GPT-5.6 Lands, EU Greenlights Mass Scanning, and Postgres Gets a Rust Clone

Good morning. It's Thursday, July 10, and we're covering OpenAI's new model family, a privacy vote that passed despite losing, a Rust Postgres clone that passes every test, and Meta's first paid AI API.

The GPT-5.6 thread topped 900 comments in under 12 hours. The model itself is three models, and the API surface changes how you think about routing.

In today's brief:

• GPT-5.6 Sol, Terra, and Luna ship with million-token context and programmatic tool calling
• EU Parliament lets suspicionless message scanning resume despite majority voting against it
• pgrust passes 100% of Postgres regression tests - but the HN discussion asks harder questions
• Meta launches Muse Spark 1.1 through its first paid API with $1.25/million input tokens

THE BIG ONE

GPT-5.6: Three Models, One API, Million-Token Context

OpenAI released GPT-5.6 as a three-tier family: Sol for complex professional work ($5/$30 per million tokens), Terra for balanced everyday agents ($2.50/$15), and Luna for high-volume extraction and classification ($1/$6). All three share a 1.05 million-token context window and 128K max output - the same inputs work across tiers, so a router can change cost and capability without rewriting prompts.

The Responses API adds two features worth noting. Programmatic tool calling lets the model invoke functions without narrating every step. The multi-agent beta allows coordinated workflows where one model instance can delegate to others. Both reduce token burn on tasks that used to require verbose chain-of-thought scaffolding.

Simon Willison's breakdown noted that Sol outperforms Claude Fable 5 on the "Agents' Last Exam" benchmark by 13.1 points - though he flags the usual concerns about benchmark validity for long-running agentic work. The practical signal is consistency: OpenAI claims fewer tokens and fewer model turns for comparable results, especially when the model can coordinate tools instead of explaining its reasoning.

The routing question is now explicit. Luna handles the repetitive bounded work, Terra gets the multi-step tool-using tasks, Sol takes the expensive high-stakes workflows. Log tier, effort, latency, and tokens so your router can learn from the traces.

Why it matters: This is the first major model release designed around explicit routing between capability tiers. If your agent runs everything through one model, you're leaving money on the table.

Our coverage: GPT-5.6 Sol, Terra, and Luna developer guide

PRIVACY

EU Parliament Lets Mass Message Scanning Resume

The European Parliament voted to allow suspicionless mass scanning of private communications to continue until 2028 - despite a majority voting against. The motion to reject needed an absolute majority of 361 votes; it got 314 against versus 276 in favor, so the regulation stands.

The practical effect: US tech companies can scan private messages on platforms like Instagram, Discord, Snapchat, and Gmail without warrants or prior suspicion. A symbolic exemption was adopted for end-to-end encrypted communications, but service providers already do not scan encrypted messages - the exemption changes nothing for WhatsApp and similar platforms.

The HN thread hit 1,405 points as the most-discussed story of the day. The procedural oddity - a measure passing despite losing the vote - drew particular attention. Patrick Breyer, the civil rights advocate leading opposition, characterized the outcome as "undemocratic" and compared it to "indiscriminately opening everyone's physical mail."

One detail from the discussion: survivors of sexual abuse paradoxically oppose the measure, arguing that surveillance removes the privacy victims need to come forward.

Why it matters: If you build communication tools in Europe, the regulatory environment just got more permissive for scanning - and the precedent for requiring it in the future is now established.

INFRASTRUCTURE

Postgres Gets an AI-Assisted Rust Clone That Passes Every Test

A project called pgrust hit the HN front page claiming a complete PostgreSQL reimplementation in Rust, passing 100% of Postgres's 46,000+ query regression suite. The discussion pulled 671 points and 569 comments, with the debate splitting between those excited about memory safety in the database layer and those skeptical about what "passing tests" actually proves.

The project is not production-ready and does not claim to be. The stated goal is to "make Postgres easier to change from the inside" - a clean-room sandbox for experimenting with architectural ideas that would be too risky to prototype against the real codebase. Planned experiments include multithreaded internals, built-in connection pooling, and no-vacuum storage designs.

The HN response captured a recurring tension in AI-assisted rewrites. One commenter noted that "the things that make software like Postgres reliable are not mostly the tests, but the real world production scars." Another pointed to the AGPL license choice - Postgres's permissive license is part of why it won enterprise adoption; pgrust's copyleft requirement changes that calculus.

