Evening brief

Thursday, July 9

2 people · 0 companies · 6 topics

TL;DR GPT-5.6 Sol, Terra, and Luna launched publicly today after a US government safety hold. Sol’s Ultra mode spawns cooperative subagents — a meaningful architectural shift for agentic work.

Models & research

  • model-releases OpenAI’s GPT-5.6 Sol/Terra/Luna went public today after the Dept. of Commerce cleared the series; they were gated since late June. Sol ($5/$30 per M tokens) is the flagship tuned for bio/chem/cybersec; Terra ($2.50/$15) is billed as 2x cheaper than GPT-5.5 at comparable quality; Luna ($1/$6) targets speed/cost. Neowin · OpenAI preview

  • agentic-ai Sol’s Ultra mode decomposes tasks and spawns multiple cooperative subagents that communicate mid-run, not just at merge time. Terminal-Bench 2.1: 88.8% standard, 91.9% Ultra. Each Ultra call can cost several times a standard request. apidog · TechTimes

  • model-releases Cerebras will run GPT-5.6 Sol at up to 750 tokens/sec in July — roughly 10x faster than a standard H100 cluster — initially to select customers. The Cerebras contract with OpenAI is reportedly worth $20B+ over multiple years. Value Add VC

  • Polymarket signal: “Which company has best AI model end of July?” currently sitting at 87% Anthropic despite today’s GPT-5.6 launch — a notable crowd read given OpenAI is shipping its strongest series. Polymarket

People

  • simonw released sqlite-utils 4.0 on July 7 — first major version since 3.0 in November 2020. Adds database schema migrations, nested transactions via db.atomic(), and compound foreign key support. He noted Claude Fable wrote much of the 4.0rc2 work for ~$149. Blog · GitHub release

Tools & dev

  • claude-code latest update adds a Dynamic Workflow Size setting in /config (small/medium/large agent count advisory, not a hard cap), workflow.run_id and workflow.name OTel attributes on workflow-spawned agents for full run reconstruction from telemetry, and reverts /review <pr> to a fast single-pass flow (multi-agent review now lives under /code-review <level>). Releasebot · CHANGELOG

  • opencode v1.17.14 ships a code-mode MCP adapter for running confined orchestration scripts against connected MCP tools, fixes paginated MCP tool catalogs losing metadata, and patches GitHub Copilot model routing to honor advertised endpoints. opencode.ai · GitHub

OSS

  • addyosmani/agent-skills (JS) — “Production-grade engineering skills for AI coding agents,” +1,297 stars today, 74k total. GitHub

  • asgeirtj/system_prompts_leaks — extracted system prompts from major AI providers, +1,218 stars today, 54k total. Worth watching for what’s actually inside frontier model system prompts. GitHub

  • bradautomates/claude-video (Python) — lets Claude watch and analyze videos by extracting frames and transcripts, +951 stars today. GitHub

Other

  • The White House is finalizing a voluntary framework with OpenAI, Google, and Anthropic giving the US government up to 30 days to review frontier model releases for national security implications before public launch. Announcement was expected this week (July 7-11). The GPT-5.6 delay was an early instance of this process in practice. FAQ.com.tw · White House EO