205
PlanBench tasks GPT-5 solved across eight domains with standard descriptions; LAMA 204, DeepSeek R1 157, Gemini 155

What the chart shows

PlanBench asks models to write a complete plan for a structured problem, such as stacking blocks or moving packages, with up to 45 tasks in each of eight domains. Across all eight, GPT-5 solved 205 tasks and LAMA, a classical planning program, 204. DeepSeek R1 solved 157 and Gemini 155.

No planner led everywhere. LAMA led or tied in six domains, including all 45 Miconic tasks. GPT-5 solved all 45 Spanner tasks, against LAMA’s 15, and GPT-5 and Gemini both solved 38 Childsnack tasks.

Key numbers

MeasureValue
GPT-5, all eight domains205
LAMA, all eight domains204
DeepSeek R1, all eight domains157
Gemini, all eight domains155
GPT-5, Spanner45
LAMA, Miconic45

How to read it

  • Totals are sums across the eight domains, calculated by us from the data.
  • LAMA searches systematically for a plan; the language models write plans from the problem description. The source data labels the Google model only as Gemini.
  • Source study: Corrêa et al. (2025), arXiv 2511.09378.
Source: Corrêa et al., 2025, via the AI Index Report 2026 (Stanford HAI), figure 2.4.9: PlanBench: task solved on standard planning. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart