Benchmark Almanac

Opus Magnum Bench

GPT-5.6 Sol 25.5 leads since July 2026
Not in model cards

Agents build machines on a hexagonal grid to make a target molecule. Each puzzle score compares the machine’s combined cost, cycles, and area with the human record.

The makers compare 20 models on 36 campaign puzzles. Unsolved puzzles earn zero. Matching the human record earns full credit.

Released
June 2026
Kind
Puzzles and chess
Level
Competition
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text → Actions
Interaction
Agentic
Environment
Opus Magnum
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
In views
Games, Agents
Strongest model tested
Claude Fable 5 #12 on the capability index
Last model added
July 2026
Scores refreshed
November 20th, 2026
20
Models scored on its own leaderboard

GPT-5.6 Sol leads with 25.5 among models released by July 2026

Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-5.6 Sol xhigh 25.5
  2. 2 Claude Fable 5 high 21.6
  3. 3 GPT-5.5 xhigh 19.4
  4. 4 GPT-5.6 Terra xhigh 17.2
  5. 5 Gemini 3.1 Pro high 11.2
  6. 6 Claude Sonnet 5 high 9.56
  7. 7 GLM-5.2 high 9.14
  8. 8 Gemini 3.5 Flash high 8.82
  9. 9 Claude Opus 4.8 high 8.21
  10. 10 Grok 4.5 high 6.55
  11. 11 GPT-5.6 Luna xhigh 6.41
  12. 12 Gemini 3 Flash high 5.54
  13. 13 DeepSeek-V4-Pro 4.82
  14. 14 Claude Sonnet 4.6 medium 4.45
  15. 15 Kimi K2.7 Code 3.93
  16. 16 GLM-5.1 max 3.88
  17. 17 GPT-5.4 Mini xhigh 3.32
  18. 18 GPT-5.4 Nano xhigh 2.35
  19. 19 Kimi K2.6 1.98
  20. 20 Grok 4.3 high 0.00

Source: Official Opus Magnum Bench leaderboard

#ModelScoreSettingDateSource
1 GPT-5.6 Sol 25.5 xhigh Jul 2026 Official leaderboard
2 Claude Fable 5 21.6 high Jun 2026 Official leaderboard
3 GPT-5.5 19.4 xhigh Apr 2026 Official leaderboard
4 GPT-5.6 Terra 17.2 xhigh Jul 2026 Official leaderboard
5 Gemini 3.1 Pro 11.2 high Feb 2026 Official leaderboard
6 Claude Sonnet 5 9.56 high Jun 2026 Official leaderboard
7 GLM-5.2 9.14 high Jun 2026 Official leaderboard
8 Gemini 3.5 Flash 8.82 high May 2026 Official leaderboard
9 Claude Opus 4.8 8.21 high May 2026 Official leaderboard
10 Grok 4.5 6.55 high Jul 2026 Official leaderboard
11 GPT-5.6 Luna 6.41 xhigh Jul 2026 Official leaderboard
12 Gemini 3 Flash 5.54 high Dec 2025 Official leaderboard
13 DeepSeek-V4-Pro 4.82 Apr 2026 Official leaderboard
14 Claude Sonnet 4.6 4.45 medium Feb 2026 Official leaderboard
15 Kimi K2.7 Code 3.93 Jun 2026 Official leaderboard
16 GLM-5.1 3.88 max Apr 2026 Official leaderboard
17 GPT-5.4 Mini 3.32 xhigh Mar 2026 Official leaderboard
18 GPT-5.4 Nano 2.35 xhigh Mar 2026 Official leaderboard
19 Kimi K2.6 1.98 Apr 2026 Official leaderboard
20 Grok 4.3 0.00 high Apr 2026 Official leaderboard

Rankings by part

We read 6 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Reward (avg), Solved (pass@k) and Avg steps. GPT-5.6 Terra leads 3 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Reward (avg)

20 models we track have a score here.

  1. 1 GPT-5.6 Sol 21.5
  2. 2 Claude Fable 5 17.0
  3. 3 GPT-5.5 14.8
  4. 4 GPT-5.6 Terra 12.2
  5. 5 Gemini 3.1 Pro 8.16
  6. 6 Claude Sonnet 5 7.35
  7. 7 GLM-5.2 6.43
  8. 8 Gemini 3.5 Flash 5.57
  9. 9 Claude Opus 4.8 5.56
  10. 10 GPT-5.6 Luna ↑1 4.62

More effort adds a median 6 score

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Claude Fable 5
low high
21.6
GPT-5.5
medium xhigh
19.4
Claude Opus 4.8
high
8.21

Source: Official Opus Magnum Bench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5.6 Sol xhigh 25.5 385,225
Claude Fable 5 high 21.6 756,266
GPT-5.5 xhigh 19.4 244,835
Claude Fable 5 low 17.7 358,886
GPT-5.6 Terra xhigh 17.2 382,336
GPT-5.5 medium 13.6 102,698
Gemini 3.1 Pro high 11.2 89,945
Claude Sonnet 5 high 9.56 228,239
GLM-5.2 high 9.14 592,156
Gemini 3.5 Flash high 8.82 121,764
Claude Opus 4.8 high 8.21 40,493
Grok 4.5 high 6.55 479,775
GPT-5.6 Luna xhigh 6.41 788,601
Gemini 3 Flash high 5.54 172,286
DeepSeek-V4-Pro 4.82 407,001
Claude Opus 4.8 4.56 29,321
Claude Sonnet 4.6 medium 4.45 23,035
Kimi K2.7 Code 3.93 172,793
GLM-5.1 max 3.88 283,971
GPT-5.4 Mini xhigh 3.32 999,663
GPT-5.4 Nano xhigh 2.35 959,796
Kimi K2.6 1.98 342,876
Grok 4.3 high 0.00 56,741

3 models have held number one on the official leaderboard since its release in June 2026

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI Anthropic
Own leaderboard Not tested since Jul 2026

Source: Official Opus Magnum Bench leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard
0.00 10.0 20.0 30.0 2026 0.00 10.0 20.0 30.0 2026

Source: Official Opus Magnum Bench leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Games and long tasks

All 261 in Games and long tasks
  • ARC-AGI-3 In use Interactive game environments where agents must explore, infer goals and rules, and act efficiently without instructions. GPT-6.2 Astra 100%
  • Orak Fading LLM agents play 12 popular video games across most genres through an MCP interface. Each game scores in-game progress or outcomes, and the board ranks models by their average rank across the games. Gemini 2.5 Pro 3.90
  • MineDojo Fading Agents complete open-ended tasks in Minecraft, such as mining blocks and fighting mobs, from a large simulation suite. Programmatic checks score many tasks. Seed2.0 Pro 49%
  • TextQuests Only in system cards Agents play 25 classic interactive fiction games. The score is harmful actions per 100 steps, where lower is better. The board also reports game progress. Llama 4 Scout 7.00
  • GameWorld Only in system cards Multimodal game agents play 34 browser games across 170 tasks. Outcomes are checked from the game state. GPT-6.1 Terra 76.8%
  • Swarm Bench Only in system cards Models orchestrate agent swarms on complex tasks that benefit from coordinated decomposition and parallel execution. Kimi K3 76.3%

Data audit

All 9 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
  • Pass Independent runs Scores from the official leaderboard (20 entries).
  • Pass Models scored 20 models we track.
  • Pass Newest score The newest is from July 2026, for GPT-5.6 Luna.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Claude Fable 5 (#12).
  • Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
  • Pass Importer We read the board on opusmagnumbench.com: 23 rows for models we track at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check