Benchmark Almanac

Agentick

GPT-6 Astra 0.63 leads since September 2026
Not in model cards

Agents solve procedurally generated gridworld tasks involving navigation, planning, reasoning, memory, generalization, and coordination. Scores normalize task performance between random and oracle reference policies.

The suite contains 37 tasks with several observation formats and difficulty levels. Published configurations vary in training, harness, modality, and seed count. These settings must accompany scores.

Released
May 2026
Kind
Puzzles and chess
Level
Everyday
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images → Actions
Interaction
Agentic
Environment
A simulator
Items from
Generated
Refresh
One fixed set
Run by
Anyone
In views
Agents, Vision, Robotics
Strongest model tested
GPT-6 Astra #4 on the capability index
Last model added
September 2026
Scores refreshed
November 20th, 2026

In the authors’ words

AI agent research spans a wide spectrum: from RL agents that learn from scratch to foundation model agents that leverage pre-trained knowledge, yet no unified benchmark enables fair comparison across these approaches. We present Agentick, a benchmark for sequential decision-making agents designed to evaluate RL, LLM, VLM, hybrid, and human agents on common ground and to power research on the fundamental challenges of sequential decision-making. Agentick provides 37 procedurally generated tasks across six capability categories, four difficulty levels, and five observation modalities, all exposed through a single Gymnasium-compatible interface. The benchmark ships with a Coding API, oracle reference policies for all tasks, pre-built SFT datasets, a composable agent harness, and a live leaderboard. An evaluation spanning 27 configurations and over 90,000 episodes reveals that no single approach dominates: GPT-5 mini leads overall at 0.309 oracle-normalized score while PPO dominates planning and multi-agent tasks; the reasoning harness multiplies LLM performance by 3-10x; and ASCII observations consistently outperform natural language. These findings highlight the substantial room for improvement that remains across all agent paradigms. Agentick's capability-decomposed, multi-modal design provides the empirical infrastructure needed to drive progress toward general autonomous agents, both as an evaluation framework and as a training ground for RL post-training of foundation models in truly sequential environments.

Abstract of Agentick: A Unified Benchmark for General Sequential Decision-Making Agents (arXiv:2605.06869, CC0)
16
Entries on its own leaderboard, 11 models we track

GPT-6 Astra leads with 0.63 among models released by September 2026 GPT-6 Astra leads with 0.63 among models released by September 2026

Best score per entry 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. 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 September 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-6 Astra ascii, high, 1 seed 0.63
  2. 2 Claude Opus 5 ascii, high, 1 seed 0.63
  3. 3 GPT-5.6 Sol ascii, high, 1 seed 0.58
  4. 4 GPT-5.6 Terra ascii, high, 1 seed 0.57
  5. 5 GPT-5.6 Luna ascii, high, 1 seed 0.55
  6. 6 Qwen3.5-4B (SFT-250k) ascii 0.45
  7. 7 Qwen3.5-4B (SFT-120k) ascii 0.35
  8. 8 GPT-5 mini ascii 0.31
  9. 9 PPO Dense (2M) rgb_array 0.29
  10. 10 Qwen3.5-4B ascii 0.23
  11. 11 PPO Dense (500k) rgb_array 0.23
  12. 12 Gemini 2.5 Flash-Lite Sep 2025, ascii 0.19
  13. 13 Qwen3.5-2B ascii 0.13
  14. 14 Qwen3.5-0.8B ascii 0.09
  15. 15 Qwen3-4B ascii 0.09
  16. 16 PPO Sparse (500k) rgb_array 0.07

