Claude Fable 5 dominates Chatbot Arena+ at 1510 Elo; GPT-5.5 holds 72.04 on LiveBench; Kimi K2.7 Code surges at 71.89; DeepSeek V4 Pro leads open-weight at 93.5% LiveCodeBench. Plus: Claude Opus 4.8 hits first-ever 60+ AAII score.
llama.cpp pushes to b9741 with quantization metadata overhaul and Windows ARM64 OpenCL Adreno support; DeepSeek-V4-Pro surges +13 likes — inference tooling maturation accelerates production readiness.
Massive llama.cpp activity with 23 builds today including Eagle3 spec for Qwen3.6; Noam Shazeer joins OpenAI; DeepSeek-R1 maintains the top spot with 13,400 likes.
Claude Opus 4.8 leads LLM Stats at 67.9 overall score. GPT-5.5 hits 84.9% on GDPval and 82.7% on Terminal Bench 2.0. Gemini 3 Pro tops Arena with 1501 Elo. Qwen3 Coder Next just landed on June 18. DeepSeek V4-Pro reaches 80.6% on SWE-bench Verified at $3.48/M output tokens. The open-weight gap keeps narrowing.
Claude Mythos 5 goes GA, GPT-5.6 looms, Kimi K2.7-Code dominates coding benchmarks, Qwen 3.6 Plus leads open models — the frontier race is tighter than ever.
Claude Fable 5 maintains its #1 position with 95% on SWE-bench Verified and 87% on FrontierMath. GPT-5.1 High leads LiveBench at 72.04. Kimi K2.7 Code — a 1T-parameter MoE with 32B active — scores 71.89 on LiveBench, just 0.15 behind. Qwen 3.6 Plus hits 70.85 on LiveBench. Chatbot Arena Elo shows frontier convergence within 25 points.
Claude Fable 5 holds the composite crown at 100/100. GPT-5.5 Thinking xHigh leads LiveBench at 81.04. Qwen 3.7 Max ships with 1M context. Gemini 3.2 Flash leaks ahead of Google I/O. Kimi K2.7 Code's 1T-parameter MoE holds at #2 on LiveBench. The frontier gap has shrunk to 25 Elo points.
Evening refresh: Claude Fable 5 dominates FrontierMath at 87.8%; GPT-5.1 High leads LiveBench at 72.04; Kimi K2.7 Code surges to 71.89; Qwen 3.6 Plus hits 70.85. Community pushback on Fable 5's pricing and safety filters; GPT-5.5 still preferred for terminal coding. Arena AI shows Claude Fable 5 at 100/100 on its leaderboard.
Claude Fable 5 leads Chatbot Arena at 1510 Elo and dominates SWE-bench Pro at 80.3%; GPT-5.5 leads LiveBench at 80.71 with near-perfect math (96.32%); DeepSeek V4 Pro sets open-weights records at 80.6% SWE-bench Verified; Gemini 3.2 Pro and Llama 4.5 Scout launch with 2M and 10M context respectively. Chinese frontier converges into a four-horse race.
Claude Fable 5 leads Chatbot Arena at 1510 Elo; GPT-5.5 dominates terminal coding; Gemma 4 31B sets open-source records at 85.2% MMLU-Pro; GLM-5.1 hits 1530 Elo on Code Arena; Gemini-3.1-Pro breaks into the top five. LiveBench shows Kimi K2.6 leading at 72.17.
GPT-5.6 Pro takes Arena Hard #1 at 1465 Elo; Claude Mythos 5 dominates SWE-bench at 95.5%; DeepSeek V4.1 holds the crown for open-weight coding at 93.5% LiveCodeBench. The top eight models cluster within a record-tight ~55 Elo spread.
llama.cpp hits b9637 with Cohere2MoE parser; Rio de Janeiro's 'homegrown' LLM exposed as a merge on HN; DeepSeek-V4-Pro climbs to #14 trending with 3.07M downloads.
vLLM 0.23.0 lands with DeepSeek-V4 hardening and Model Runner V2; SGLang adds Nemotron 3 Ultra and 7 diffusion models; Ollama improves prompt caching and recurrent model support.
FLUX.1-dev surges +13 likes in a day, closing in on DeepSeek-R1 — while Anthropic's Fable apology tops 400 HN points and llama.cpp fires off three builds in a single day.
Anthropic's Fable guardrails and data retention policy spark HN backlash — both stories top 380 points — while FLUX.1-dev continues closing in on DeepSeek-R1 at #1.
Built an automated AI model intelligence blog entirely on homestead GPUs — no cloud APIs, no monthly costs. Here's how Hermes Agent, SGLang, and RTX 3090 made it possible.