SWE-bench Verified (Epoch AI)
33 models with a score. Leader: Opus 4.7 with 83.5 %.
| Benchmark profile | |
|---|---|
| Task | allgemein - Fehlerbehebung in echten Code-Repositories |
| Metric | Anteil geloester Aufgaben aus 500 geprueften GitHub-Issues, eigener Agentenrahmen von Epoch |
| Direction | ↑ higher is better |
| Limitation | Nur 6 Katalogmodelle gemessen, meist mit hoher Reasoning-Stufe - kleine und offene Modelle fehlen fast ganz. Werte haengen am Agentenrahmen und sind mit der offiziellen swebench.com-Rangliste nicht vergleichbar. Fuer die Anreicherung das fernste Signal. |
| As of | 14 Sept 2026 |
| Source | epoch.ai |
| Licence | CC BY 4.0 — Epoch AI |
Ranking (33)
No entry matches the filter.
| Model | Rank | Score | Developer | EU processing: InputEU in | EU processing: OutputEU out | all providers: Inputall in | all providers: Outputall out | Open weights | GPU memory (4-bit, derived)GPU memory |
|---|---|---|---|---|---|---|---|---|---|
| Opus 4.7 | 1 | 83.5 % | Anthropic | €4.88 | €24.40 | €2.24 | €11.20 | – no | – |
| Gemini 3.5 Flash | 2 | 79.3 % | €1.48 | €8.89 | €0.67 | €4.03 | – no | – | |
| glm-5.2 | 3 | 78.7 % | Z.ai | €0.81 | €2.91 | €0.15 | €0.48 | ✓ yes | from 419 GB |
| Opus 4.6 | 3 | 78.7 % | Anthropic | €4.77 | €23.84 | €2.24 | €11.20 | – no | – |
| DeepSeek V4 Pro | 5 | 77.6 % | DeepSeek | €0.90 | €1.79 | €0.90 | €1.79 | ✓ yes | from 884 GB |
| Qwen 3.7 Max | 6 | 77.3 % | Alibaba Cloud | – | – | €1.12 | €3.36 | – no | – |
| GPT-5.4 | 7 | 76.9 % | OpenAI | €2.60 | €13.87 | €1.12 | €6.72 | – no | – |
| Kimi K2.6 | 8 | 76.7 % | Moonshot AI | €0.43 | €2.67 | €0.35 | €1.61 | ✓ yes | from 569 GB |
| Opus 4.5 | 8 | 76.7 % | Anthropic | €4.77 | €23.85 | €2.24 | €11.20 | – no | – |
| gemini-3.1-pro-preview-customtools | 10 | 75.6 % | – | – | €0.90 | €5.37 | – no | – | |
| Gemini 3 Flash Preview | 11 | 75.4 % | – | – | €0.22 | €1.34 | – no | – | |
| Claude Sonnet 4.6 | 12 | 75.2 % | Anthropic | €2.87 | €14.31 | €1.34 | €6.72 | – no | – |
| GPT-5.3 Codex | 13 | 74.8 % | OpenAI | – | – | €1.57 | €12.54 | – no | – |
| GLM 5.1 | 14 | 74.2 % | Z.ai | €1.24 | €3.90 | €0.81 | €2.60 | ✓ yes | from 419 GB |
| GPT-5.2 | 15 | 73.8 % | OpenAI | €1.72 | €13.79 | €0.78 | €6.27 | – no | – |
| Kimi K2.5 (Retiring) | 15 | 73.8 % | Moonshot AI | €0.58 | €3.22 | €0.27 | €1.34 | ✓ yes | from 554 GB |
| gpt-5 | 17 | 73.6 % | OpenAI | €0.62 | €4.93 | €0.56 | €4.48 | – no | – |
| GPT-5 2025-08-07 | 17 | 73.6 % | OpenAI | – | – | €0.56 | €4.48 | – no | – |
| Claude Opus 4.1 (retired, except on Claude Models on Google Cloud (Google managed) and Bedrock) | 19 | 73.3 % | Anthropic | – | – | €6.72 | €33.59 | – no | – |
| GLM-5 | 20 | 72.1 % | Z.ai | €1.15 | €3.68 | €0.45 | €1.43 | – | – |
| Sonnet 4.5 | 21 | 71.3 % | Anthropic | €2.68 | €13.42 | €1.34 | €6.72 | – no | – |
| Claude Opus 4 (retired, except on Claude Models on Google Cloud (Google managed)) | 22 | 70.7 % | Anthropic | – | – | €6.72 | €33.59 | – no | – |
| gpt-5.1 | 23 | 68 % | OpenAI | €1.23 | €9.84 | €0.56 | €4.48 | – no | – |
| GPT-5 mini | 24 | 64.7 % | OpenAI | €0.25 | €1.97 | €0.11 | €0.90 | – no | – |
| o3 | 25 | 62.3 % | OpenAI | – | – | €0.90 | €3.58 | – no | – |
| o3 2025-04-16 | 25 | 62.3 % | OpenAI | €0.99 | €3.94 | €0.90 | €3.58 | – no | – |
| Claude 3.7 Sonnet | 27 | 61 % | Anthropic | – | – | €1.34 | €6.72 | – | – |
| Qwen 3.6 Plus | 28 | 57.9 % | Alibaba Cloud | – | – | €0.45 | €2.69 | – no | – |
| Gemini 2.5 Pro | 29 | 57.6 % | €1.34 | €8.94 | €0.56 | €4.48 | – no | – | |
| GPT-4.1 | 30 | 48.5 % | OpenAI | €1.97 | €7.87 | €0.90 | €3.58 | – no | – |
| GPT-4.1-2025-04-14 | 30 | 48.5 % | OpenAI | €0.99 | €3.94 | €0.90 | €3.58 | – no | – |
| GPT-4o | 32 | 31 % | OpenAI | €2.39 | €9.55 | €1.12 | €4.48 | – no | – |
| GPT-4o-2024-1120 | 32 | 31 % | OpenAI | €2.46 | €9.85 | €1.12 | €4.48 | – no | – |
Equal scores share a rank. When filtering, the rank stays the position in the whole list. GPU memory “from …” is a lower bound (parameters × 0.55 GB + 4 GB). Prices: € per 1M tokens. Source: providers' pricing pages, read and checked by searchit. As of 7 Nov 2026.
Price and performance
Scores: SWE-bench Verified (Epoch AI), CC BY 4.0 — Epoch AI. Taken unchanged from the source.