SolidWorks Bench
Agents model the custom parts of 73 real machines in SolidWorks and assemble them from raster drawing packages, through screenshots and the SolidWorks API. A grader checks the geometry against the reference and reads mates, motion and feature trees from the native files.
Launched November 17, 2026 with nine models at high reasoning effort, one attempt per machine. Machines have 4 to 185 custom parts and time limits of 2.7 to 24 hours. The reward is a weighted sum of 11 graded components, set to zero when the STEP export is missing or does not match the native files.
- Made by
- Markov
- Released
- November 2026
- Kind
- CAD models
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Kept private
- Measures
- Capability
- Subject
- Mechanical engineering
- Language
- English
- Interaction
- Agentic
- Environment
- SolidWorks, APIs and tools
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- A host
- In views
- CAD, Agents, Vision, Image, video and 3D generation
- Strongest model tested
- Claude Fable 5.2 #2 on the capability index
- Last model added
- November 2026
- Scores refreshed
- November 19th, 2026
Claude Fable 5.2 leads with 0.57
Best score per model from the benchmark's own leaderboard, read on November 19th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 1 Claude Fable 5.2 high 0.57
- 2 GPT-6 Astra high 0.43
- 3 Claude Sonnet 5.5 high 0.35
- 4 Claude Opus 5.5 high 0.34
- 5 Kimi K3.1 Pro high 0.20
- 6
DeepSeek V4.1 Flash high 0.13
- 7
Muse Spark Open high 0.07
- 8 GLM-5.3-Flash high 0.06
- 9
MiniMax-M3.1-Flash high 0.02
Source: Official SolidWorks Bench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 | 0.57 | high | Nov 2026 | Official leaderboard |
| 2 | GPT-6 Astra | 0.43 | high | Sep 2026 | Official leaderboard |
| 3 | Claude Sonnet 5.5 | 0.35 | high | Sep 2026 | Official leaderboard |
| 4 | Claude Opus 5.5 | 0.34 | high | Sep 2026 | Official leaderboard |
| 5 | Kimi K3.1 Pro | 0.20 | high | Oct 2026 | Official leaderboard |
| 6 | | 0.13 | high | Sep 2026 | Official leaderboard |
| 7 | | 0.07 | high | Oct 2026 | Official leaderboard |
| 8 | GLM-5.3-Flash | 0.06 | high | Aug 2026 | Official leaderboard |
| 9 | | 0.02 | high | Oct 2026 | Official leaderboard |
Rankings by part
We read 14 parts from the benchmark’s own leaderboard in 4 groups: area, geometry, assembly function and part practice. Each ranks models on one part of the test. Claude Fable 5.2 leads 11 of them, and GPT-6 Astra leads the other 3. Arrows show how far a model moved from its overall rank among the models we track.
Geometry
Parts and placement, read from the STEP export. 45% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.64
- 2 Claude Opus 5.5 ↑2 0.49
- 3 GPT-6 Astra ↓1 0.40
- 4 Claude Sonnet 5.5 ↓1 0.39
- 5 Kimi K3.1 Pro 0.14
- 6 GLM-5.3-Flash ↑2 0.08
- 7
DeepSeek V4.1 Flash ↓1 0.06
- 8
MiniMax-M3.1-Flash ↑1 0.06
- 9
Muse Spark Open ↓2 0.04
Assembly function
Ties, motion, lock-free mates and patterns, read from the native files. 30% of the reward. 9 models we track have a score here.
