Coding · Benchmark profile

ProgramBench: Can Language Models Rebuild Programs From Scratch?

Reference2026Active4 models

Data verified 1 Aug 2026 · Methodology 2.3.0

Benchmark score on ProgramBench: Can Language Models Rebuild Programs From Scratch?

Leaderboard · 4 models

Claude Opus 5 leads at 93%, followed by Kimi K3 (77.8%) and GLM-5.2 (63.7%).

Display onlyCoding2026

#1Closed

Claude Opus 5

Anthropic

claude-opus-5

93%

Overall #1 · Context 1M tokens

#2Open

Kimi K3

Moonshot AI

kimi-k3

77.8%

Overall #7 · Context 1M tokens

#3Open

GLM-5.2

Z.AI

glm-5-2

63.7%

Overall #20 · Context 1M tokens

ViewModelsColours

AnthropicMoonshot AIZ.AI

Lumina

ProgramBench: Can Language Models Rebuild Programs From Scratch? leaderboard

Accessible data for ProgramBench: Can Language Models Rebuild Programs From Scratch? leaderboard
Model	Rank	Model ID	Provider	Score	Unit
Claude Opus 5	#1	claude-opus-5	Anthropic	93	percent
Kimi K3	#2	kimi-k3	Moonshot AI	77.8	percent
GLM-5.2	#3	glm-5-2	Z.AI	63.7	percent
Kimi K2.7 Code	#4	kimi-k2-7-code	Moonshot AI	53.6	percent

Benchmark score table (4 models)

One best compatible score per canonical product · higher is better

ProgramBench: Can Language Models Rebuild Programs From Scratch? leaderboard ranked by score
Rank	Model	Provider	License	Evidence use	Score
#1	Claude Opus 5 claude-opus-5	Anthropic	closed	Source carrier	93%
#2	Kimi K3 kimi-k3	Moonshot AI	open	Source carrier	77.8%
#3	GLM-5.2 glm-5-2	Z.AI	open	Source carrier	63.7%
#4	Kimi K2.7 Code kimi-k2.7-code	Moonshot AI	open	Source carrier	53.6%

The top of this snapshot is led by Claude Opus 5 at 93%; third place is 29.3 points behind. The top-4 spread is 39.4 points.

Configuration-preserving evidencePublished source rows (11)Exact source labels, configurations, versions and dated observations. Same-snapshot canonical/configuration projections are shown once.

Every distinct dated source-published row for ProgramBench: Can Language Models Rebuild Programs From Scratch?, including exact model configuration labels
Exact published model / configuration	Canonical product	Score	Version / system	Evidence	Source / observed
Exact BenchLM registry variant Claude Opus 5; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	Claude Opus 5	93%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01
Kimi K3 public default configuration (reasoning_effort=max; temperature=1.0) kimi-k3-max	Kimi K3	77.8%	ProgramBench public evaluationKimi Code	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-29
Exact BenchLM registry variant GLM-5.2; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	GLM-5.2	63.7%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01
Exact BenchLM registry variant Kimi K2.7 Code; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	Kimi K2.7 Code	53.6%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01
Exact BenchLM registry variant Claude Opus 5; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	Claude Opus 5	93%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-07-27 ↗Observed 2026-07-27 · checked 2026-07-27
Exact BenchLM registry variant Kimi K3; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	Kimi K3	77.8%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-07-27 ↗Observed 2026-07-27 · checked 2026-07-27
Exact BenchLM registry variant GLM-5.2; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	GLM-5.2	63.7%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-07-27 ↗Observed 2026-07-27 · checked 2026-07-27
Exact BenchLM registry variant Kimi K2.7 Code; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	Kimi K2.7 Code	53.6%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 2026-07-27 ↗Observed 2026-07-27 · checked 2026-07-27
Exact BenchLM registry variant Kimi K3; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	Kimi K3	77.8%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 21 July 2026 ↗Observed 2026-07-21 · checked 2026-07-21
Exact BenchLM registry variant GLM-5.2; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	GLM-5.2	63.7%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 21 July 2026 ↗Observed 2026-07-21 · checked 2026-07-21
Exact BenchLM registry variant Kimi K2.7 Code; bulk export does not retain a complete upstream harness configuration. source label without registered configuration ID	Kimi K2.7 Code	53.6%	2026Source-native system	Source carrier · no additional scoring weight source-checked	BenchLM public datasets — 21 July 2026 ↗Observed 2026-07-21 · checked 2026-07-21

Compare top models

Claude Opus 5 vs Kimi K3 Kimi K3 vs GLM-5.2 GLM-5.2 vs Kimi K2.7 Code

About ProgramBench: Can Language Models Rebuild Programs From Scratch?

Definition and scoring

Organisation: John Yang, Kilian Lieret, Jeffrey Ma, Parth Thakkar, Dmitrii Pedchenko, Sten Sootla, Emily McMilin, Pengcheng Yin, Rui Hou, Gabriel Synnaeve, Diyi Yang, Ofir Press
Category: Coding
Version: 2026
Direction: higher is better
Ranking use: Reference
Contamination risk: Unknown

A cleanroom software-engineering benchmark where agents receive only a compiled executable and documentation, then must architect and implement a complete codebase that reproduces the original program's behavior. Every genuine source row stays tied to its exact model label, configuration, benchmark version and evaluation system. The summary chart shows one best compatible score per canonical product; Score 2.0 admits only explicitly mapped, frozen protocols and keeps incompatible configurations separate.

Open benchmark source ↗

FAQ

Quick answers

What does ProgramBench: Can Language Models Rebuild Programs From Scratch? measure?

Which model scores highest on ProgramBench: Can Language Models Rebuild Programs From Scratch??

Claude Opus 5 by Anthropic currently leads with 93%.

How many models are evaluated on ProgramBench: Can Language Models Rebuild Programs From Scratch??

4 canonical model products have a qualifying score, backed by 11 distinct dated source rows.

Does this affect overall Lumina rank?

No. This benchmark is display-only and does not enter the overall Lumina composite.

Coding · Benchmark profile

ProgramBench: Can Language Models Rebuild Programs From Scratch?

Reference2026Active4 models

Data verified 1 Aug 2026 · Methodology 2.3.0

Benchmark score on ProgramBench: Can Language Models Rebuild Programs From Scratch?