LiveBench update: Claude Fable 5.1 tops real-world leaderboard at 85.7, DeepSeek leads open source
bindureddy · x · 2026-10-28
bindureddy shared the latest LiveBench leaderboard, arguing it now reflects real-world performance.
Key rankings (with cost per successful task):
- Claude Fable 5.1 tops the board at 85.7, costing $6.904 per successful task
- GPT-6 Astra follows closely at 85.6 ($4.548)
- Claude Opus 5.5 at 85.1, best in Design (86.9)
- Gemini 4 Argon beats GPT-6.1 Terra on Data Analysis (85.2)
- DeepSeek V4.1 Flash 1017 is the top open-source model at 74.0, also the cheapest at $0.181 per task
The board scores Agentic, Design, Browser Use, Data Analysis and Research separately; GPT-6.1 Sol stands out on Browser Use (93.3) and cost-efficiency.
More from Models
- Permissive flagship open-model licenses fell from 75% to 35% in months — bookwormengr · 2026-10-28
- Anthropic Gets 80% of Revenue From API Pricing; DeepSeek Costs 20x Less — bindureddy · 2026-10-28
- Anthropic models burn 2-3x tokens and cost 10-50x more on FrontierCode 2.0 — teortaxesTex · 2026-10-28
- Last full day on grandfathered 20x OpenAI plans before 10x usage switch — nptacek · 2026-10-28
- Dev: Claude lags far behind OpenAI models on memory and computer use — iruletheworldmo · 2026-10-28
- Tweak one neuron and Qwen3.5-4B turns into a screaming oracle: PEFT activation steering demos — LysandreJik · 2026-10-28