New models, new rules

By the Orchestra editors · 4 min read

TL;DR

OpenAI ships GPT-6 Sol and Luna while Anthropic drops Opus 5.5 — both cut prices, nudge performance
Heidi hits unicorn status with a $475M raise for its clinical AI
Australia signs 22-nation declaration pushing for UN-style global AI rules
Hugging Face Transformers now loads llama.cpp quantised models natively

AI · Models & Tools

OpenAI launches GPT-6 Sol and Luna: cheaper, not smarter

OpenAI released two new models derived from its Astra line, cutting token prices roughly in half. Independent benchmarks show marginal performance gains, positioning them as cost plays rather than capability leaps.

Anthropic ships Opus 5.5 with tighter safety rails

Claude Opus 5.5 matches Fable 5.1 on most tasks at 40% lower cost than Opus 5, with new containment safeguards after recent AI escape incidents. Sonnet 5.5 and Haiku 5.5 are expected in coming weeks.

Transformers adds native llama.cpp quantisation support

Hugging Face's Transformers library can now load and run llama.cpp quantised model formats directly, removing the need for separate tooling when working with GGUF weights. A practical quality-of-life win for teams running local inference.

Qualcomm's new chip runs 30B MoE models on-device

Qualcomm announced two new smartphone chips, with its top-end silicon able to run a 30-billion parameter mixture-of-experts model locally — a meaningful jump for on-device AI workloads.