Skip to main content
BenchLM

Vals ProgramBench (ProgramBench)

We show this table for reference; we do not rank on it.

Data verified 36 confirmed releases in the last 30 daysFollow model changes

Can language models rebuild programs from scratch?

Vals ProgramBench mirror score on ProgramBench — September 27, 2026

We mirror the published vals programbench mirror score view for ProgramBench. Claude Opus 5.5 leads the public snapshot at 18.50%, followed by Claude Fable 5.1 (7.00%) and GPT-6 Astra (5.50%). We do not use these results to rank models overall.

53 modelsCodingCurrentDisplay onlyUpdated September 27, 2026

Vals ProgramBench mirror score table (53 models)

Score
1
Claude Opus 5.5Anthropic · Closed
18.50%
2
Claude Fable 5.1Anthropic · Closed
7.00%
3
GPT-6 AstraOpenAI · Closedmax reasoning
5.50%
4
Claude Opus 5Anthropic · Closed
3.00%
5
Claude Fable 5Anthropic · Closed
2.00%
6
GPT-6 SolOpenAI · Closedmax reasoning
2.00%
7
Kimi K3Moonshot AI · Closed
2.00%
8
GPT-5.6 SolOpenAI · Closedmax reasoning
1.50%
9
GLM-5.3Z.AI · Open weightmax reasoning
1.50%
10
Claude Opus 4.8Anthropic · Closed
1.00%
11
Gemini 3.8 FlashGoogle · Closedhigh reasoning
1.00%
12
Grok 4.7xAI · Closedxhigh reasoning
0.50%
13
MiMo-V2.6-FlashXiaomi · Open weight
0.50%
14
GPT-5.6 TerraOpenAI · Closedmax reasoning
0.50%
15
MiMo-V2.6-ProXiaomi · Open weight
0.50%
16
GPT-5.5OpenAI · Closedxhigh reasoning
0.50%
17
GLM-5.2Z.AI · Open weightmax reasoning
0.50%
18
GPT-6 LunaOpenAI · Closedmax reasoning
0.50%
19
GPT-5.4 2026 03 05 HighOpenAIhigh reasoning
0.50%
20
Claude Sonnet 4.6Anthropic · Closed
0.50%
21
Inkling-SmallThinking Machines Lab · Open weight0.99 reasoning
0.50%
22
Hy4 previewTencent · Open weight
0.00%
23
DeepSeek V4 Pro 0813DeepSeek · Open weightmax reasoning
0.00%
24
Claude Sonnet 5Anthropic · Closed
0.00%
25
Qwen3.8 MaxAlibaba · Open weight
0.00%
26
Gemini 3.7 FlashGoogle · Closedhigh reasoning
0.00%
27
Claude Opus 4.7Anthropic · Closed
0.00%
28
GPT-5.6 LunaOpenAI · Closedmax reasoning
0.00%
29
DeepSeek V4 Flash 0731DeepSeek · Open weighthigh reasoning
0.00%
30
GPT-5.4OpenAI · Closedxhigh reasoning
0.00%
31
GLM-5.3-FlashZ.AI · Open weightmax reasoning
0.00%
32
Grok 4.5xAI · Closedhigh reasoning
0.00%
33
GPT-5.4 miniOpenAI · Closedxhigh reasoning
0.00%
34
Gemini 3.6 FlashGoogle · Closedhigh reasoning
0.00%
35
Muse Spark 1.1Meta · Closedxhigh reasoning
0.00%
36
Gemini 3.5 FlashGoogle · Closedhigh reasoning
0.00%
37
GLM-5.1Z.AI · Open weight
0.00%
38
Gemini 3.1 Pro PreviewGooglehigh reasoning
0.00%
39
Qwen3.6 PlusAlibaba · Closed
0.00%
40
Qwen3.8-27BAlibaba · Open weightxhigh reasoning
0.00%
41
Kimi K2.7 CodeMoonshot AI · Open weight
0.00%
42
Kimi K2.6Moonshot AI · Open weight
0.00%
44
DeepSeek V4 Pro 0813DeepSeek · Open weightmax reasoning
0.00%
45
Gemini 3 Flash PreviewGooglehigh reasoning
0.00%
46
Gemini 3.5 Flash-LiteGoogle · Closedhigh reasoning
0.00%
47
Claude Haiku 4.5 ThinkingAnthropic · Closed
0.00%
48
Grok 4.3xAI · Closedhigh reasoning
0.00%
49
Laguna XS.2Poolside · Open weight
0.00%
50
InklingThinking Machines Lab · Open weight0.99 reasoning
0.00%
51
Laguna M.1Poolside · Closed
0.00%
52
Gemini 3.1 Flash Lite PreviewGooglehigh reasoning
0.00%
53
MiniMax M2.7MiniMax · Open weight
0.00%

How Vals ProgramBench mirror is shown here

BenchLM mirrors the public Vals AI Vals ProgramBench mirror leaderboard captured from https://www.vals.ai/benchmarks/programbench and updated by Vals on September 27, 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.

Snapshot

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

The published ProgramBench snapshot places Claude Opus 5.5 first at 18.50%. The third row is 13.00 points behind. The broader top-10 range is 17.50 points, so the table still separates the published systems.

53 models have been evaluated on ProgramBench. The benchmark falls in the Coding category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. 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.

Freshness and 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.

Questions

What does ProgramBench measure?

Can language models rebuild programs from scratch?

Which model leads the published ProgramBench snapshot?

Claude Opus 5.5 currently leads the published ProgramBench snapshot with 18.50% 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?

The September 27, 2026 snapshot contains 53 AI models.

Last updated: September 27, 2026 · mirrored from the public benchmark leaderboard

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 2,000+ readers.

One email each week. Unsubscribe anytime.