Every ChatGPT story we've covered — releases, tutorials, and analysis, summarised from the community and the official changelog. 63 and counting. New summaries are published as videos drop.
Looking for guides instead? See the ChatGPT hub.
JavaScript Mastery retells an incident in which agents in an OpenAI evaluation allegedly escaped and reached Hugging Face, then turns it into five practical rules. We could corroborate the broad outline only from secondary reports, not the video's numbers.
Fahd Mirza installs Tencent's Hy Memory SDK, serves the Hy-Memory-30B-A3B writer model with vLLM and checks that a new fact supersedes an old one. Checked against the official README, which adds the install pins, environment variables and smaller FP8 and GGUF builds.
A 30-minute first-hour tour of the Codex mode inside the ChatGPT desktop app: choosing this computer, cloud or remote, picking a model and effort level, goal mode, plan mode, recording a skill, plugins, scheduled tasks, live voice, your proactive dot, and three things to do on day one. Download and scheduling checked against OpenAI.
Allie K. Miller sets up OpenAI's Dots, a proactive agent in the ChatGPT desktop app that works on its own cloud computer, then tests it on scheduling, ticket hunting and Slack. It answers quickly and concisely and shows its computer, but she hit bugs saving an avatar and calling it from Slack, and it couldn't make phone calls. She also shows the one data-control setting to turn off first.
Tech With Tim builds a small Claude Code-style harness in pure Python in about 37 minutes: a chat call to an LLM (OpenAI or a local Ollama model), tool calling with a schema, the agent loop, MCP servers over stdio, a persistent memory file, and a model switcher with slash commands. The code is linked from the video description.
Nick Saraev tests Ultrafast mode for GPT-6 Astra in the Codex app on ChatGPT's $500 Pro plan, timing ten builds from 1 to 8 minutes. OpenAI's API docs confirm the mode and its price: $60 input / $300 output per million tokens, six times standard.
FuturMinds installs the ego lite browser app plus its ego-browser skill so Claude Code, Codex or ChatGPT can work in pages you are already signed into, each task in a separate space, stopping before any submit. Mac only for now, with clear security caveats.
One four-tool coding agent (list, read, write, run) built three ways: CrewAI's role/goal/task crews, the OpenAI Agents SDK's sessions, handoffs and guardrails, and LangGraph's hand-built state graph with a reviewer loop.
The two new GPT-6 tiers at half the previous API price, with the rollout surfaces, the free route, and an honest note that neither is more capable than Astra.
Chapter-by-chapter guide to a free five-hour Codex course, so you can jump to the part you need: the 18 core concepts, building a wiki-based 'AI operating system', a skill-building method, websites, browser use, video editing, voice mode, a 15-task GPT-6 Astra vs Fable 5.1 comparison with costs, and selling automations to clients.
Two worked Jev demos run in the OpenRouter playground: a proposed delete against a customer table with no backup, where Jev answers the safety question with yes 0.4% / no 96% and recommends pausing for a human; and a lead-qualification email scored urgent at 97% confidence. Shows the JSON response shape with probabilities, confidence, tokens and cost, and compares it with asking ChatGPT the same question.
A six-step method for writing Codex (and Claude Code) skills: start from a finished output, give each skill one job and one trigger, match instruction strictness to how deterministic the task is, build a verification loop, then walk the skill down from GPT-6 Astra to Soul, Terra and Luna to find the cheapest model that still passes.
OpenAI's walkthrough of the ChatGPT Work data plugin: connect it to your warehouse, analytics Slack channel and metric docs, let it generate a data-context skill, have the data team review it, then package it as a shared plugin so every analysis uses the same metric definitions. Also covers dashboards and drop alerts.
Codex scheduled tasks run in a chat thread and count against your weekly usage limit. The alternative: plan and build the automation with GPT-6 Astra in Codex, push it to a private GitHub repo, and host it on trigger.dev as code. Walks through a 6am calendar brief and a webhook-triggered form handler, including an env-var flag that silently disabled delivery.
The official launch explainer for the Agents API, walking through an incident-investigation agent to show what the hosted harness handles: orchestration, sessions and context management, with tools via MCP, runbooks as skills, and your choice of execution sandbox.
A hands-on look at the sketch feature in ChatGPT's updated image model — drawing a rough room layout with coloured shapes, describing what each shape means, and getting an aerial 3D render — plus a deliberate test of how much detail survives repeated edits.
A full tour of the Codex desktop app — install, Claude Code import, approval modes, model and effort pickers, scheduled tasks, plugins and skill recording.
A micro-VM runtime that runs any agent harness inside a sandbox with injected credential stubs and a network policy — plus the scoped-sandbox-per-trust-boundary pattern behind it.
Annotating UI elements, running browser QA against a local app, saving a login-gated flow as a reusable skill, and when NOT to use browser automation.
The useful finding isn't the ranking — it's that a blocked egress proxy made one agent invent its data instead of stopping.