Source: Official Agentick leaderboard

#EntryScore ± 95% CISettingReleasedRun onSource
1 GPT-6 Astra 0.63 0.5 to 0.8 ascii, high, 1 seed, markovian_reasoner Sep 2026 Sep 2026 Official leaderboard
2 Claude Opus 5 0.63 0.5 to 0.7 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
3 GPT-5.6 Sol 0.58 ±0.1 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
4 GPT-5.6 Terra 0.57 ±0.1 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
5 GPT-5.6 Luna 0.55 ±0.1 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
6 Qwen3.5-4B (SFT-250k) open weights 0.45 0.4 to 0.5 ascii, markovian_zero_shot May 2026 Official leaderboard
7 Qwen3.5-4B (SFT-120k) open weights 0.35 0.3 to 0.4 ascii, markovian_zero_shot May 2026 Official leaderboard
8 GPT-5 mini 0.31 ascii, MarkovianReasoner Aug 2025 Mar 2026 Official leaderboard
9 PPO Dense (2M) open weights 0.29 0.2 to 0.4 rgb_array Mar 2026 Official leaderboard
10 Qwen3.5-4B 0.23 0.2 to 0.3 ascii, markovian_reasoner Feb 2026 Mar 2026 Official leaderboard
11 PPO Dense (500k) open weights 0.23 ±0.1 rgb_array Mar 2026 Official leaderboard
12 Gemini 2.5 Flash-Lite Sep 2025 0.19 ±0.1 ascii, markovian_reasoner Sep 2025 Mar 2026 Official leaderboard
13 Qwen3.5-2B 0.13 ±0.1 ascii, markovian_reasoner Feb 2026 Mar 2026 Official leaderboard
14 Qwen3.5-0.8B 0.09 ±0 ascii, markovian_reasoner Feb 2026 Mar 2026 Official leaderboard
15 Qwen3-4B 0.09 ascii, markovian_reasoner Apr 2025 Mar 2026 Official leaderboard
16 PPO Sparse (500k) open weights 0.07 ±0 rgb_array Mar 2026 Official leaderboard
#ModelScore ± 95% CISettingReleasedRun onSource
1 GPT-6 Astra 0.63 0.5 to 0.8 ascii, high, 1 seed, markovian_reasoner Sep 2026 Sep 2026 Official leaderboard
2 Claude Opus 5 0.63 0.5 to 0.7 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
3 GPT-5.6 Sol 0.58 ±0.1 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
4 GPT-5.6 Terra 0.57 ±0.1 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
5 GPT-5.6 Luna 0.55 ±0.1 ascii, high, 1 seed, markovian_reasoner Jul 2026 Sep 2026 Official leaderboard
6 GPT-5 mini 0.31 ascii, MarkovianReasoner Aug 2025 Mar 2026 Official leaderboard
7 Qwen3.5-4B 0.23 0.2 to 0.3 ascii, markovian_reasoner Feb 2026 Mar 2026 Official leaderboard
8 Gemini 2.5 Flash-Lite Sep 2025 0.19 ±0.1 ascii, markovian_reasoner Sep 2025 Mar 2026 Official leaderboard
9 Qwen3.5-2B 0.13 ±0.1 ascii, markovian_reasoner Feb 2026 Mar 2026 Official leaderboard
10 Qwen3.5-0.8B 0.09 ±0 ascii, markovian_reasoner Feb 2026 Mar 2026 Official leaderboard
11 Qwen3-4B 0.09 ascii, markovian_reasoner Apr 2025 Mar 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 Generalization, Memory and Multi agent. Claude Opus 5 leads 4 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.

Generalization

11 models we track have a score here.

  1. 1 Claude Opus 5 ↑1 0.58
  2. 2 GPT-6 Astra ↓1 0.58
  3. 3 GPT-5.6 Luna ↑2 0.50
  4. 4 GPT-5.6 Terra 0.50
  5. 5 GPT-5 mini ↑1 0.44
  6. 6 GPT-5.6 Sol ↓3 0.42
  7. 7 Qwen3.5-4B 0.34
  8. 8 Gemini 2.5 Flash-Lite Sep 2025 0.29
  9. 9 Qwen3.5-2B 0.17
  10. 10 Qwen3.5-0.8B 0.14

Scores at each effort level

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
Qwen3.5-4B markovian_reasoner
language ascii
0.23
Qwen3.5-2B markovian_reasoner
language
0.13
Qwen3.5-0.8B markovian_reasoner
language ascii
0.09
Qwen3.5-4B markovian_zero_shot
language rgb_array
0.09
Qwen3-4B markovian_reasoner
language ascii
0.09
Gemini 2.5 Flash-Lite markovian_zero_shot
ascii
0.06
Qwen3.5-2B markovian_zero_shot
language ascii
0.06
Qwen3-4B markovian_zero_shot
language
0.02
Qwen3.5-0.8B markovian_zero_shot
language
0.02

