AI News DigestArchive · Every weekday or Monday
Issue · 2026-09-16

Most AI news is noise.Here is the rest.

MODELS

Opus 5.2 is already routed to some accounts

  • First impressions: best model yet at understanding intent, strong on multi-step reasoning (Gegam245074).
  • A second independent tester: better than Opus 5 for visual/3D tasks, still behind Astra (pankajkumar_dev).
  • Rumored to be a real leap from Opus 5 but not reaching Fable 5 level (kimmonismus).
MODELS

GPT-6 Sol: the cheap-and-fast tier is close

  • Users already routed to it via 5.6 Sol call it consistently fast (kimmonismus).
  • It's positioned as the mid-tier: faster and cheaper than Astra, already spotted in the API (pankajkumar_dev).
DIGEST

MCP is winning the tool-integration argument

  • Core claim: models tool-call well enough now that MCP's statelessness and tool deferral beat CLI wrappers for most integrations (trq212).
  • Independent agreement, with one honest caveat: CLIs still win for piping local file data (RhysSullivan).
ORCHESTRATION

The harness, not the model, is where capability comes from

  • A harness does four concrete things: runs the loop, hands out tools, manages memory, checks the work (gregisenberg).
  • A live example: deciding whether something is a skill concern or a harness concern in an actual codebase (mattpocockuk).
  • Evidence it's not just talk: a custom harness on a graph database beat SoTA on a public benchmark in 3 days (hyperparticle).
  • Counterpoint: most of what makes a harness good is old guardrail engineering, not AI-specific tricks (dexhorthy).
DIGEST

Grok's roadmap talk is outrunning what's shipped

  • Elon's own Grok 4.7 benchmark claim already quietly moved from "beats Opus 5.1" to "roughly matches Opus 5.0" (shawnchauhan1).
  • Ground truth: no model card, API ID, price, or benchmark exists for 4.7 yet; 4.8 was announced but unshipped (realrunai).
TOOLS

Two tools worth a look

  • Inspo: a design MCP server for Claude Code, Codex, and OpenCode that searches 800+ sites for UI inspiration, one npx away (nutlope).
  • security-audit-skill: a Cloudflare skill for Cursor/Codex that runs multi-phase security audits and outputs machine-readable findings (LLMpsycho).
DIGEST

One workflow change worth copying

  • /retro now turns fuzzy rules into deterministic checks on its own — proposes lint rules, pre-commit hooks, or CI workflows from a found error (mattpocockuk).
Get tomorrow’s issue