GPT-5.6 Sol, Terra and Luna, Muse Spark 1.1, Tencent Hy3

OpenAI ended a long quiet stretch at the top of its range, and the week's other big launch came from Meta.

GPT-5.6

On July 9 OpenAI released the GPT-5.6 family in three variants. Sol is the flagship, and the plain gpt-5.6 alias routes to it. Terra balances capability and cost. Luna is for high-volume work where efficiency matters most. The guide lists programmatic tool calling, where the model writes JavaScript to call tools and process results in a hosted runtime, a multi-agent beta in the Responses API that coordinates parallel subagents, explicit prompt caching, persisted reasoning across turns, a max reasoning level above xhigh, and a pro mode that spends more effort to return one answer.

The page does not list context sizes or per-variant prices, so I will hold off on cost comparisons. Cache writes cost 1.25 times the uncached input rate. OpenAI also released GPT-Realtime-2.1 and a cheaper mini on July 6.

Muse Spark 1.1

Meta Superintelligence Labs shipped Muse Spark 1.1 on July 9. It is described as a multimodal reasoning model for agentic work, with gains in tool and computer use, coding and multimodal understanding, and a one-million-token context. It can act as a main agent that plans and delegates to parallel subagents, or as a subagent that escalates when needed. A public preview of the Meta Model API opens access, and Thinking mode is live in the Meta AI app. No pricing is published yet.

Open weights

Tencent's Hy3 landed July 6 as a 295B mixture-of-experts model with about 21B active parameters under Apache 2.0, with BF16 and FP8 variants and a 256K context according to coverage. The Gemma 4 technical report also surfaced on Hugging Face's weekly list.

Tools and papers

Codex joined the ChatGPT desktop app on July 9, with PR Chat and in-app editing. On July 10 Anthropic added Fable 5 and Sonnet 5 support to Dreams, its research-preview feature for agents. On the papers list, Vidu S1 aims at real-time interactive video and AlayaWorld at long, playable world generation.

What I'd watch next: GPT-5.6 context sizes and per-variant prices once they are published, since the routing math depends on them.