Benchmark Almanac

ISA-Bench

Gemma 4 31B IT 60.9% leads since September 2026
Not in model cards

Models write programs for five games with restricted instruction sets. A verifier checks each program and gives feedback before the next attempt.

Microsoft Research India published results for 14 models on 92 tasks, including seven TIS-100 tasks. The authors report solves after one and five attempts.

Released
September 2026
Kind
Puzzles and chess
Level
University
Scored
Checked automatically
Answers
Code
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code → Code
Interaction
Agentic
Environment
Code sandbox
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
In views
Agents
Strongest model tested
Gemma 4 31B IT #101 on the capability index
Last model added
April 2026
Scores refreshed
November 20th, 2026
14
Entries in its paper, 10 models we track

Gemma 4 31B IT leads with 60.9% among models released by April 2026 Gemma 4 31B IT leads with 60.9% among models released by April 2026

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.

The newest model with a score here came out in April 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 Gemma 4 31B IT Up to five attempts with execution feedback 60.9%
  2. 2 Qwen3-32B Up to five attempts with execution feedback 31.5%
  3. 3 Qwen3-14B Up to five attempts with execution feedback 26.1%
  4. 4 Phi-4-reasoning Up to five attempts with execution feedback 26.1%
  5. 5 DeepSeek-R1-Distill-Qwen-32B Up to five attempts with execution feedback 25%
  6. 6 Devstral Small 2 Up to five attempts with execution feedback 23.9%
  7. 7 Qwen3-Coder-30B-A3B-Instruct Up to five attempts with execution feedback 22.8%
  8. 8 Qwen3-8B Up to five attempts with execution feedback 21.7%
  9. 9 Q2.5-Coder 32B Up to five attempts with execution feedback 14.1%
  10. 10 Granite 4.1 8B Up to five attempts with execution feedback 13%
  11. 11 Q2.5-Coder 14B Up to five attempts with execution feedback 7.61%
  12. 12 Codestral 22B Up to five attempts with execution feedback 6.52%
  13. 13 Phi-4 Up to five attempts with execution feedback 6.52%
  14. 14 DS-Coder-V2 16B Up to five attempts with execution feedback 5.43%

Source: The ISA-Bench paper

#EntryScoreSettingDateSource
1 Gemma 4 31B IT 60.9% Up to five attempts with execution feedback Apr 2026 Official leaderboard
2 Qwen3-32B 31.5% Up to five attempts with execution feedback Apr 2025 Official leaderboard
3 Qwen3-14B 26.1% Up to five attempts with execution feedback Apr 2025 Official leaderboard
4 Phi-4-reasoning 26.1% Up to five attempts with execution feedback Apr 2025 Official leaderboard
5 DeepSeek-R1-Distill-Qwen-32B 25% Up to five attempts with execution feedback Jan 2025 Official leaderboard
6 Devstral Small 2 23.9% Up to five attempts with execution feedback Dec 2025 Official leaderboard
7 Qwen3-Coder-30B-A3B-Instruct 22.8% Up to five attempts with execution feedback Jul 2025 Official leaderboard
8 Qwen3-8B 21.7% Up to five attempts with execution feedback Apr 2025 Official leaderboard
9 Q2.5-Coder 32B 14.1% Up to five attempts with execution feedback Official leaderboard
10 Granite 4.1 8B 13% Up to five attempts with execution feedback Apr 2026 Official leaderboard
11 Q2.5-Coder 14B 7.61% Up to five attempts with execution feedback Official leaderboard
12 Codestral 22B 6.52% Up to five attempts with execution feedback Official leaderboard
13 Phi-4 6.52% Up to five attempts with execution feedback Dec 2024 Official leaderboard
14 DS-Coder-V2 16B 5.43% Up to five attempts with execution feedback Official leaderboard
#ModelScoreSettingDateSource
1 Gemma 4 31B IT 60.9% Up to five attempts with execution feedback Apr 2026 Official leaderboard
2 Qwen3-32B 31.5% Up to five attempts with execution feedback Apr 2025 Official leaderboard
3 Qwen3-14B 26.1% Up to five attempts with execution feedback Apr 2025 Official leaderboard
4 Phi-4-reasoning 26.1% Up to five attempts with execution feedback Apr 2025 Official leaderboard
5 DeepSeek-R1-Distill-Qwen-32B 25% Up to five attempts with execution feedback Jan 2025 Official leaderboard
6 Devstral Small 2 23.9% Up to five attempts with execution feedback Dec 2025 Official leaderboard
7 Qwen3-Coder-30B-A3B-Instruct 22.8% Up to five attempts with execution feedback Jul 2025 Official leaderboard
8 Qwen3-8B 21.7% Up to five attempts with execution feedback Apr 2025 Official leaderboard
9 Granite 4.1 8B 13% Up to five attempts with execution feedback Apr 2026 Official leaderboard
10 Phi-4 6.52% Up to five attempts with execution feedback Dec 2024 Official leaderboard

Rankings by part

We read 10 parts from the benchmark’s paper, each ranking models on one part of the test, such as HRM solve@1, HRM solve@5 and SIC-1 solve@1. Gemma 4 31B IT leads 5 of them, and 4 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

HRM solve@1

10 models we track have a score here.

  1. 1 Gemma 4 31B IT 50%
  2. 2 Qwen3-8B ↑6 25%
  3. 3 DeepSeek-R1-Distill-Qwen-32B ↑2 22.2%
  4. 4 Qwen3-14B ↓1 22.2%
  5. 5 Qwen3-Coder-30B-A3B-Instruct ↑2 22.2%
  6. 6 Qwen3-32B ↓4 19.4%
  7. 7 Devstral Small 2 ↓1 16.7%
  8. 8 Granite 4.1 8B ↑1 8.33%
  9. 9 Phi-4 ↑1 2.78%
  10. 10 Phi-4-reasoning ↓6 0%

Number one over time

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

Google DeepMind
Benchmark paper Not tested since Sep 2026

Source: The ISA-Bench paper

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
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Gemma 4 31B IT Qwen3-32B DeepSeek-R1-Distill-Qwen-32B Phi-4 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Gemma 4 31B IT Qwen3-32B

Source: Official ISA-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

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 7 months ago, in April 2026, for Granite 4.1 8B. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the benchmark’s paper (14 entries).
  • Pass Models scored 10 models we track, plus 4 other entries on its leaderboard.
  • Warning Newest score The newest came 7 months ago, in April 2026, for Granite 4.1 8B.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Gemma 4 31B IT (#101).
  • Pass Headroom The best is 60.9%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 10 rows for models we track and 4 other entries 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