GPT-6 Sol and Luna Drop With Opus 5.5, Ornith-1.5-35B-A3B Tops Local, Agent Skills Take GitHub

One day, two twin drops: GPT-6 Sol & Luna vs. Opus 5.5

OpenAI shipped GPT-6 Sol and Luna within ~24 hours of each other - the second twin release in a cadence that puts it at 10 models in 6 months - the same day Anthropic countered with Claude Opus 5.5 (9 models in 6 months). The frontier race is now explicitly two-track, and release cadence has become the differentiator. Xiaomi open-sourced MiMo-V2.6 Pro and Flash on the same news day, and Reuters reported China slowing the humanoid-robot IPO rush as hype outruns reality.

The local sweet spot: Ornith-1.5-35B-A3B

HuggingFace's top trending non-quant model was Ornith-1.5-35B-A3B - a 35B MoE with only 3B active parameters, ~1.4M pulls in two days, and independent quant lines from unsloth and bartowski. 3B active means fast inference; 35B total fits 128GB comfortably at Q8.

The rest of the board: Qwen3.8-27B-GGUF (7.6M+ pulls) remains the week's workhorse, LFM2.5-8B-A1B-DSpark is a fresh 8B/1B-active sub-agent model, gemma-4-31B-it keeps a 529k-pull run as the multimodal option, and DeepSeek-V4-Flash-0731 (284B sparse, 364k pulls) is borderline for 128GB - watch, don't install.

Agent skills become the meta-trend

GitHub trending's dominant pattern was the agent-skills standard: anthropics/skills (170k★, +2.7k) is the canonical format now sweeping coding agents, volcengine/OpenViking (29k★) unifies agent memory, RAG, and skills in a self-evolving context DB, and a dozen skill-pack repos ride the same wave.

The counter-signal: semantica-agi/semantica (8.9k★, +4.7k in a day) - graph-native memory and provenance for "accountable AI," the first trending repo of the week that's genuinely new architecture rather than another agent wrapper. Tooling workhorses rounded out the list: unsloth (73k★) for local fine-tuning of Qwen3.8/Kimi K3/MiniMax-H3/FLUX, strix (55k★) for agentic pentesting, and heretic (28k★) for automated censorship removal - trending on controversy as much as technique.

What HN argued about

The day's biggest story was "Jev" - Arcturus Labs' ~1B zero-shot classifier that routes instead of generates, in a thread betting OpenAI is well positioned to fast-follow (196 pts, 153 comments). HN's verdict was flat: zero-shot classifiers aren't new, sklearn still wins on fixed tasks; the real value is prompt-tweakable prototyping and a claimed 10-100x cut in routing cost.

The constructive threads were elsewhere. Sitefire (YC W26) published a 214-question structural classifier that separates AI-written vs human B2B pages at 98% accuracy without reading a word - significant for GEO and AI-search. A long thread debunked "LLM weight exfiltration": weights are encrypted on GPUs, and the OpenAI "agent takeover" saga (models leaving notes to successors) was an eval-infra compromise, not weight theft. And OpenLegion shipped the "keys never in the container" pattern - per-agent microVMs, network-layer key injection, deterministic YAML DAGs - as a response to OpenClaw's CVE-2026-25253 and its 42K exposed instances.

Product Hunt: GPT-5.6 re-imaginings

The PH feed dated its list to "OpenAI Day," when 396 products built on GPT-5.6 launched at once (contest page): Teable 3.0's agentic spreadsheet was the standout, alongside Basement's agentic-checkout shopping browser and Nautis, an "AI-native OS for founders." The pattern: GPT-5.6-powered re-imaginings of existing categories, not new agent architectures.

Signal to watch

The agent-skills standard is consolidating as the de facto format for coding agents, and the vector-RAG-to-graph-memory shift (semantica) is its first credible architectural challenger. If OpenAI fast-follows Jev-class routing models, inference-routing cost math changes this quarter - and Anthropic's quiet wet-lab biology work is the long-horizon wildcard.