DeepSeek-V4 Pro benchmark: Top performance at ultra-low cost
zainhas · x · 2026-08-15
An analysis by Together AI reveals that DeepSeek-V4 Pro 0813 achieves an 88.5% pass@4 score on DeepSWE software engineering tasks, outperforming GPT-5.6 Sol and Fable 5. It costs only $0.24 per task, making it 35x cheaper than Sol and 90x cheaper than Fable.
Related event: DeepSeek-V4 Pro Leads in Software Engineering Tasks(2 posts)→
More from Models
- DeepSeek Developing Flash Variants to Match 3T Model Coding Performance — bindureddy · 2026-08-15
- Qwen3.8-27B performance sparks buzz, netizens joke about Google's reaction — max_paperclips · 2026-08-15
- Developer quantizes AI9Stars G9v3-39A5B to GGUF and creates llama.cpp fork for support — linuxid10t · 2026-08-15
- Gemini 3.5 Pro checkpoint renamed to Gemini 3.7 Flash High, sparking speculation — Rare_Bunch4348 · 2026-08-15
- Perplexity releases Agent API and web search benchmarks — AravSrinivas · 2026-08-15
- Debate: DeepSeek Performance and Skepticism About "10T Models" — teortaxesTex · 2026-08-15