Coding · Benchmark profile

MLE-Bench

Machine-learning engineering benchmark where agents compete on Kaggle-style tasks; Google reports Average Position Score on the Partial-30 subset.

ReferencePartial-30Active6 models

Data verified 21 Jul 2026 · Methodology 2.3.0

Benchmark score on MLE-Bench

Leaderboard · 6 models

Claude Sonnet 5 leads at 66.9%, followed by Gemini 3.6 Flash (63.9%) and Gemini 3.5 Flash (49.7%).

Display onlyCodingPartial-30

#1Closed

Claude Sonnet 5

Anthropic

claude-sonnet-5

66.9%

Overall #16 · Context 1M tokens

#2Closed

Gemini 3.6 Flash

Google

gemini-3.6-flash

63.9%

Overall #15 · Context 1.05M tokens

#3Closed

Gemini 3.5 Flash

Google

gemini-3.5-flash

49.7%

Overall #24 · Context 1.05M tokens

ViewModelsColours

AnthropicGoogleOpenAIxAI

Lumina

MLE-Bench leaderboard

Accessible data for MLE-Bench leaderboard
Model	Rank	Model ID	Provider	Score	Unit
Claude Sonnet 5	#1	claude-sonnet-5	Anthropic	66.9	percent
Gemini 3.6 Flash	#2	gemini-3-6-flash	Google	63.9	percent
Gemini 3.5 Flash	#3	gemini-3-5-flash	Google	49.7	percent
GPT-5.6 Luna	#4	gpt-5-6-luna	OpenAI	47.6	percent
Grok 4.5	#5	grok-4-5	xAI	43.2	percent
Gemini 3.1 Pro Preview	#6	gemini-3-1-pro-preview	Google	42.6	percent

Benchmark score table (6 models)

One best compatible score per canonical product · higher is better

MLE-Bench leaderboard ranked by score
Rank	Model	Provider	License	Evidence use	Score
#1	Claude Sonnet 5 claude-sonnet-5	Anthropic	closed	Public reference	66.9%
#2	Gemini 3.6 Flash gemini-3.6-flash	Google	closed	Public reference	63.9%
#3	Gemini 3.5 Flash gemini-3.5-flash	Google	closed	Public reference	49.7%
#4	GPT-5.6 Luna gpt-5.6-luna	OpenAI	closed	Public reference	47.6%
#5	Grok 4.5 grok-4.5	xAI	closed	Public reference	43.2%
#6	Gemini 3.1 Pro Preview gemini-3.1-pro-preview	Google	closed	Public reference	42.6%

The top of this snapshot is led by Claude Sonnet 5 at 66.9%; third place is 17.2 points behind. The top-6 spread is 24.3 points.

Configuration-preserving evidencePublished source rows (6)Exact source labels, configurations, versions and dated observations. Same-snapshot canonical/configuration projections are shown once.

Every distinct dated source-published row for MLE-Bench, including exact model configuration labels
Exact published model / configuration	Canonical product	Score	Version / system	Evidence	Source / observed
As published in Gemini 3.6 Flash evaluation table. source label without registered configuration ID	Claude Sonnet 5	66.9%	Partial-30Provider evaluation harness	Public reference · not scored provider-reported	Gemini 3.6 Flash product page with evaluation table ↗Observed 2026-07-21 · checked 2026-07-21
gemini-3.6-flash via Gemini API; default sampling; single-attempt / pass@1 unless noted in Google eval methodology (July 2026). MLE-Bench Partial-30 Average Position Score; k=2 independent runs; bash terminal harness. source label without registered configuration ID	Gemini 3.6 Flash	63.9%	Partial-30Provider evaluation harness	Public reference · not scored provider-reported	Gemini 3.6 Flash product page with evaluation table ↗Observed 2026-07-21 · checked 2026-07-21
As published in Gemini 3.6 Flash evaluation table for Gemini 3.5 Flash comparison. source label without registered configuration ID	Gemini 3.5 Flash	49.7%	Partial-30Provider evaluation harness	Public reference · not scored provider-reported	Gemini 3.6 Flash product page with evaluation table ↗Observed 2026-07-21 · checked 2026-07-21
As published in Gemini 3.6 Flash evaluation table. source label without registered configuration ID	GPT-5.6 Luna	47.6%	Partial-30Provider evaluation harness	Public reference · not scored provider-reported	Gemini 3.6 Flash product page with evaluation table ↗Observed 2026-07-21 · checked 2026-07-21
As published in Gemini 3.6 Flash evaluation table. source label without registered configuration ID	Grok 4.5	43.2%	Partial-30Provider evaluation harness	Public reference · not scored provider-reported	Gemini 3.6 Flash product page with evaluation table ↗Observed 2026-07-21 · checked 2026-07-21
As published in Gemini 3.6 Flash evaluation table for Gemini 3.1 Pro comparison. source label without registered configuration ID	Gemini 3.1 Pro Preview	42.6%	Partial-30Provider evaluation harness	Public reference · not scored provider-reported	Gemini 3.6 Flash product page with evaluation table ↗Observed 2026-07-21 · checked 2026-07-21

Compare top models

Claude Sonnet 5 vs Gemini 3.6 Flash Gemini 3.6 Flash vs Gemini 3.5 Flash Gemini 3.5 Flash vs GPT-5.6 Luna GPT-5.6 Luna vs Grok 4.5

About MLE-Bench

Definition and scoring

Organisation: OpenAI
Category: Coding
Version: Partial-30
Direction: higher is better
Ranking use: Reference
Contamination risk: Low

Machine-learning engineering benchmark where agents compete on Kaggle-style tasks; Google reports Average Position Score on the Partial-30 subset. 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 MLE-Bench measure?

Machine-learning engineering benchmark where agents compete on Kaggle-style tasks; Google reports Average Position Score on the Partial-30 subset.

Which model scores highest on MLE-Bench?

Claude Sonnet 5 by Anthropic currently leads with 66.9%.

How many models are evaluated on MLE-Bench?

6 canonical model products have a qualifying score, backed by 6 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

MLE-Bench

Machine-learning engineering benchmark where agents compete on Kaggle-style tasks; Google reports Average Position Score on the Partial-30 subset.

ReferencePartial-30Active6 models

Data verified 21 Jul 2026 · Methodology 2.3.0

Benchmark score on MLE-Bench