The new GPT-5.6 family: Luna, Terra, Sol
GPT-5.6 reached general availability in three sizes, priced per token from $1/$6 (input/output) for the smallest tier, Luna.
AI Daily Recap
20 articles · 5 categories
The finishable daily brief
What happened in AI — Jul 9, 2026
Thursday, Jul 9, 2026
20 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- OpenAI shipped GPT-5.6 in three tiers (Luna, Terra, Sol) and it's already the default model across Microsoft 365 Copilot.
- Meta shipped Muse Spark 1.1, its first Spark model with an API, claiming gains in agentic tool calling and computer use.
- xAI shipped Grok 4.5, its first Opus-class model since acquiring Cursor.
- OpenAI's new ChatGPT Work agent can act across your apps and files and stay on a task for hours.
- AWS published field guidance on where MCP tool design goes wrong and how to fix it.
- A new "Agents' Last Exam" benchmark tests AI agents on real-world professional workflows.
OpenAI shipped GPT-5.6 in three tiers — Luna, Terra, and Sol — priced from $1/$6 per million tokens for Luna, and Microsoft made it the default model across every 365 Copilot app within hours. Meta answered with Muse Spark 1.1 (its first Spark model with an API, claiming agentic tool-calling and computer-use gains), and xAI shipped Grok 4.5, its first Opus-class model since acquiring Cursor.
Around it, the day's agent-tooling flow kept building the harness: OpenAI's ChatGPT Work agent now runs unattended on a task for hours, AWS published field guidance on MCP tool design, and four Show HN launches (CodeAlmanac, Bytechef, SoulOS, LangDrift) target coding-agent memory, orchestration, and cross-language testing.
Models & Releases: GPT-5.6 Ships, Meta and xAI Answer 4 items
OpenAI's GPT-5.6 launch was the day's biggest story — three priced tiers and instant adoption inside Microsoft 365 Copilot — while Meta and xAI both answered with agent-capable releases of their own.
GPT-5.6 is now the preferred model in Microsoft 365 Copilot
Microsoft made GPT-5.6 the default model across Word, Excel, PowerPoint, Chat, and Cowork within hours of its release.
Introducing Muse Spark 1.1
Meta's first Spark model to ship an API claims significant gains in agentic tool calling and computer use over the original Muse Spark.
[AINews] SpaceXAI launches Grok 4.5, first Opus-class model post Cursor acquisition
xAI shipped Grok 4.5, its first Opus-class model since acquiring Cursor, continuing to ship faster than any other frontier lab.
Building and Reviewing Code with Agents 4 items
Four separate posts converged on handing code review and repo knowledge to an agent, from a build tutorial to a structured local tool to AWS's tool-design playbook.
How to build a GitHub code review agent
A step-by-step build guide for a GitHub code review agent.
Local codereview tool for handing structured review to agent
An open-source local tool that hands structured code review work to an agent.
Show HN: CodeAlmanac – Self-updating wiki for your coding agent (local, Apache)
CodeAlmanac is a self-updating, local wiki that lives in your repo as Markdown and updates itself from your coding agent's conversations.
MCP tool design: Practical approaches and tradeoffs
AWS's field guide to where MCP tool design goes wrong and how to fix it with context-engineering techniques.
Agent Orchestration, State, and Evals 5 items
New releases targeted the harness around the model: an agent that runs unattended for hours, an open orchestration platform, a persona/state layer, cross-language testing, and a fresh benchmark for real-world agent work.
ChatGPT is now a partner for your most ambitious work
ChatGPT Work is a new OpenAI agent that acts across your apps and files and can stay on a project for hours, turning a goal into finished work.
Agents' Last Exam: AI Agent Benchmark for Real-World Professional Workflows
A new benchmark tests AI agents against real-world professional workflows rather than synthetic tasks.
Bytechef open source platform for AI agent orchestration and workflow automation
An open-source platform for orchestrating AI agents and automating workflows.
Show HN: SoulOS–State&persona management for agents without giving up your LLM
SoulOS manages agent state and persona without requiring a switch away from your existing LLM.
Show HN: LangDrift – test AI agents across languages
LangDrift tests whether an AI agent's behavior holds consistent across different languages.
AI Infrastructure: Optimization and Inference 3 items
Infrastructure news ranged from optimization AI going generally available to a database that skips the LLM call entirely, plus a forensic root-cause fix for a ChatGPT crash.
Solve harder problems with AlphaEvolve, now available to everyone on Google Cloud
AlphaEvolve, Google's optimization AI for problems like chip design and delivery-network planning, is now generally available on Google Cloud.
AlloyDB Ships Proxy Models That Replace LLM Calls with Local Inference inside the Database
AlloyDB AI's new proxy models train a lightweight local model from LLM outputs and run queries at database speed, with smart batching claiming up to 2,400x speedups over external LLM calls.
OpenAI Fixes 18-Year-Old GNU libunwind Bug by Treating Crash Debugging Like Epidemiology
OpenAI traced a ChatGPT data-infrastructure crash to two unrelated bugs at once: silent hardware corruption on one Azure host and an 18-year-old race condition in GNU libunwind's setcontext function.
Safety, Governance, and Transparency 4 items
Anthropic and OpenAI both made governance moves today — a public commitment to answer hard questions, a new trust appointment, a biosecurity bug bounty, and a usage-reflection tool for Claude.
Inviting hard questions
Anthropic is asking the public for its hardest questions about AI and committing to publish its work addressing them.
Ben Bernanke appointed to Anthropic's Long-Term Benefit Trust
Former Federal Reserve chair Ben Bernanke joins the trust that oversees Anthropic's long-term mission commitments.
GPT-5.5 Bio Bug Bounty
OpenAI detailed its biosecurity-focused bug bounty program for GPT-5.5.
A new way to reflect on how you use Claude
Anthropic shipped a feature that tracks and visualizes Claude usage so people can check it against their own goals.
You are caught up for this edition