Atlas · C15 · Executive mind
Planning & organising
Breaking goals into steps, sequencing, adapting the plan when things change.
AI Reliable score
4.5 / 10
Average adult level · peak 7 · trend ↑↑ · grade B
Why this score
Short structured plans: agents now beat the human baseline on desktop tasks (OSWorld-Verified ~86-90% vs 72%) and near-saturate customer-service tool use. Long horizon: in a simulated year of running a business the best model earned ~$11k vs ~$63k for a good human strategy. Consistency drops sharply on repeated attempts.
On the human scale
2Beginner. Follows given steps
4Average adult. Average adult: plans a day or trip, adapts to changes ← AI today
6Competent pro. Professional: plans and runs multi-week projects with dependencies
8Expert. Expert: runs complex programmes with many people and resources
10Best humans. Best human planners
Future warning
At its best, AI already reaches 7 on this capability. Peak results are never counted in the index; they show where the water may rise next.
Evidence
OSWorld-Verified leaderboardB
E16 · 2026-10
- Measures
- Real desktop computer tasks (369 tasks, 100-step limit)
- AI result
- Top 90.2% (Intelligence-Indeed Agent); Claude Fable 5 86.0%; Claude Opus 5 83.4%.
- Humans
- Humans 72.36%
- On our scale
- Above the human baseline on short structured computer tasks.
presenc.ai ↗
Vending-Bench 2 (Andon Labs)B
E17 · 2026
- Measures
- Running a simulated vending business for a year
- AI result
- Best: Claude Opus 5 ~$11,182 average final balance from $500.
- Humans
- Good human strategy estimated ~$63k per year (~10x the best model)
- On our scale
- Long-horizon planning and judgment far below a competent human: key limit on Reliable.
epoch.ai ↗
Artificial Analysis benchmark snapshot (via modelgrep)B
E14 · 2026-11
- Measures
- Independent re-runs of several benchmarks
- AI result
- Long-Context Reasoning 88.7% (Kimi K3); Tau2-Bench 99.1%; IFBench 82.9%; Terminal-Bench Hard 65.9%.
- Humans
- Not published
- On our scale
- Supporting evidence for working memory, language and short agentic plans.
modelgrep.com ↗
tau-bench pass^k (Sierra)B
E18 · 2024 (models of 2024)
- Measures
- Consistency: succeeding on every one of k repeated attempts
- AI result
- Success fell by about a third from pass@1 to 4 repeated attempts (e.g. 0.69 to 0.46).
- Humans
- Not applicable
- On our scale
- Shows why 'Reliable' is lower than 'Peak'. Dated: needs a 2026 re-check.
beancount.io ↗
History
First reading: Atlas v0.1, 5 Nov 2026. Every later change will be listed here with its reason.
Disagree? Send a correction