- 1 GPT-6 Astra ↑1 0.54
- 2 Claude Fable 5.2 ↓1 0.51
- 3 Kimi K3.1 Pro ↑2 0.37
- 4
DeepSeek V4.1 Flash ↑2 0.32
- 5 Claude Sonnet 5.5 ↓2 0.32
- 6
MiniMax-M3.1-Flash ↑3 0.30
- 7 GLM-5.3-Flash ↑1 0.27
- 8
Muse Spark Open ↓1 0.24
- 9 Claude Opus 5.5 ↓5 0.09
Part practice
Flex, materials, sketches, holes and part patterns. 25% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.56
- 2 Claude Opus 5.5 ↑2 0.37
- 3 GPT-6 Astra ↓1 0.34
- 4 Claude Sonnet 5.5 ↓1 0.30
- 5 Kimi K3.1 Pro 0.12
- 6 GLM-5.3-Flash ↑2 0.06
- 7
DeepSeek V4.1 Flash ↓1 0.06
- 8
MiniMax-M3.1-Flash ↑1 0.05
- 9
Muse Spark Open ↓2 0.02
Parts
Every face of each modeled part checked against the reference at 0.025 mm and 0.1°, plus the exact volume difference. 25% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.73
- 2 Claude Opus 5.5 ↑2 0.65
- 3 Claude Sonnet 5.5 0.57
- 4 GPT-6 Astra ↓2 0.53
- 5 Kimi K3.1 Pro 0.22
- 6 GLM-5.3-Flash ↑2 0.12
- 7
DeepSeek V4.1 Flash ↓1 0.09
- 8
MiniMax-M3.1-Flash ↑1 0.08
- 9
Muse Spark Open ↓2 0.06
Placement
Every pair of parts that touch in the reference, graded on how the two sit against each other. 20% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.53
- 2 Claude Opus 5.5 ↑2 0.28
- 3 GPT-6 Astra ↓1 0.24
- 4 Claude Sonnet 5.5 ↓1 0.17
- 5 Kimi K3.1 Pro 0.04
- 6 GLM-5.3-Flash ↑2 0.03
- 7
DeepSeek V4.1 Flash ↓1 0.03
- 8
MiniMax-M3.1-Flash ↑1 0.03
- 9
Muse Spark Open ↓2 0.02
Ties
Each component mated to a neighbor it really touches, weighted by how well it sits. 10% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.34
- 2 GPT-6 Astra 0.29
- 3
MiniMax-M3.1-Flash ↑6 0.17
- 4 GLM-5.3-Flash ↑4 0.11
- 5 Kimi K3.1 Pro 0.10
- 6
DeepSeek V4.1 Flash 0.09
- 7 Claude Sonnet 5.5 ↓4 0.09
- 8 Claude Opus 5.5 ↓4 0.07
- 9
Muse Spark Open ↓2 0.07
Motion
SolidWorks drags each joint: hinges turn about the right axis, slides slide and rigid parts hold. 10% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.46
- 2 GPT-6 Astra 0.44
- 3 Kimi K3.1 Pro ↑2 0.39
- 4 Claude Sonnet 5.5 ↓1 0.34
- 5
DeepSeek V4.1 Flash ↑1 0.32
- 6 GLM-5.3-Flash ↑2 0.24
- 7
Muse Spark Open 0.14
- 8
MiniMax-M3.1-Flash ↑1 0.14
- 9 Claude Opus 5.5 ↓5 0.07
Lock-free
No Lock-mate shortcuts. 5% of the reward. 9 models we track have a score here.
- 1 GPT-6 Astra ↑1 1.00
- 2
MiniMax-M3.1-Flash ↑7 1.00
- 3 Kimi K3.1 Pro ↑2 1.00
- 4 Claude Fable 5.2 ↓3 1.00
- 5
DeepSeek V4.1 Flash ↑1 0.98
- 6
Muse Spark Open ↑1 0.91
- 7 Claude Sonnet 5.5 ↓4 0.84
- 8 GLM-5.3-Flash 0.80
- 9 Claude Opus 5.5 ↓5 0.21
Patterns
Repeated components placed by patterns. 5% of the reward. 9 models we track have a score here.
