DeepSeek V4 Pro launches, rivals Fable5 in agent benchmarks at a fraction of the price
AGI Hunt · wechat · 2026-08-12
DeepSeek V4 Pro is now officially available (version DeepSeek-V4-Pro-0813), with agent capabilities close to Fable5: TerminalBench 87.9 vs 88.0, and even surpassing Fable on Cybergym. Pricing is 3 CNY per million input tokens and 6 CNY per million output tokens, compared to Fable's $10 and $50, a tens-fold difference. API access is live. DeepSeek Harness is also coming, with some users in beta. Additionally, Alibaba's Qwen3.8-2.4T-A95B weights have been open-sourced on HuggingFace: a 2.4T-parameter MoE with 95B active, natively supporting vision and agent tool calling.
Related event: DeepSeek V4 Pro Officially Released with Disruptive Pricing and 1M Context(12 posts)→
More from Models
- Dev Reflects on AI Coding: Models Make Basic Reasoning Errors, Hand-Coding Wins — jsuarez · 2026-08-13
- Grok Hits SOTA on Databricks' OfficeQA Pro V2 Evaluation — GavinSBaker · 2026-08-13
- Grok 4.6 hits Vercel AI Gateway with 500K token context window — soleio · 2026-08-13
- xAI Grants SuperGrok Users a Free Weekly Usage Reset — XFreeze · 2026-08-13
- Krea, FLUX, and MiniMax Open-Weights: Promises vs Reality — felixsanz · 2026-08-13
- Anthropic Slammed by Users for Extra 'Fast Mode' Fees on Claude Code — Neel_MynO · 2026-08-13