BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases9

SWE-bench (Vals)

Vals AI's own SWE-bench run. Shown for reference.

As of 25 Oct 2026, Claude Opus 5 leads SWE-bench (Vals) on BenchLeader with 97.0%, ahead of DeepSeek V4 Pro at 96.4%, across 85 model configurations with a published result.

Published by
Vals AI
Category
Coding
Index weight
Reference only
Models
85
Data as of
25 Oct 2026
Newest result
1 Sept 2026

Vals AI (vals.ai).

What the test looks like

SWE-bench Verified is 500 real GitHub issues from popular Python repositories, each screened by human annotators to make sure the task is well specified and the tests are fair. The model reads the issue, edits the repository, and its patch is accepted only if the hidden tests pass.

How it is scored

Percent of the 500 issues resolved in Vals AI's own agent harness, run by Vals AI, not the bash-only mini-SWE-agent harness of the official board.

What to keep in mind

Vals AI's own run of SWE-bench Verified. Shown for reference; the official bash-only board is counted in the index. See also the SWE-bench Verified (bash only) board.

Top 15
  1. 1Claude Opus 597.0%
  2. 2DeepSeek V4 Pro (max)96.4%
  3. 3GPT-5.6 Sol (max)96.2%
  4. 4Grok 4.6 (high)95.6%
  5. 5GPT-5.6 Terra (max)95.4%
  6. 6GLM 5.3 (max)95.4%
  7. 7Claude Fable 595.0%
  8. 8Kimi K393.4%
  9. 9GPT-5.6 Luna (max)93.0%
  10. 10GLM 5.3 Flash (max)92.0%
  11. 11DeepSeek V4 Flash (high)88.8%
  12. 12Claude Opus 4.888.6%
  13. 13Muse Spark 1.2 (xhigh)86.6%
  14. 14Grok 4.5 (high)86.6%
  15. 15Qwen3.8 27B (xhigh)86.0%
85 of 85
#
1Claude Opus 597.0%65.12026-09-01
2DeepSeek V4 Promax96.4%60.82026-09-01
3GPT-5.6 Solmax96.2%68.02026-09-01
4Grok 4.6high95.6%62.92026-09-01
5GPT-5.6 Terramax95.4%64.42026-09-01
6GLM 5.3max95.4%63.52026-09-01
7Claude Fable 595.0%66.52026-09-01
8Kimi K393.4%58.82026-09-01
9GPT-5.6 Lunamax93.0%59.92026-09-01
10GLM 5.3 Flashmax92.0%54.22026-09-01
11DeepSeek V4 Flashhigh88.8%57.32026-09-01
12Claude Opus 4.888.6%60.12026-09-01
13Muse Spark 1.2xhigh86.6%63.22026-09-01
14Grok 4.5high86.6%59.32026-09-01
15Qwen3.8 27Bxhigh86.0%56.32026-09-01
16Claude Opus 4.8 Claude Code85.8%–2026-09-01
17Qwen3.8 Maxmax85.6%63.32026-09-01
18GLM 5.2max82.8%62.32026-09-01
19GPT-5.5xhigh82.6%66.12026-09-01
20Inkling Small82.2%54.92026-09-01
21Claude Opus 4.782.0%63.22026-09-01
22Muse Spark 1.1xhigh82.0%61.62026-09-01
23Gemini 3.7 Flashhigh80.8%63.52026-09-01
24Gemini 3.8 Flashhigh80.0%64.42026-09-01
25Gemini 3.6 Flashhigh79.6%62.02026-09-01
26Claude Sonnet 579.6%55.92026-09-01
27Composer 2.579.6%–2026-09-01
28Gemini 3.5 Flashhigh78.8%63.12026-09-01
29Gemini 3.1 Prohigh78.8%60.32026-09-01
30GPT-5.4xhigh78.2%64.72026-09-01
31Claude Opus 4.6thinking78.2%62.52026-09-01
32Kimi K2.7 Code78.2%55.82026-09-01
33GPT-5.3-Codexxhigh78.0%63.82026-09-01
34Claude Sonnet 4.677.4%57.12026-09-01
35Gemini 3 Prohigh76.4%60.02026-09-01
36Claude Opus 4.5thinking76.4%59.92026-09-01
37GLM 5.176.4%55.92026-09-01
38GPT 5.5 Codex76.4%–2026-09-01
39Kimi K2.676.2%55.32026-09-01
40GPT 5.5 Factory76.2%–2026-09-01
41GPT-5.2xhigh75.8%60.72026-09-01
42Gemini 3 Flashhigh75.0%58.32026-09-01
43MiniMax M375.0%56.42026-09-01
44Gemini 3.5 Flash Litehigh75.0%47.92026-09-01
45MiniMax-M2.174.8%53.82026-09-01
46Muse Spark74.4%64.32026-09-01
47MiniMax M2.574.2%53.62026-09-01
48MiMo-V2.5-Pro74.0%54.92026-09-01
49MiniMax M2.773.8%55.62026-09-01
50Qwen3.6 Plus73.4%58.82026-09-01
51GPT-5.4 Minixhigh73.0%56.32026-09-01
52Qwen3.6 Maxmax72.8%62.32026-09-01
53GPT-5.2-Codexhigh72.4%–2026-09-01
54Grok 4.20thinking72.2%58.62026-09-01
55GLM-5thinking71.4%58.92026-09-01
56Grok 4.3high71.4%57.02026-09-01
57Qwen3.5 Plusthinking71.2%57.32026-09-01
58Kimi K2.5thinking70.0%57.62026-09-01
59Claude Sonnet 4.5thinking70.0%52.62026-09-01
60Qwen3.6 27B70.0%48.32026-09-01

More coding benchmarks · All LLM benchmarks

Cite as: BenchLeader, “SWE-bench (Vals) leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/vals_swebench.

SWE-bench (Vals): questions

What does SWE-bench (Vals) measure?
SWE-bench Verified is 500 real GitHub issues from popular Python repositories, each screened by human annotators to make sure the task is well specified and the tests are fair. The model reads the issue, edits the repository, and its patch is accepted only if the hidden tests pass. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads SWE-bench (Vals)?
Claude Opus 5 leads SWE-bench (Vals) with 97.0% as of 25 Oct 2026, ahead of DeepSeek V4 Pro at 96.4%.
How many models have SWE-bench (Vals) results?
85 model configurations have a SWE-bench (Vals) result on BenchLeader, all taken from Vals AI.
Who runs SWE-bench (Vals) and how often is it updated?
SWE-bench (Vals) is published by Vals AI. BenchLeader re-reads the published results once a day and records the date each result was published.
Does SWE-bench (Vals) count toward the BenchLeader Index?
No. SWE-bench (Vals) is shown for reference but left out of the composite index.