GPT-5.6 家族、Muse Spark 1.1 等 AI 代理周报

In today’s issue:

• OpenAI ships the GPT-5.6 family
• Meta releases Muse Spark 1.1
• OpenAI launches ChatGPT Work agent
• xAI releases Grok 4.5 for coding
• Cognition ships SWE-1.7 at 1000 tok/s
• Mistral drops Robostral Navigate
• Harness design sets agent economics
• OpenAI launches GPT-Live voice
• Google open-sources Gemma 4
• Tencent open-sources 295B Hy3
• Google ships Cloud Run sandboxes
• Nous puts Hermes Agent in the cloud
• Microsoft releases Flint for agents
• Ternlight runs embeddings in-browser
• GPT-5.6 proves 50-year math conjecture
• Databricks benchmarks coding agents
• OpenAI audits SWE-Bench Pro
• FrontierFinance benchmarks agent analysts
• Paper turns memory into navigation
• GitLost tricks GitHub’s AI agent
• Anthropic finds a global workspace
• Sakana replays Picbreeder with VLMs

And all the top AI dev news, papers, and tools.

Top Stories

OpenAI Ships the GPT-5.6 Family

OpenAI began rolling out its GPT-5.6 family, Sol, Terra, and Luna, across ChatGPT, Codex, and the API.

• Capability tiers: The number marks the generation while Sol, Terra, and Luna are durable tiers that advance on their own cadence. Sol is the flagship for the hardest tasks, Terra matches GPT-5.5 at lower cost, and Luna is the fastest and cheapest.
• Built for agents: GPT-5.6 is the new default brain behind Codex and ChatGPT Work, tuned for long-horizon tool use and coding.
• Pricing: Sol runs 5 dollars/30 dollars per million input/output tokens, Terra 2.50 dollars/15 dollars, and Luna 1 dollar/6 dollars.
• Rollout: Live now in ChatGPT, Codex, and the API, with the Codex desktop app merging into the ChatGPT app on Windows and Mac.

Blog

Meta Releases Muse Spark 1.1

Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, and opened the Meta Model API to developers for the first time.

• Agent orchestration: Works with native tools, MCP servers, and custom skills, and can act as a main agent that plans and delegates work to parallel subagents.
• Agentic benchmarks: Posts SOTA scores on MCP Atlas (88.1), JobBench (54.7 vs Opus 4.8 at 48.4 and GPT-5.5 at 38.3), and Humanity’s Last Exam with tools (62.1 vs Opus 4.8 at 57.9), plus FinanceBench.
• Long context: Supports a 1M-token context window for long-horizon, multimodal work.
• Open API and pricing: Meta Model API is in public preview at 1.25 dollars/4.25 dollars per million input/output tokens, with 20 dollars in free credits for new accounts.

Blog

Read more