Skip to main content

Vals ProgramBench (ProgramBench)

Can language models rebuild programs from scratch?

Data verified

How BenchLM shows Vals ProgramBench mirror

BenchLM mirrors the public Vals AI Vals ProgramBench mirror leaderboard captured from https://www.vals.ai/benchmarks/programbench and updated by Vals on July 8, 2026. The snapshot preserves overall scores, uncertainty, latency, cost-per-test metadata, and task-level scores where Vals publishes them.

Vals ProgramBench mirror is display only on BenchLM. Vals proprietary or Vals-hosted aggregate views are useful context, but BenchLM does not use them as weighted ranking inputs or as a replacement for benchmark-native source records.

27 Vals rows4 task viewspublic datasetTasks: Overall, Fully Resolved, Almost Resolved, Raw Pass RateDisplay only

Vals ProgramBench mirror score on ProgramBench — July 8, 2026

BenchLM mirrors the published vals programbench mirror score view for ProgramBench. Claude Fable 5 leads the public snapshot at 2.00% , followed by GPT-5.6 Sol (1.50%) and Claude Opus 4.8 (1.00%). BenchLM does not use these results to rank models overall.

27 modelsExternal benchmark mirrorsCurrentDisplay onlyUpdated July 8, 2026

The published ProgramBench snapshot is tightly clustered at the top: Claude Fable 5 sits at 2.00%, while the third row is only 1.00 points behind. The broader top-10 spread is 2.00 points, so many of the published scores sit in a relatively narrow band.

27 models have been evaluated on ProgramBench. The benchmark falls in the External benchmark mirrors category. BenchLM tracks this category separately from its weighted global scoring system, so these results are best compared on the dedicated Korean benchmark views. ProgramBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About ProgramBench

Year

2026

Tasks

Program reconstruction tasks

Format

Accuracy score

Difficulty

Cleanroom software engineering

BenchLM mirrors the public Vals AI ProgramBench leaderboard as display-only external evidence. The captured snapshot preserves overall scores, task-level scores where Vals publishes them, uncertainty, latency, and cost-per-test metadata. It is excluded from BenchLM weighted rankings.

BenchLM freshness & provenance

Version

ProgramBench 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Vals ProgramBench mirror score table (27 models)

1
Claude Fable 5anthropic/claude-fable-5
2.00%
2
GPT-5.6 Solopenai/gpt-5.6-sol
1.50%
3
Claude Opus 4.8anthropic/claude-opus-4-8
1.00%
4
GPT-5.5openai/gpt-5.5
0.50%
5
GLM 5.2zai/glm-5.2
0.50%
6
GPT-5.6 Terraopenai/gpt-5.6-terra
0.50%
7
GPT-5.4 2026 03 05 Highopenai/gpt-5.4-2026-03-05-high
0.50%
8
Claude Sonnet 4.6anthropic/claude-sonnet-4-6
0.50%
9
Claude Sonnet 5anthropic/claude-sonnet-5
0.00%
10
Claude Opus 4.7anthropic/claude-opus-4-7
0.00%
11
GPT-5.4openai/gpt-5.4-2026-03-05
0.00%
12
GPT-5.4 Miniopenai/gpt-5.4-mini-2026-03-17
0.00%
13
Gemini 3.5 Flashgoogle/gemini-3.5-flash
0.00%
14
GLM 5.1zai/glm-5.1
0.00%
15
Gemini 3.1 Pro Previewgoogle/gemini-3.1-pro-preview
0.00%
16
Qwen3.6 Plusalibaba/qwen3.6-plus
0.00%
17
Kimi K2.7 Codekimi/kimi-k2.7-code
0.00%
18
Kimi K2.6kimi/kimi-k2.6
0.00%
19
Nemotron 3 Ultra 550b A55bnvidia/nemotron-3-ultra-550b-a55b
0.00%
20
DeepSeek V4 Prodeepseek/deepseek-v4-pro
0.00%
21
Gemini 3 Flash Previewgoogle/gemini-3-flash-preview
0.00%
22
Claude Haiku 4.5 20251001 Thinkinganthropic/claude-haiku-4-5-20251001-thinking
0.00%
23
Grok 4.3grok/grok-4.3
0.00%
24
Laguna Xs.2poolside/laguna-xs.2
0.00%
25
Laguna M.1poolside/laguna-m.1
0.00%
26
Gemini 3.1 Flash Lite Previewgoogle/gemini-3.1-flash-lite-preview
0.00%
27
MiniMax M2.7minimax/MiniMax-M2.7
0.00%

FAQ

What does ProgramBench measure?

Can language models rebuild programs from scratch?

Which model leads the published ProgramBench snapshot?

Claude Fable 5 currently leads the published ProgramBench snapshot with 2.00% vals programbench mirror score. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on ProgramBench?

27 AI models are included in BenchLM's mirrored ProgramBench snapshot, based on the public leaderboard captured on July 8, 2026.

Last updated: July 8, 2026 · mirrored from the public benchmark leaderboard

The AI models change fast. We track them for you.

For engineers, researchers, and the plain curious — a weekly brief on new models, ranking shifts, and pricing changes.

Free. No spam. Unsubscribe anytime.