Benchmark profile

FLTEval

A repository-level Lean 4 proof engineering benchmark that measures whether a model can complete formal proofs and correctly define new mathematical concepts inside realistic FLT project pull requests.

Data verified July 23, 2026

How BenchLM shows FLTEval right now

BenchLM is tracking FLTEval in the local dataset, but exact-source verification records for these rows are still being attached. To avoid a blank benchmark page, BenchLM shows the current tracked rows below as a display-only reference table.

These tracked rows are useful for inspection and spot-checking, but until exact-source attachments are completed they should not be treated as fully verified public benchmark rows.

4 tracked modelsLocal tracked rowsAwaiting exact-source attachmentsDisplay only

Leanstral: Open-Source foundation for trustworthy vibe-coding

Tracked score on FLTEval — July 23, 2026

BenchLM mirrors the published tracked score view for FLTEval. Claude Opus 4.6 leads the public snapshot at 39.6% , followed by Claude Sonnet 4.6 (23.7%) and Claude Haiku 4.5 (23%). BenchLM does not use these results to rank models overall.

Claude Opus 4.6

Anthropic

claude-opus-4-6

39.6%

Overall —

Claude Sonnet 4.6

Anthropic

claude-sonnet-4-6

23.7%

Overall —

Claude Haiku 4.5

Anthropic

claude-haiku-4-5

23%

Overall —

4 modelsCodingCurrentDisplay onlyUpdated July 23, 2026

Tracked score table (4 models)

Score

Claude Opus 4.6Anthropic

39.6%

Claude Sonnet 4.6Anthropic

23.7%

Claude Haiku 4.5Anthropic

23%

LeanstralMistral

21.9%

The published FLTEval snapshot places Claude Opus 4.6 first at 39.6%. The third row is 16.6 points behind. The broader top-10 range is 17.7 points, so the table still separates the published systems.

4 models have been evaluated on FLTEval. The benchmark falls in the Coding category. This category carries a 20% weight in BenchLM.ai's overall scoring system. FLTEval is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About FLTEval

Year

2026

Tasks

FLT project pull requests

Format

Lean 4 repository task completion

Difficulty

Formal verification / proof engineering

FLTEval is designed to move evaluation beyond isolated competition-math problems. Instead of proving one-off statements, models must operate inside realistic formal repositories and finish pull-request-style Lean 4 work with Lean itself acting as a verifier.

Leanstral: Open-Source foundation for trustworthy vibe-coding Public benchmark source

BenchLM freshness & provenance

Version

FLTEval 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.

FAQ

What does FLTEval measure?

Which model leads the published FLTEval snapshot?

Claude Opus 4.6 currently leads the published FLTEval snapshot with 39.6% tracked score. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on FLTEval?

4 AI models are included in BenchLM's mirrored FLTEval snapshot, based on the public leaderboard captured on July 23, 2026.

Last updated: July 23, 2026 · mirrored from the public benchmark leaderboard

Choose a model with this week’s evidence

Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.

One email each week. Unsubscribe anytime.