- 1 GPT-6 Astra ↑1 0.93
- 2 Claude Fable 5.2 ↓1 0.45
- 3 Claude Sonnet 5.5 0.14
- 4 Kimi K3.1 Pro ↑1 0.10
- 5 GLM-5.3-Flash ↑3 0.00
- 6
DeepSeek V4.1 Flash 0.00
- 7 Claude Opus 5.5 ↓3 0.00
- 8
Muse Spark Open ↓1 0.00
- 9
MiniMax-M3.1-Flash 0.00
Sketches
Sketches fully defined without Fix relations. 4% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.71
- 2 Claude Opus 5.5 ↑2 0.49
- 3 GPT-6 Astra ↓1 0.47
- 4 Claude Sonnet 5.5 ↓1 0.38
- 5 Kimi K3.1 Pro 0.13
- 6 GLM-5.3-Flash ↑2 0.03
- 7
Muse Spark Open 0.03
- 8
DeepSeek V4.1 Flash ↓2 0.02
- 9
MiniMax-M3.1-Flash 0.02
Holes
Hole features where the drawing calls for them. 4% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.58
- 2 GPT-6 Astra 0.43
- 3 Claude Opus 5.5 ↑1 0.19
- 4 Claude Sonnet 5.5 ↓1 0.17
- 5 Kimi K3.1 Pro 0.11
- 6
DeepSeek V4.1 Flash 0.01
- 7
MiniMax-M3.1-Flash ↑2 0.01
- 8 GLM-5.3-Flash 0.00
- 9
Muse Spark Open ↓2 0.00
Part patterns
Repeated features patterned, not copied. 4% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.39
- 2 GPT-6 Astra 0.20
- 3 Claude Opus 5.5 ↑1 0.08
- 4 Claude Sonnet 5.5 ↓1 0.06
- 5 Kimi K3.1 Pro 0.02
- 6 GLM-5.3-Flash ↑2 0.00
- 7
DeepSeek V4.1 Flash ↓1 0.00
- 8
Muse Spark Open ↓1 0.00
- 9
MiniMax-M3.1-Flash 0.00
Materials
The material from the bill of materials on each part. 7% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.61
- 2 Claude Opus 5.5 ↑2 0.57
- 3 GPT-6 Astra ↓1 0.38
- 4 Claude Sonnet 5.5 ↓1 0.34
- 5 Kimi K3.1 Pro 0.16
- 6
DeepSeek V4.1 Flash 0.06
- 7
MiniMax-M3.1-Flash ↑2 0.05
- 8 GLM-5.3-Flash 0.04
- 9
Muse Spark Open ↓2 0.02
Flex
A key dimension changed by 5 to 30%, rebuilt, and compared with the reference changed the same way. 6% of the reward. 9 models we track have a score here.
- 1 Claude Fable 5.2 0.43
- 2 Claude Sonnet 5.5 ↑1 0.38
- 3 Claude Opus 5.5 ↑1 0.26
- 4 GPT-6 Astra ↓2 0.24
- 5 GLM-5.3-Flash ↑3 0.18
- 6
DeepSeek V4.1 Flash 0.17
- 7
MiniMax-M3.1-Flash ↑2 0.13
- 8 Kimi K3.1 Pro ↓3 0.10
- 9
Muse Spark Open ↓2 0.02
The best score, 0.57, costs $69 a task
Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official SolidWorks Bench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Claude Fable 5.2 | high | 0.57 | $69 | ||
| GPT-6 Astra | high | 0.43 | $47 | ||
| Claude Sonnet 5.5 | high | 0.35 | $17 | ||
| Claude Opus 5.5 | high | 0.34 | $25 | ||
| Kimi K3.1 Pro | high | 0.20 | $11 | ||
| high | 0.13 | $1.12 | |||
| high | 0.07 | $13 | |||
| GLM-5.3-Flash | high | 0.06 | $3.04 | ||
| high | 0.02 | $15 |
2 models have held number one on the official leaderboard since its release in November 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official SolidWorks 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 SolidWorks 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 Engineering and CAD
- EEBench In use 13 electrical-engineering circuit and PCB design tasks graded by deterministic build checks, simulation and BOM constraints. No LLM judge.
Gemini 4 Carbon 83.9%
- 3DCodeBench In use Agentic procedural 3D modeling via code: author engine-ready 3D assets through software APIs. Graded for executability and shape fidelity.
LongCat-2.5 55.4%
- CADGenBench Only in system cards Agents construct CAD models from design prompts. The generation split is graded for executability and geometric correctness. Claude Fable 5 45.1%
- PartBench Only in system cards Parametric mechanical part design and CAD-style generation, graded for geometric and manufacturing fidelity. Grok 4.6 58.5%
- Razavi-Bench Not in model cards Models explain what happens in analog circuits and why. A language model awards up to four points per answer against a reference and a rubric.
Gemini 4 Carbon 98.3%
- CircuitRubric Not in model cards Models write analog circuit netlists from design requests. A program compares transistor connections and relative device sizes with accepted reference circuits. Claude Opus 4.8 96%
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 (9 entries).
- Pass Models scored 9 models we track.
- Pass Newest score The newest is from November 2026, for Claude Fable 5.2.
- Pass Frontier models 4 of the top 10 have a score, led by Claude Fable 5.2 (#2).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on markovstudios.com: 9 rows for models we track at the last read.
- Pass Last read November 19th, 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.