Benchmark Almanac

VibePrevisBench

GPT-6 Astra 77.7 leads since November 2026
Not in model cards

Coding agents write Blender Python that turns a script, photos, an opening frame or reference footage into a graybox 4D previs scene and render it. Rule checks score the exported scene against hidden facts, and a VLM judge checks that those facts are visible in the render.

242 tasks. The overall score is half Scene Check and half Screen Check on a 0 to 100 scale, with failed deliveries scored zero. The judge, Gemini 4 Argon, is also one of the 11 agents tested. Annotations are CC BY-NC-SA 4.0.

Released
November 2026
Kind
CAD models
Level
Expert
Scored
Graded by a model
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images, Video → 3D, Code, Video, Actions
Interaction
Agentic
Environment
Blender
Refresh
One fixed set
Run by
Anyone
Strongest model tested
GPT-6 Astra #4 on the capability index
Last model added
September 2026
Scores refreshed
November 20th, 2026
11
Models scored on its own leaderboard

GPT-6 Astra leads with 77.7 among models released by September 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 September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

Source: Official VibePrevisBench leaderboard

#ModelScoreSettingDateSource
1 GPT-6 Astra 77.7 reasoning: max; 4-hour task limit, Codex CLI Sep 2026 Official leaderboard
2 GPT-6.1 Sol 76.0 reasoning: max; 4-hour task limit, Codex CLI Sep 2026 Official leaderboard
3 Gemini 4 Argon 75.0 reasoning: high; 4-hour task limit, Claude Code Sep 2026 Official leaderboard
4 Claude Opus 5.5 74.1 reasoning: high; 4-hour task limit, Claude Code Sep 2026 Official leaderboard
5 GPT-5.6 Sol 70.2 reasoning: max; 4-hour task limit, Codex CLI Jul 2026 Official leaderboard
6 DeepSeek V4.1 Flash 69.6 reasoning: max; 4-hour task limit; model snapshot: 1017, Claude Code Sep 2026 Official leaderboard
7 Kimi K3 67.5 reasoning: xhigh; 4-hour task limit, Claude Code Jul 2026 Official leaderboard
8 Gemini 3.8 Flash 64.7 reasoning: high; 4-hour task limit, Claude Code Sep 2026 Official leaderboard
9 GLM-5.3-Flash 63.7 reasoning: high; 4-hour task limit, Claude Code Aug 2026 Official leaderboard
10 Doubao Seed 2.1 Pro 61.3 reasoning: max; 4-hour task limit, Claude Code Jun 2026 Official leaderboard
11 Muse Spark 1.3 61.1 reasoning: xhigh; 4-hour task limit, Claude Code Sep 2026 Official leaderboard

Rankings by part

We read 189 parts from the benchmark’s own leaderboard in 5 groups: modality, environment, dimension, category and length. Each ranks models on one part of the test. GPT-6 Astra leads 128 of them, and 6 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Text

11 models we track have a score here.

  1. 1 GPT-6 Astra 81.0
  2. 2 GPT-6.1 Sol 78.8
  3. 3 Claude Opus 5.5 ↑1 78.0
  4. 4 Gemini 4 Argon ↓1 77.4
  5. 5 DeepSeek V4.1 Flash ↑1 75.5
  6. 6 Kimi K3 ↑1 75.3
  7. 7 GPT-5.6 Sol ↓2 75.1
  8. 8 Gemini 3.8 Flash 68.9
  9. 9 GLM-5.3-Flash 68.5
  10. 10 Muse Spark 1.3 ↑1 67.8

The best score, 77.7, costs $33 a task

Runs on the cost frontier score higher than any run that's cheaper. Gemini 4 Argon at reasoning: high; 4-hour task limit comes within 2.7 of the best for $4.45 a task, 7 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Z.ai OpenAI DeepSeek Google DeepMind Cost frontier One model at each effort level
0.00 20.0 40.0 60.0 80.0 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra, reasoning: max; 4-hour task limit DeepSeek V4.1 Flash, reasoning: max; 4-hour task limit; model snapshot: 1017 GLM-5.3-Flash, reasoning: high; 4-hour task limit 0.00 20.0 40.0 60.0 80.0 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra, reasoning: max; 4-hour task limit DeepSeek V4.1 Flash, reasoning: max; 4-hour task limit; model snapshot: 1017

Source: Official VibePrevisBench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra reasoning: max; 4-hour task limit Codex CLI 77.7 $33 148,594
GPT-6.1 Sol reasoning: max; 4-hour task limit Codex CLI 76.0 $6.71 195,254
Gemini 4 Argon reasoning: high; 4-hour task limit Claude Code 75.0 $4.45 180,978
Claude Opus 5.5 reasoning: high; 4-hour task limit Claude Code 74.1 $7.82 137,249
GPT-5.6 Sol reasoning: max; 4-hour task limit Codex CLI 70.2 $6.19 54,309
DeepSeek V4.1 Flash reasoning: max; 4-hour task limit; model snapshot: 1017 Claude Code 69.6 $1.89 260,679
Kimi K3 reasoning: xhigh; 4-hour task limit Claude Code 67.5 $18 226,367
Gemini 3.8 Flash reasoning: high; 4-hour task limit Claude Code 64.7 $2.73 166,256
GLM-5.3-Flash reasoning: high; 4-hour task limit Claude Code 63.7 $0.73 134,957
Doubao Seed 2.1 Pro reasoning: max; 4-hour task limit Claude Code 61.3 $5.96 143,224
Muse Spark 1.3 reasoning: xhigh; 4-hour task limit Claude Code 61.1 $3.84 157,436

Number one over time

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

OpenAI
Own leaderboard 1 leader

Source: Official VibePrevisBench 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
60.0 65.0 70.0 75.0 80.0 2026 60.0 65.0 70.0 75.0 80.0 2026

Source: Official VibePrevisBench 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

All 112 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 All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (11 entries).
  • Pass Models scored 11 models we track.
  • Pass Newest score The newest is from September 2026, for Gemini 4 Argon.
  • Pass Frontier models 4 of the top 10 have a score, led by 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 vibeprevisbench.github.io: 11 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