Skip to main content
Radar

Every change to the models you run, with its source and its date. Releases, price changes, retirements, API changes, and incidents.Every change to the models you run, with its source.

Follow model changes

VulcanBench v3

We mirror this table; we do not rank on it.

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

An open software-engineering benchmark built from real merged post-cutoff pull requests across Python, Rust, TypeScript, JavaScript, and Go repositories.

Benchmark score on VulcanBench v3 — September 18, 2026

We mirror the published score view for VulcanBench v3. Grok 4.5 leads the public snapshot at 89.9%, followed by Claude Fable 5 (89.5%) and DeepSeek V4 Flash 0731 (88.4%). We do not use these results to rank models overall.

15 modelsCodingCurrentDisplay onlyUpdated September 18, 2026

Benchmark score table (15 models)

Score
1
Grok 4.5xAI · Closed
89.9%
2
Claude Fable 5Anthropic · Closed
89.5%
3
DeepSeek V4 Flash 0731DeepSeek · Closed
88.4%
4
Claude Opus 5Anthropic · Closed
87.0%
5
GPT-5.6 SolOpenAI · Closed
87.0%
6
GPT-5.6 TerraOpenAI · Closed
87.0%
7
Grok 4.6xAI · Closed
87.0%
8
Muse Spark 1.2Meta · Closed
87.0%
9
DeepSeek V4 Pro (High)DeepSeek · Open weight
87.0%
10
GPT-5.6 LunaOpenAI · Closed
85.5%
11
Qwen3.8-27BAlibaba · Open weight
82.6%
12
Qwen3.8 MaxAlibaba · Open weight
81.2%
13
GLM-5.3Z.AI · Open weight
78.3%
14
Claude Haiku 4.5Anthropic · Closed
76.2%
15
Kimi K3Moonshot AI · Closed
73.7%

The published VulcanBench v3 snapshot places Grok 4.5 first at 89.9%. The third row is 1.5 points behind. The broader top-10 range is 4.4 points, so many of the published results sit in a relatively narrow band.

15 models have been evaluated on VulcanBench v3. The benchmark falls in the Coding category. This category carries a 20% weight in BenchLM.ai's overall scoring system. VulcanBench v3 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About VulcanBench v3

Year

2026

Tasks

23 post-cutoff repository tasks in the v3 report

Format

Pass@1 with low, medium, and high effort

Difficulty

Professional multi-file software engineering

VulcanBench v3 evaluates repository-level patches with deterministic hidden tests inside network-isolated Docker sandboxes. The August 24 board covers 23 tasks, 14 models, 39 model-by-effort columns, and 1,582 runs. We mirror each benchmark-owner best-per-model row as display-only evidence. Product-harness studies, extended-budget ablations, and other rows outside the raw uniform-harness board stay separate.

Freshness and provenance

Version

VulcanBench v3 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 VulcanBench v3 measure?

An open software-engineering benchmark built from real merged post-cutoff pull requests across Python, Rust, TypeScript, JavaScript, and Go repositories.

Which model scores highest on VulcanBench v3?

Grok 4.5 by xAI currently leads with a score of 89.9% on VulcanBench v3.

How many models are evaluated on VulcanBench v3?

15 AI models have been evaluated on VulcanBench v3 on BenchLM.

Last updated: September 18, 2026 · BenchLM version VulcanBench v3 2026

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.