Back to benchmarks
Coding BenchmarkVerified 39 days ago

SWE-bench Verified Leaderboard — June 2026

Real-world coding benchmark update: Claude Mythos 5 takes the top spot at ~78%, DeepSeek V4.1 Pro sits within 6 points of frontier closed models.

Claude Mythos 5, Claude Opus 4.7, GPT-5.6 Pro, GPT-5.6, DeepSeek V4.1 Pro, Qwen 3.7Multiple2026-06-21
Winner:Claude Mythos 5

Results dashboard

Overall score

Claude Mythos 5
78
Claude Opus 4.7
75
GPT-5.6 Pro
73
GPT-5.6
70
DeepSeek V4.1 Pro
69
Claude Sonnet 4.6
68
Qwen 3.7
66

Results table

AgentScoreTimeCostPass
Claude Mythos 578—— Fail
Claude Opus 4.775—— Fail
GPT-5.6 Pro73—— Fail
GPT-5.670—— Fail
DeepSeek V4.1 Pro69—— Fail
Claude Sonnet 4.668—— Fail
Qwen 3.766—— Fail

The task

SWE-bench Verified is a curated subset of real GitHub issues from popular Python repositories. The benchmark measures whether an agent can reproduce a bug, implement a fix, and pass the repository's own test suite end-to-end. It is the most-cited real-world coding benchmark for frontier LLMs.

The June 2026 leaderboard snapshot shows the top of the field clustered within a 12-point band, with open-weight models now within single digits of the frontier.

Scoring rubric

Criterion Weight Notes
Resolved issue rate 100% Percentage of benchmark instances where the agent's patch passes all tests.

Scores are rounded/directional; treat as approximate pending independent verification.

Methodology

  • Source: vendor disclosures, public SWE-bench Verified leaderboard at swebench.com, and third-party replication where available.
  • Updated monthly by Presenc AI.
  • Scores reflect the strongest reported configuration for each model/agent pairing.

Key findings

  • Claude Mythos 5 GA takes the top spot at ~78%, displacing Claude Opus 4.7.
  • DeepSeek V4.1 Pro closes the gap: open-weight performance is now ~6 points behind the leader.
  • Qwen 3.7 leads the Chinese frontier set on coding tasks.
  • GPT-5.6 Pro is competitive at ~73%, but lags Anthropic's top Claude variants on this benchmark.
  • The gap between top closed and top open-weight models has narrowed to single digits over the last 12 months.

Honest caveats

  • Scores are vendor-reported or aggregated; independent replication is limited.
  • Scaffold and tool access heavily influence results. The same model with different agent tooling may score differently.
  • SWE-bench focuses on Python open-source bugs; it does not cover proprietary codebases, security fixes, or cross-language maintenance.

When to choose which

  • Claude Mythos 5 / Opus 4.7: highest reported success rate on real GitHub issues; best for high-stakes code repair.
  • DeepSeek V4.1 Pro: strong open-weight option when data residency or cost matters.
  • Qwen 3.7: strong option for Chinese-language or regionally hosted deployments.
  • GPT-5.6 Pro: good balance of general capability and coding; verify in your own stack before betting on it.

Source