470
- Vals AI, vals.ai/benchmarks/hlab
- Mercor, mercor.com/blog/human-baselines-for-benchmarks-ai-now-outperforms-junior-accountants
- Samaya AI, samaya.ai/blog/frontier-ai-models-outperform-human-experts-on-earnings-prediction
- Column Tax, taxcalcbench.ai
- Stanford Digital Economy Lab, digitaleconomy.stanford.edu/news/canariesaug26
- Anthropic, anthropic.com/research/employment-at-anthropic