--- format: "aidr-story-markdown/v1" id: "d05d8bf7ca19e0001a4ab98c3020a6707c61bf5b14e834a177779b810e9f3279" canonical_url: "https://aidr.today/d05d8bf7?lang=en" title: "DeepSeek V4 Pro Hits 88.5% on Coding Benchmarks, Beating Fable 5 at 90x Lower Cost" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-08-15T05:35:32.000Z" category: "Models" topics: ["deepseek","benchmarks","unreliable"] source_urls: ["https://huggingnews.com/ai/update-deepseek-v4-pro-hits-885percent-on-coding-benchmarks-beating-fabl-bc8a1d2b","https://x.com/togethercompute/status/2088468211159171246"] summary: "Together AI tested the DeepSeek-V4 Pro 0813 model against GPT-5.6 Sol and Fable 5 using the DeepSWE benchmark for software engineering. The model achieved an 88.5% pass@4 rate at a cost of $0.24 per task, making it 35 times cheaper than Sol and 90 times cheaper than Fable. DeepSeek transitioned the V4 Pro model to a production release on Aug. 13 to support performance gains for agent workflows. New dynamic pricing for the V4 lineup takes effect on Aug. 16 to move usage away from congested periods." --- # DeepSeek V4 Pro Hits 88\.5% on Coding Benchmarks, Beating Fable 5 at 90x Lower Cost > [Open the canonical story]() **Published:** 2026-08-15T05:35:32.000Z **Category:** Models **Topics:** deepseek, benchmarks, unreliable ## Summary Together AI tested the DeepSeek\-V4 Pro 0813 model against GPT\-5\.6 Sol and Fable 5 using the DeepSWE benchmark for software engineering\. The model achieved an 88\.5% pass@4 rate at a cost of $0\.24 per task, making it 35 times cheaper than Sol and 90 times cheaper than Fable\. DeepSeek transitioned the V4 Pro model to a production release on Aug\. 13 to support performance gains for agent workflows\. New dynamic pricing for the V4 lineup takes effect on Aug\. 16 to move usage away from congested periods\. ## Sources - [Story source]() - [Story source]()