Source: Official Agentick leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra ascii, high, 1 seed markovian_reasoner 0.63
Claude Opus 5 ascii, high, 1 seed markovian_reasoner 0.63
GPT-5.6 Sol ascii, high, 1 seed markovian_reasoner 0.58
GPT-5.6 Terra ascii, high, 1 seed markovian_reasoner 0.57
GPT-5.6 Luna ascii, high, 1 seed markovian_reasoner 0.55
GPT-5 mini ascii MarkovianReasoner 0.31
Qwen3.5-4B ascii markovian_reasoner 0.23
Gemini 2.5 Flash-Lite Sep 2025 ascii markovian_reasoner 0.19
Qwen3.5-4B language markovian_reasoner 0.18
Qwen3.5-2B ascii markovian_reasoner 0.13
Qwen3.5-2B language markovian_reasoner 0.12
Qwen3.5-0.8B ascii markovian_reasoner 0.09
Qwen3.5-4B rgb_array markovian_zero_shot 0.09
Qwen3-4B ascii markovian_reasoner 0.09
Gemini 2.5 Flash-Lite Sep 2025 language markovian_zero_shot 0.06
Qwen3.5-2B ascii markovian_zero_shot 0.06
Qwen3.5-0.8B language markovian_reasoner 0.06
Gemini 2.5 Flash-Lite Sep 2025 ascii markovian_zero_shot 0.05
Qwen3-4B language markovian_reasoner 0.05
Qwen3.5-2B language markovian_zero_shot 0.03
Qwen3.5-4B ascii markovian_zero_shot 0.02
Qwen3-4B ascii markovian_zero_shot 0.02
Qwen3.5-0.8B ascii markovian_zero_shot 0.02
Qwen3.5-4B language markovian_zero_shot 0.02
Qwen3-4B language markovian_zero_shot 0.02
Qwen3.5-0.8B language markovian_zero_shot 0.02

Scores at each effort level

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
Qwen3.5-4B markovian_reasoner
language ascii
0.23
Qwen3.5-2B markovian_reasoner
language
0.13
Qwen3.5-0.8B markovian_reasoner
language ascii
0.09
Qwen3.5-4B markovian_zero_shot
language rgb_array
0.09
Qwen3-4B markovian_reasoner
language ascii
0.09
Gemini 2.5 Flash-Lite markovian_zero_shot
ascii
0.06
Qwen3.5-2B markovian_zero_shot
language ascii
0.06
Qwen3-4B markovian_zero_shot
language
0.02
Qwen3.5-0.8B markovian_zero_shot
language
0.02

Source: Official Agentick leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra ascii, high, 1 seed markovian_reasoner 0.63
Claude Opus 5 ascii, high, 1 seed markovian_reasoner 0.63
GPT-5.6 Sol ascii, high, 1 seed markovian_reasoner 0.58
GPT-5.6 Terra ascii, high, 1 seed markovian_reasoner 0.57
GPT-5.6 Luna ascii, high, 1 seed markovian_reasoner 0.55
GPT-5 mini ascii MarkovianReasoner 0.31
Qwen3.5-4B ascii markovian_reasoner 0.23
Gemini 2.5 Flash-Lite Sep 2025 ascii markovian_reasoner 0.19
Qwen3.5-4B language markovian_reasoner 0.18
Qwen3.5-2B ascii markovian_reasoner 0.13
Qwen3.5-2B language markovian_reasoner 0.12
Qwen3.5-0.8B ascii markovian_reasoner 0.09
Qwen3.5-4B rgb_array markovian_zero_shot 0.09
Qwen3-4B ascii markovian_reasoner 0.09
Gemini 2.5 Flash-Lite Sep 2025 language markovian_zero_shot 0.06
Qwen3.5-2B ascii markovian_zero_shot 0.06
Qwen3.5-0.8B language markovian_reasoner 0.06
Gemini 2.5 Flash-Lite Sep 2025 ascii markovian_zero_shot 0.05
Qwen3-4B language markovian_reasoner 0.05
Qwen3.5-2B language markovian_zero_shot 0.03
Qwen3.5-4B ascii markovian_zero_shot 0.02
Qwen3-4B ascii markovian_zero_shot 0.02
Qwen3.5-0.8B ascii markovian_zero_shot 0.02
Qwen3.5-4B language markovian_zero_shot 0.02
Qwen3-4B language markovian_zero_shot 0.02
Qwen3.5-0.8B language markovian_zero_shot 0.02

4 models have held number one on the official leaderboard since its release in May 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 Sep 2026

Source: Official Agentick 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 0.20 0.40 0.60 0.80 2025 2026 0.00 0.20 0.40 0.60 0.80 2025 2026

Source: Official Agentick 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 (16 entries).
  • Pass Models scored 11 models we track, plus 5 other entries on its leaderboard.
  • Pass Newest score The newest is from September 2026, for GPT-6 Astra.
  • Note Frontier models Only 1 of the top 10 has a score. The strongest tested is GPT-6 Astra (#4).
  • Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
  • Pass Importer We read the board on roger-creus.github.io: 26 rows for models we track and 5 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