The gap between "can write a function" and "can fix a real bug in a real codebase" is where SWE-Bench Verified lives. As of July 2026, 103 models have been evaluated on this benchmark, and the leaderboard tells a fascinating story about who can actually ship code, not just generate snippets.
What Is SWE-Bench Verified?
SWE-Bench Verified is a curated subset of 500 software engineering problems drawn from real GitHub issues across 12 popular Python repositories. Each problem requires the model to understand an issue description, navigate an actual codebase, and produce a patch that resolves the issue. Human annotators verified every problem to ensure quality.
This is not LeetCode. This is not "write a function to reverse a string." A model scoring 80% here can genuinely contribute to a development workflow — debugging, refactoring, and resolving issues in production codebases with multiple files, classes, and dependencies.
The benchmark was introduced in the paper "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" by Jimenez et al. When published in October 2023, the best model (Claude 2) solved just 1.96% of issues. Today's leader solves 95%.
The Top 10 as of July 2026
| Rank | Model | Developer | SWE-Bench Verified | Open Source? |
|---|---|---|---|---|
| 1 | Claude Fable 5 | Anthropic | 95.0% | No |
| 2 | Claude Opus 4.8 | Anthropic | 88.7% | No |
| 3 | GPT-5.6 Sol (preview) | OpenAI | 87.4% | No |
| 4 | Claude Sonnet 5 | Anthropic | 85.2% | No |
| 5 | Gemini 3.5 Pro | Google DeepMind | 83.9% | No |
| 6 | GPT-5.5 | OpenAI | 82.1% | No |
| 7 | Claude Sonnet 4.6 | Anthropic | 80.8% | No |
| 8 | DeepSeek-V4-Pro-Max | DeepSeek | 80.6% | Yes (MIT) |
| 9 | Grok 4 | xAI | 78.3% | No |
| 10 | LongCat-2.0 | Meituan | 59.5% | Yes (MIT) |
Scores sourced from llm-stats.com and developer publications. Some scores are self-reported.
What the Rankings Tell Us
Anthropic Dominates the Top
Four of the top seven models are Anthropic's. Claude Fable 5 at 95% is a remarkable number — it means the model resolves 475 out of 500 real-world engineering issues correctly. Fable 5 is Anthropic's specialized coding and reasoning model, and while it's the most expensive Claude variant, the performance gap is substantial.
But the practical sweet spot is Sonnet 5 at 85.2%. It delivers nearly equivalent real-world coding ability at a fraction of Fable 5's cost. For teams building agentic coding pipelines, Sonnet 5 is the value leader.
GPT-5.6 Is Coming
GPT-5.6 Sol sits at #3 with 87.4% — but only in restricted preview. Currently limited to roughly 20 government and partner organizations due to U.S. export control considerations, GPT-5.6 is expected to get a broader public release in mid-July 2026. If its preview numbers hold, it will compete directly with Opus 4.8 for the #2 spot.
The Open-Source Champion: DeepSeek V4 Pro Max
DeepSeek V4 Pro Max is the highest-ranked open-source model at 80.6% — and it carries an MIT license. That means you can run it locally, fine-tune it, and use it commercially without API restrictions. For organizations that need data sovereignty or want to avoid vendor lock-in, DeepSeek V4 is the answer.
The official V4 launch (graduating from preview) is expected mid-July 2026, with new peak/off-peak API pricing. The open weights are already available.
Google and xAI Are Close Behind
Gemini 3.5 Pro at 83.9% and Grok 4 at 78.3% round out the frontier tier. Google's strength is its massive context window and multimodal capabilities, while Grok 4 brings real-time data access through X integration.
Beyond the Top 10: The Open-Source Landscape
The open-source coding model space has matured significantly in 2026:
- DeepSeek V4 Pro Max (80.6%) — The clear leader. 1.6T parameters, MIT license, trained on a mix of Chinese and international hardware.
- LongCat-2.0 (59.5%) — Meituan's 1.6T Mixture-of-Experts model. Lower on the leaderboard but impressive for a model trained from scratch on Chinese infrastructure.
- Various Llama derivatives — Community fine-tunes continue to push Llama-based models into the 60-70% range.
The open-source gap to proprietary models has narrowed from ~40 points in early 2025 to roughly 15 points today. DeepSeek's next iteration could close it further.
How to Choose a Coding Model in July 2026
For maximum capability: Claude Fable 5
If budget is not a concern and you need the absolute best automated code repair and generation, Fable 5 is unmatched. Best for enterprise pipelines where correctness is critical.
For value and production use: Claude Sonnet 5
At 85.2% with Sonnet-tier pricing, this is the model most teams should default to. It plans, uses tools, and handles long-horizon coding tasks autonomously.
For open-source and self-hosting: DeepSeek V4 Pro Max
The only open-source model above 80% on SWE-Bench Verified. MIT license means no restrictions. Run it on your own hardware for complete data control.
For cost-sensitive API use: DeepSeek V4 Flash
DeepSeek's smaller variant (284B) trades some accuracy for significantly lower cost. With the upcoming peak/off-peak pricing, off-peak hours offer exceptional value.
For Google ecosystem teams: Gemini 3.5 Pro
Strong scores across coding, reasoning, and multimodal tasks, with the advantage of deep Google Cloud integration and massive context windows.
What's Next
With GPT-5.6's public launch imminent and DeepSeek V4's official release coming mid-July, the coding model landscape will shift again within weeks. The trend is clear: scores that seemed impossible in 2023 (when Claude 2 managed 1.96%) are now table stakes. The frontier is moving toward 100%, and the open-source community is keeping pace.
The question is no longer "can AI fix real bugs?" It's "how much of your engineering workflow should you hand over?"