The honest answer from the project: you probably should not use it for anything real. It is a research vehicle that happens to have high test coverage.

Why it matters: AI can now produce code-compatible rewrites that pass comprehensive test suites. The question of what that proves - and who maintains parity as the upstream evolves - is still open.

Our coverage: pgrust passes 100% of Postgres regression tests

PLATFORMS

Meta Ships Its First Paid AI Model API

Meta released Muse Spark 1.1 through a new Meta Model API - the company's first paid AI model. Pricing comes in at $1.25 per million input tokens and $4.25 per million output, with $20 in free credits for new developers. The API uses an OpenAI-compatible format, so existing SDK code can switch endpoints with minimal changes.

The model targets personal agentic tasks with a 1 million-token context window and active context management - the model can compact context while preserving critical steps for later work. Zero-shot tool generalization means it works with MCP servers and custom skills without fine-tuning. Meta claims top rankings against GPT-5.5 and Opus 4.8 on agentic evaluations while being "10x cheaper and twice as fast."

Simon Willison created an LLM plugin for CLI access within hours of the announcement. The HN thread hit 375 points, with discussion focusing on whether Meta's entry will pressure API pricing across the market.

Why it matters: Meta entering the paid API market with aggressive pricing and OpenAI-compatible endpoints creates a new budget option for high-volume agentic workloads.

Our coverage: Meta Muse Spark 1.1 developer guide

RESEARCH

GLM 5.2 Matches Human Bookkeeper Accuracy on VAT Returns

A benchmark from Toot Books showed GLM 5.2 preparing quarterly VAT returns for a UK small business with near-human accuracy - off by 7 pence on the net position after processing 59 transactions. Cost: $2.73 versus typical accounting fees of $1,000+. The HN discussion hit 203 points, but the conversation pivoted quickly from "wow this works" to "but who goes to prison when it doesn't?"

The benchmark documented 20 failures across 18 transactions, including misclassified founder capital and 14 instances of mixing up zero-rated versus exempt VAT categories. The authors acknowledge that "the job performed by the humans was broader than what was requested of the model" - humans also had to find invoices and reason through context not in the bank feed.

The liability question dominated the discussion. If you hire an accountant and they commit fraud, your liability is limited by good-faith engagement of a professional. If your LLM decides to commit tax fraud, you are in uncharted legal territory.

Why it matters: The benchmark shows concrete cost savings for structured financial compliance work, but the legal and liability frameworks have not caught up.

Our coverage: GLM 5.2 VAT benchmark analysis

TOOLS WORTH A LOOK

1. Colibri - Run GLM 5.2's 744B MoE model on consumer hardware with 32GB RAM through aggressive memory optimization; only ~40B parameters active per forward pass. OSS, 692 points on HN.

2. FableCut - Zero-dependency browser video editor designed for AI agents to manipulate video through structured commands. OSS, 91 points.

3. Context.dev (YC S26) - API for structured data extraction from any website; aimed at product teams and agents that need web data without scraping infrastructure. Paid, 96 points.

WHAT ELSE IS HAPPENING

• Bun rewrites 535K lines of Zig to Rust in 11 days: Andrew Kelley's thoughts on the AI-assisted rewrite; 712 points, 619 comments. Our analysis.
• Mitchell Hashimoto interview on Ghostty and Zig: Deep dive on terminal architecture and language choice; 262 points. (45 min read)
• Anthropic asks for "hard questions": Public call for difficult AI questions with commitment to show their work; announcement July 9.
• Ben Bernanke joins Anthropic's Long-Term Benefit Trust: Former Fed chairman appointed to governance board; announcement July 9.
• Building a real-time AI tutor for 5-year-olds: Ello's streaming interpreter architecture for sub-second educational responses; 90 points.
• TLS certificates for internal services done right: Practical guide to internal PKI without the pain; 154 points. (10 min read)

FROM THE SITE

What We Published Yesterday

The GPT-5.6 and Muse Spark launches drove a full day of coverage: the GPT-5.6 developer guide with routing recommendations, the Meta Muse Spark 1.1 guide covering pricing and API access, the pgrust analysis on what the AI-assisted rewrite actually proves, the Bun Rust rewrite breakdown, and the GLM 5.2 VAT benchmark analysis on liability questions.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.