Wynter
AI Product Report Every Wednesday

AI Product Report: GPT-6.2 Twins, Agent Identity Wars, and the MCP Discovery Problem

By Wynter Jones 4 mins read

In short

GPT-6.2 ships as two models instead of one, Product Hunt's leaderboard turns into agent infrastructure, and SnapLogic tackles the MCP discovery problem.

The headline: OpenAI ships two brains instead of one

On November 17, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol simultaneously — a heavy model for research, coding, and multi-step reasoning, and a cheap, fast sibling for everyday work (about a fifth of Astra's per-token cost). Both are live now in ChatGPT (Plus/Pro/Enterprise/Edu, plus Codex) and the API as gpt-6.2-astra and gpt-6.2-sol. Independent evaluator Artificial Analysis ranks Astra #2 on its Intelligence Index, just behind Anthropic's Claude Fable 5.2 — and OpenAI is explicitly selling token efficiency as a feature now, not just raw capability. Splitting the lineup into a premium/efficiency pair instead of one model is the more interesting signal: it's an admission that "smartest possible" and "cheapest usable" are now different products, not different settings.

Agentic AI is eating the operating system layer

Two launches this week point at agents moving from chat widgets into the actual infrastructure they operate on:

  • Stitch AI launched agentic banking tools built into the core banking OS rather than bolted on top — Copilot for plain-language ops, pre-built Agents for onboarding/credit decisioning/collections, and an Agent Harness for banks to build their own, all inside existing permission and audit controls. The pitch is squarely about regulated-industry trust: data never leaves the core system, every agent gets scoped permissions and a full audit trail.
  • SnapLogic launched a Discovery Hub — a searchable, org-wide registry of MCP servers and tools, plus expanded SnapCode agent-building and new Agent Observability for tracing individual agent runs. This is the clearest evidence yet that MCP sprawl inside large orgs has become its own problem: once you've got dozens of MCP servers, the bottleneck shifts from "can I connect a tool" to "can I find out what's already been built."

Product Hunt: the week the category became "infrastructure for agents"

This week's Product Hunt leaderboard (week of Nov 16) reads less like consumer app launches and more like plumbing for a world full of AI agents:

  • Zoom Slides — turns Zoom meeting recordings into ready-to-present slide decks, the week's #1.
  • Boost.space AgentOS — "one operating system for all your AI agents," #2 and climbing.
  • readywhen — a personal assistant that lives in your email inbox around the clock.
  • OpenBot — an open-source, self-hostable alternative to Grok Bot, any stack.
  • OpenFrame — AI agents that run IT operations, with a human approving every fix — a sane middle ground between full autonomy and ticket queues.
  • ChatGPT Wallet — lets ChatGPT agents make purchases without ever seeing your card number.
  • Puffin — identity and inboxes for AI agents, not humans.
  • Render Sandboxes — spins up a dedicated, sub-second compute sandbox for every agent.
  • UI SDK by Vercel — building blocks so agents can assemble their own UI.

Taken together: identity, wallets, sandboxes, observability, and discovery hubs for agents are now a full product category on their own, independent of whatever model is powering them.

On GitHub: the open-source agent stack keeps consolidating

Trending repo roundups this week (findarepo.com, awesome-repositories.com) point to continued momentum in three lanes rather than any single breakout repo: agent orchestration frameworks, RAG pipelines, and coding-agent harnesses/eval tooling. The theme matches the Product Hunt data — open-source infrastructure for running agents reliably is getting more attention than new foundation models themselves right now.

Also worth knowing

  • Samsung confirmed a December 10 launch for Gemini-powered smart glasses (Galaxy + Gentle Monster/Warby Parker frames), with Google committing that captured visuals won't be used for ad targeting or model training, and automatic bystander face redaction.
  • fileAI opened a Tokyo office and launched fileForensics, a tool that screens business documents for AI-generated or manipulated content before they're used to approve a payment or claim — a direct response to rising AI-generated document fraud.

The takeaway

The frontier-model race (GPT-6.2 Astra/Sol vs. Claude Fable 5.2) is still the marquee story, but the bulk of this week's actual product launches were about agent infrastructure: identity, wallets, sandboxes, discovery, and observability for the agents themselves. If last month's story was "which model is smartest," this week's story is "who's going to manage the fleet of agents running on top of it."

Past editions of the AI Product Report

  1. AI Product Report — 11 November
  2. AI Product Report — 4 November
  3. AI Product Report — 28 October
  4. AI Product Report — 21 October
  5. AI Product Report — 14 October
  6. AI Product Report — 8 October
  7. AI Product Report — 30 September
  8. AI Product Report — 23 September
  9. AI Product Report — 16 September
  10. AI Product Report — 9 September
  11. The week AI video stopped needing an editor