News archive
Weekday briefing October 2, 2026 · 8:00 AM ET

GPT-6 choices, Gemini skills, and AI-built sites

OpenAI's GPT-6 guide offers a practical way to choose a model: test real tasks and measure the cost of a successful result. Google is replacing Gems with reusable Gemini skills, while new site-building tools from ChatGPT and Shopify are worth testing with their publishing and feature limits in mind.

AI-assisted and human-reviewed Last updated October 2, 2026
01

The Big Story

OpenAI News

A practical guide to building with GPT-6

OpenAI published a guide to choosing among GPT-6 Astra (hard reasoning), GPT-6.1 Sol (complex coding and research), and GPT-6 Luna (repeated, focused tasks). It recommends evaluating task success, latency, and cost per successful task before deployment, while using prompt caching and context compaction to manage expense. The guide also covers decision boundaries for agents, asynchronous tools, mid-run steering, and Sol's multi-agent support (currently in beta).

Takeaway

For a recurring AI task, test the cheapest suitable model on real examples, measure cost per successful result, and specify which decisions require human approval before putting it into production.

Read the original at OpenAI News
02

For Everyone

Google Gemini Blog

Let skills in Gemini tackle your most repetitive tasks

Google is rolling out reusable “skills” in Gemini chat globally, with availability for Workspace business, enterprise, nonprofit and education customers expected in the coming weeks. Users can save instructions, invoke them with a slash command, attach reference files and combine skills for tasks such as drafting in a brand voice. Skills will replace Gems, which Google says it will migrate automatically when support ends (March 2027 for Workspace business, enterprise and nonprofit accounts; November for personal accounts).

Takeaway

If your team relies on Gems, inventory the important ones and test their migrated skills before the March 2027 cutoff; new repeatable writing or presentation prompts are worth trying once Workspace access arrives.

Read the original at Google Gemini Blog

Google Gemini Blog

Guided Vision in Gemini Live: built for accessibility

Google launched Guided Vision in Gemini Live for compatible Android phones running Android 9 or later where Gemini Live is supported. By sharing a camera, users can ask spoken questions about their surroundings, read small labels, find objects, and get verbal prompts to reposition the camera for a better view. Google developed it with accessibility partner Aira and more than 1,000 testers, but says it can make mistakes and must not be used for navigation or obstacle detection.

Takeaway

Android users who need help reading labels or locating objects can try the setting in Gemini Live; don't treat its descriptions as a substitute for a mobility aid or safety check.

Read the original at Google Gemini Blog
03

For Builders

Hacker News

Clef: Open-weight decision models, and new RL fine-tuning platform

Cloudflare released Clef and Clef-flash, decision models that classify text or images into structured answers such as whether a support request is urgent and which team should handle it. Both are available through Workers AI and as Apache 2.0 open-weight models; Cloudflare's own tests report lower latency than the competing Jev model, though results vary by benchmark. Hands-on customer fine-tuning is being offered first, while a self-service training platform is still planned.

Takeaway

If your business already automates support-ticket routing, a developer could test Clef on labeled tickets and keep uncertain cases with a person. There's no need to adopt it just to use an AI chatbot.

Read the original at Hacker News

GitHub Trending

codegraph: Pre-indexed code knowledge graph, auto syncs on code changes, for Claude Code, Codex, Gemini, Cursor, OpenCode, AntiGravity, Kiro, CoPilot, and Hermes Agent — fewer tokens, fewer tool calls, 100% local

CodeGraph is an MIT-licensed local index that maps a codebase's symbols and dependencies so AI coding agents can retrieve relevant source and see what a change might affect. It supports several popular coding assistants and updates the index as files change. In the project's own seven-repository question-answering benchmark, agents used 88% fewer tool calls and 62% fewer tokens, but their sessions also retained about 80% more retrieved context, which can crowd long conversations.

Takeaway

A business with an active software team could test it on one repository and compare answer quality and actual usage cost before adding another agent integration. It has little immediate use outside software development.

Read the original at GitHub Trending

TechCrunch AI

Amazon releases its own Jev clone as decision models flood the web

AWS released Strands Decider 2B, an open-source model that chooses among predefined options and reports confidence instead of generating text. Based on Qwen3.5-2B, it can run locally and is intended for faster, cheaper decisions within automated workflows. Amazon says the approach can improve reliability, but balancing speed with accurate, well-calibrated choices remains an open challenge.

Takeaway

If an AI workflow repeatedly selects from a fixed set of next steps, benchmark a small decision model against your current LLM on accuracy, latency, and cost before switching.

Read the original at TechCrunch AI
04

For Business

Hacker News

Sites in ChatGPT

OpenAI has put Sites in ChatGPT into public beta for eligible Plus, Pro, Business, Enterprise and Edu subscribers. Users can describe a website or app, refine it through conversation, then publish it on a ChatGPT-hosted address or custom domain; options include team editing, visitor sign-in, saved files and, in eligible workspaces, read-only access to visitors' connected tools with their approval. Access, limits and workspace controls vary by plan and region.

Takeaway

A small team could prototype a customer-facing page or internal dashboard, but check plan limits and test sharing permissions and stored customer data before publishing.

Read the original at Hacker News

TechCrunch AI

Shopify debuts Canvas, a way to build online stores by chatting with AI

Shopify introduced Canvas, a desktop-only tool that lets merchants build and edit stores by chatting with its Sidekick AI agent. Changes render from the store's actual code in real time, so merchants can test interactions and view different screen sizes. The initial release doesn't support third-party themes, app blocks or extensions, markets, translations, rollouts, or theme updates.

Takeaway

Shopify merchants with straightforward stores can try Canvas for layout changes, but should check its unsupported features before relying on it for an existing storefront.

Read the original at TechCrunch AI