Skip to main content

Vals-hosted SWE-bench mirror (Vals SWE-bench mirror)

Vals AI hosted SWE-bench view for solving production software engineering tasks.

How BenchLM shows Vals SWE-bench mirror

BenchLM mirrors the public Vals AI Vals SWE-bench mirror leaderboard captured from https://www.vals.ai/benchmarks/swebench and updated by Vals on May 16, 2026. The snapshot preserves overall scores, uncertainty, latency, cost-per-test metadata, and task-level scores where Vals publishes them.

Vals SWE-bench 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.

48 Vals rows5 task viewspublic datasetTasks: Overall, 1-4 hours, 15 min - 1 hour, <15 min fix, >4 hoursDisplay only

Vals SWE-bench score on Vals SWE-bench mirror — May 16, 2026

BenchLM mirrors the published vals swe-bench score view for Vals SWE-bench mirror. GPT-5.5 leads the public snapshot at 82.60% , followed by Claude Opus 4.7 (82.00%) and Gemini 3.5 Flash (78.80%). BenchLM does not use these results to rank models overall.

48 modelsExternal benchmark mirrorsCurrentDisplay onlyUpdated May 16, 2026

The published Vals SWE-bench mirror snapshot is tightly clustered at the top: GPT-5.5 sits at 82.60%, while the third row is only 3.80 points behind. The broader top-10 spread is 6.20 points, so many of the published scores sit in a relatively narrow band.

48 models have been evaluated on Vals SWE-bench mirror. The benchmark falls in the External benchmark mirrors category. BenchLM tracks this category separately from its weighted global scoring system, so these results are best compared on the dedicated Korean benchmark views. Vals SWE-bench mirror is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About Vals SWE-bench mirror

Year

2026

Tasks

Software engineering issue-resolution tasks

Format

Accuracy score

Difficulty

Production software engineering

BenchLM keeps this separate from its canonical SWE-bench Verified page so Vals-hosted results remain secondary context rather than source-of-record data.

BenchLM freshness & provenance

Version

Vals SWE-bench mirror 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.

Vals SWE-bench score table (48 models)

1
GPT-5.5openai/gpt-5.5
82.60%
2
Claude Opus 4.7anthropic/claude-opus-4-7
82.00%
3
Gemini 3.5 Flashgoogle/gemini-3.5-flash
78.80%
4
Gemini 3.1 Pro Previewgoogle/gemini-3.1-pro-preview
78.80%
5
GPT-5.4openai/gpt-5.4-2026-03-05
78.20%
6
Claude Opus 4.6 Thinkinganthropic/claude-opus-4-6-thinking
78.20%
7
GPT-5.3 Codexopenai/gpt-5.3-codex
78.00%
8
Claude Sonnet 4.6anthropic/claude-sonnet-4-6
77.40%
9
DeepSeek V4 Prodeepseek/deepseek-v4-pro
77.40%
10
Claude Opus 4.5 20251101 Thinkinganthropic/claude-opus-4-5-20251101-thinking
76.40%
11
Gemini 3 Pro Previewgoogle/gemini-3-pro-preview
76.40%
12
GLM 5.1 Thinkingzai/glm-5.1-thinking
76.40%
13
Kimi K2.6 Thinkingkimi/kimi-k2.6-thinking
76.20%
14
GPT-5.2openai/gpt-5.2-2025-12-11
75.80%
15
Gemini 3 Flash Previewgoogle/gemini-3-flash-preview
75.00%
16
MiniMax M2.1minimax/MiniMax-M2.1
74.80%
17
Muse Sparkmeta/muse_spark
74.40%
18
MiniMax M2.5 Lightningminimax/MiniMax-M2.5-Lightning
74.20%
19
MiniMax M2.7minimax/MiniMax-M2.7
73.80%
20
Qwen3.6 Plusalibaba/qwen3.6-plus
73.40%
21
GPT-5.4 Miniopenai/gpt-5.4-mini-2026-03-17
73.00%
22
Qwen3.6 Max Previewalibaba/qwen3.6-max-preview
72.80%
23
GPT-5.2 Codexopenai/gpt-5.2-codex
72.40%
24
Grok 4.20 0309 Reasoninggrok/grok-4.20-0309-reasoning
72.20%
25
Grok 4.3grok/grok-4.3
71.40%
26
GLM 5 Thinkingzai/glm-5-thinking
71.40%
27
Qwen3.5 Plus Thinkingalibaba/qwen3.5-plus-thinking
71.20%
28
Kimi K2.5 Thinkingkimi/kimi-k2.5-thinking
70.00%
29
Claude Sonnet 4.5 20250929 Thinkinganthropic/claude-sonnet-4-5-20250929-thinking
70.00%
30
Qwen3.6 27balibaba/qwen3.6-27b
70.00%
31
GPT-5.4 Nanoopenai/gpt-5.4-nano-2026-03-17
69.80%
32
GPT-5.1openai/gpt-5.1-2025-11-13
69.80%
33
GLM 4.7zai/glm-4.7
69.40%
34
GPT-5openai/gpt-5-2025-08-07
69.00%
35
DeepSeek V3p2 Thinkingfireworks/deepseek-v3p2-thinking
67.60%
36
Claude Haiku 4.5 20251001 Thinkinganthropic/claude-haiku-4-5-20251001-thinking
66.60%
37
Qwen3.5 Flashalibaba/qwen3.5-flash
64.40%
38
Gemini 3.1 Flash Lite Previewgoogle/gemini-3.1-flash-lite-preview
62.80%
39
Devstral 2512mistralai/devstral-2512
62.80%
40
GPT-5 Miniopenai/gpt-5-mini-2025-08-07
60.80%
41
Kimi K2 Thinkingkimi/kimi-k2-thinking
60.20%
42
Grok 4 0709grok/grok-4-0709
57.80%
43
Gemini 2.5 Progoogle/gemini-2.5-pro
54.40%
44
Grok 4 Fast Reasoninggrok/grok-4-fast-reasoning
45.40%
45
Grok 4.1 Fast Reasoninggrok/grok-4-1-fast-reasoning
41.40%
46
Mistral Large 2512mistralai/mistral-large-2512
41.40%
47
GPT Oss 120bfireworks/gpt-oss-120b
33.60%
48
Command A 03 2025cohere/command-a-03-2025
7.80%

FAQ

What does Vals SWE-bench mirror measure?

Vals AI hosted SWE-bench view for solving production software engineering tasks.

Which model leads the published Vals SWE-bench mirror snapshot?

GPT-5.5 currently leads the published Vals SWE-bench mirror snapshot with 82.60% vals swe-bench score. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on Vals SWE-bench mirror?

48 AI models are included in BenchLM's mirrored Vals SWE-bench mirror snapshot, based on the public leaderboard captured on May 16, 2026.

Last updated: May 16, 2026 · mirrored from the public benchmark leaderboard

The AI models change fast. We track them for you.

For engineers, researchers, and the plain curious — a weekly brief on new models, ranking shifts, and pricing changes.

Free. No spam. Unsubscribe anytime.