Taste-Bench
Models choose the better next step at 502 decision points in software engineering and research trajectories. A question scores only when both option orders receive the correct answer. The overall score equally weights engineering and research.
Dataset access requires approval to limit training contamination. Later trajectory outcomes supply labels. Request errors and unparseable responses count as wrong. The paired-order protocol gives random guessing an expected score of 25 percent.
- Made by
- City University of Hong Kong and Microsoft
- Released
- September 2026
- Kind
- Machine learning engineering
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- Kept private
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
- Claude Opus 5 #11 on the capability index
- Last model added
- July 2026
- Scores refreshed
- November 18th, 2026
In the authors’ words
LLM agents increasingly work on long-horizon tasks, and the decisions they make along the way, such as which hypothesis to test or which implementation to build on, determine the outcome of the whole run. Making these decisions well is becoming a key capability for both engineering and research agents. We refer to the ability to make good long-horizon decisions as the taste of an agent. While existing benchmarks measure the end-to-end success of agents on long-horizon tasks, none of them measures the taste of an agent. To address this problem, we build Taste-Bench, a benchmark of taste questions constructed automatically from trajectories that agents produced in engineering and research tasks. Each question presents a decision fork, a point in a trajectory where multiple directions are available and one of them leads to a better outcome, and the evaluated model chooses among these directions without seeing what happens after the fork. We mine these forks automatically from parallel attempts at the same task and from detours inside a single trajectory, without needing human annotation. We evaluate frontier models on Taste-Bench and find that the best model answers only 59.7% of the questions correctly. We further find that forks whose deciding evidence appears later in the trajectory are much harder for every model, and that a larger reasoning budget does not improve the accuracy. Finally, we show that taste can be trained. We distill the judgment of a teacher that has seen the outcome into a student model, and the student makes better decisions on unseen tasks and improves end-to-end success on held-out SWE-bench Pro tasks.
GPT-5.6 Sol leads with 59.7% among models released by July 2026
Best score per model from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best.
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:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 13 more
- 1 GPT-5.6 Sol 59.7%
- 2 GPT-5.5 59.5%
- 3 Claude Opus 5 55.5%
- 4 Grok 4.5 54.6%
- 5 GPT-5.6 Terra 54%
- 6 GLM-5.2 53.9%
- 7 Claude Sonnet 5 51.6%
- 8 GPT-5.6 Luna 49%
- 9
MiniMax-M3 45.3%
- 10
DeepSeek-V4-Flash 43.3%
- 11 GPT-5.4 Mini 40.1%
- 12
Mistral Medium 3.5 37.7%
- 13 GPT-5.4 Nano 36.6%
- 14 Grok 4.20 reasoning 15.7%
Source: Official Taste-Bench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.6 Sol | 59.7% | Jul 2026 | Official leaderboard | |
| 2 | GPT-5.5 | 59.5% | Apr 2026 | Official leaderboard | |
| 3 | Claude Opus 5 | 55.5% | Jul 2026 | Official leaderboard | |
| 4 | Grok 4.5 | 54.6% | Jul 2026 | Official leaderboard | |
| 5 | GPT-5.6 Terra | 54% | Jul 2026 | Official leaderboard | |
| 6 | GLM-5.2 | 53.9% | Jun 2026 | Official leaderboard | |
| 7 | Claude Sonnet 5 | 51.6% | Jun 2026 | Official leaderboard | |
| 8 | GPT-5.6 Luna | 49% | Jul 2026 | Official leaderboard | |
| 9 | | 45.3% | Jun 2026 | Official leaderboard | |
| 10 | | 43.3% | Apr 2026 | Official leaderboard | |
| 11 | GPT-5.4 Mini | 40.1% | Mar 2026 | Official leaderboard | |
| 12 | | 37.7% | Apr 2026 | Official leaderboard | |
| 13 | GPT-5.4 Nano | 36.6% | Mar 2026 | Official leaderboard | |
| 14 | Grok 4.20 | 15.7% | reasoning | Feb 2026 | Official leaderboard |
Rankings by part
We read 4 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Detour engineering, Detour research and Parallel engineering. GPT-5.6 Sol leads 2 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.
Detour engineering
14 models we track have a score here.
- 1 GPT-5.6 Sol 48.1%
- 2 Grok 4.5 ↑2 47.7%
- 3 GPT-5.5 ↓1 47.4%
- 4
Mistral Medium 3.5 ↑8 40.6%
- 5 GPT-5.6 Terra 40.6%
- 6 GLM-5.2 40.2%
- 7 Claude Sonnet 5 36.1%
- 8 Claude Opus 5 ↓5 35.3%
- 9
MiniMax-M3 34.6%
- 10 GPT-5.6 Luna ↓2 32.3%
Detour research
14 models we track have a score here.
- 1 GPT-5.5 ↑1 68.8%
- 2 GPT-5.6 Sol ↓1 67.2%
- 3 Claude Opus 5 64.1%
- 4 Claude Sonnet 5 ↑3 60.9%
- 5 GLM-5.2 ↑1 59.4%
- 6 Grok 4.5 ↓2 57.8%
- 7 GPT-5.6 Luna ↑1 57.8%
- 8 GPT-5.6 Terra ↓3 57.8%
- 9 GPT-5.4 Mini ↑2 54.7%
- 10
DeepSeek-V4-Flash 46.9%
Parallel engineering
14 models we track have a score here.
- 1 GPT-5.6 Sol 75.8%
- 2 GPT-5.5 73.4%
- 3 Claude Opus 5 71%
- 4 GPT-5.6 Terra ↑1 70.2%
- 5 GPT-5.6 Luna ↑3 67.7%
- 6
MiniMax-M3 ↑3 64.5%
- 7 GLM-5.2 ↓1 64.5%
- 8 Claude Sonnet 5 ↓1 62.1%
- 9 Grok 4.5 ↓5 61.3%
- 10
DeepSeek-V4-Flash 58.1%
Parallel research
14 models we track have a score here.
- 1 Claude Opus 5 ↑2 64.6%
- 2 GLM-5.2 ↑4 60.4%
- 3 GPT-5.6 Terra ↑2 58.3%
- 4 GPT-5.5 ↓2 56.2%
- 5
MiniMax-M3 ↑4 56.2%
- 6 Claude Sonnet 5 ↑1 56.2%
- 7 Grok 4.5 ↓3 56.2%
- 8 GPT-5.6 Sol ↓7 56.2%
- 9 GPT-5.4 Mini ↑2 54.2%
- 10
DeepSeek-V4-Flash 50%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official Taste-Bench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Taste-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 AI research
- PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics.
Qwen 3.8 Max 93%
- MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
- PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
- MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6.2 Astra 96.1%
- PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Kimi K3 0.71
- KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup.
MiniMax-M3 28.8%
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access The tasks are private.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the official leaderboard (14 entries).
- Pass Models scored 14 models we track.
- Pass Newest score The newest is from July 2026, for Claude Opus 5.
- Note Frontier models None of the top 10 has a score. The strongest tested is Claude Opus 5 (#11).
- Pass Headroom The best is 59.7%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 14 rows for models we track at the last read.
- Pass Last read November 18th, 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.