Methodology

How we run and score models on our RL environments.

Version 17, updated 4 November 2026.

The environments: Each environment starts from real records. Data arrive on a clock, and the agent works with tools: it reads the records, runs code, asks people and acts. Each environment is a task with its own grader.
Runs: Every model runs in the same harness, with the same tools and budget, at a high reasoning setting. Each model runs several times on each environment, and a model sees a task fresh each time.
Grading: Each answer key is checked against the records. Numeric checks need no judge. Two judges from different labs compare each written answer with the key from the records, and the score is their mean.
Controls: Each task's reference answer scores 100% under both judges. Wrong answers, including answers that take a shortcut, score 10% or less.
Scores: A model's score is its mean over the environments of a family, from 0% to 100%, with its 95% interval.

Changes

31 October 2026Version 1. First public leaderboard, with 13 frontier and open models.
1 November 2026Version 2. Added Muse Wonder on all three families.
1 November 2026Version 3. Added Claude Haiku 5.5 and DeepSeek-V4.1-Flash on Industry and on Science and compute.
1 November 2026Version 4. Added DeepSeek-V4-Pro on Personal agents.
2 November 2026Version 5. Added Qwen3.8-Flash-Next on Personal agents.
2 November 2026Version 6. Added MAI-Thinking-1 on Industry and on Science and compute.
2 November 2026Version 7. Added Nemotron 3 Ultra on Personal agents.
2 November 2026Version 8. Added Mistral Large 4 on Personal agents.
2 November 2026Version 9. Added Mistral Large 4 on Industry and on Science and compute.
2 November 2026Version 10. Added Command A+ on Industry and on Science and compute.
2 November 2026Version 11. Added Beam 1.0 on Industry and on Science and compute.
2 November 2026Version 12. Added Beam 1.0 on Personal agents.
2 November 2026Version 13. Added Command A+ on Personal agents.
3 November 2026Version 14. Added MAI-Thinking-1 on Personal agents.
3 November 2026Version 15. Added GPT-6.1 Terra and GPT-6 Luna on Personal agents.
3 November 2026Version 16. Added GPT-6.1 Terra and GPT-6 Luna on Industry and on Science and compute.
4 November 2026Version 17. Updated Mistral Large 4 and Command A+ on Industry and on Science and compute.

Cite

@misc{reific_rl_environments_2026,
  title  = {Reific RL environments leaderboard},
  author = {{Reific}},
  year   = {2026},
  note   = {Version 17, updated 4 November 2026},
  url    = {https://reific.co/environments/methodology}
}

The scores are on each family's page: Personal agents, Industry and Science and compute.

Try them on your models.

Talk to a researcher