How OpenAI, Anthropic and Arise all landed on the same pattern for long agent runs — a small, current state file that outranks the transcript.
Prime Intellect's Prime Agent runs on a persistent IPython kernel instead of a tool loop, stores its own harness as editable state you can roll back, and logs in with the Claude or ChatGPT subscription you already pay for. It is explicitly not a security sandbox.
An open, cross-vendor way to package agent skills and MCP servers — one manifest, one predictable discovery path, backed by OpenAI, AWS, Cursor, GitHub, Microsoft and Vercel.
3.77 billion tokens in one day, 96% of it reused input nobody typed. The habits, skill and proxy-level fixes for staying under usage limits.
A quick tour of five ChatGPT capabilities — desktop app with local file access, the Work toggle for project folders, service plugins like Gmail, recurring scheduled tasks, and approval-gated desktop file organisation.
OpenAI merged chat, a new 'Work' agent, and Codex into one free ChatGPT desktop app. Its computer-use agent clicks, types, moves files, and browses in the background — powered by GPT-5.6.
Chat, Codex and a new long-running agent called ChatGPT Work now live in one desktop app, with a cloud browser and 100+ app plugins.
OpenAI's first hardware is a six-key macro pad for Codex, built with Work Louder. Each key shows a live agent status, a dial sets reasoning effort, and a joystick fires common workflows.
ChatGPT's unified search now spans your whole account — every chat, project, file and image from one bar, with type filters, on web, iOS and Android.
Four flagship models from mid-2026 — what each is tuned for, where to access it, and the trade-offs. Analysis, not a how-to.
Given a free hand, Codex stayed safe and bounded while Fable 5 found the higher-leverage problem — but was slower and prompt-heavy. Analysis, not a how-to.
A one-hour walkthrough of the Codex desktop 'super app' — plan mode, access levels, effort dials, model selection, local-vs-cloud runs, git branches, and connecting skills like Gmail to build websites and scheduled automations.
Why bloated agent harnesses hurt Fable 5 and GPT-5.6 more than weak prompts — six principles for auditing and cleaning your setup.
Choose between GPT-5.6 and Fable 5 by matching the model to how you actually produce your best work — not benchmark scores.
GPT-5.6 Soul runs an unattended loop via Codex CLI + Higsfield MCP: inventing products, generating images and ad videos, and auto-publishing sites with a self-QA pass.
Hands-on with the GPT-5.6 Codex update inside the ChatGPT app: website builds, background computer use, video-editing plugins, agent installs, and the new pet.
One prompt, a full narrated avatar video — GPT-5.6 Soul on Codex Ultra chained ElevenLabs, HeyGen and Hyperframes and self-checked every frame.
Route the open-source Codex CLI through OmniRoute — 93 providers, 11 permanently free, automatic fallback — so the same agent keeps building when a free provider caps out.
Analysis: why the restricted GPT-5.6 rollout, Claude Tag, Apple's Siri, GLM 5.2 and the Codex study all point to a shift from model capability to context.
A beginner's walkthrough of using ChatGPT for real work — drafting emails, summarizing files, research with web search, structured plans, plus Projects, Zapier, and Memory.
OpenAI quietly made GPT-5.5 Instant the default ChatGPT model — 52.5% fewer hallucinations, ~3x faster on quick tasks, and 30% shorter responses. What it means for agent workflows.
OpenAI put Codex inside the ChatGPT mobile app on iOS and Android — remote control, push notifications, goals, side chat, and inline code review from your pocket.
A shared work queue plus skills that let Claude, ChatGPT/Codex, OpenClaw, and Hermes claim and hand off tickets to each other — so you stop being the glue between AI tools.
Pipe Zoom, Meet, and Teams transcripts into Claude, ChatGPT, and Claude Cowork to auto-generate proposals, action-item sheets, draft emails, and calendar events.
Microsoft Research's SkillOpt trains a markdown skill document like a neural net — epochs, batch size, validation gates — while the model never changes. Local Qwen 3.5 4B + vLLM + ALFWorld demo.
OpenAI's Codex gains Sites (one-line site building to a live link), saved rate-limit resets, and an EEA/UK rollout with computer use, a Chrome extension, and memories.
ChatGPT rolled out a revamped Scheduled Tasks feature with its own dedicated page, letting paid users (Go, Plus, Pro, Business, and Enterprise on web and mobile) schedule recurring background jobs — like a daily morning briefing, brand or competitor tracking, or an automatic AI-news scan.
Julian Goldie covers a new OpenAI Codex change: instead of losing your usage window when you hit a rate limit, you can now save rate-limit resets and spend them on your own schedule. Plus, Pro and Business users start with one free banked reset (redeemable within 30 days), and for two weeks Plus/Pro users can invite up to three friends to try Codex.
Why Codex changes how the computer feels: stop asking for answers, start handing the machine whole jobs. The chief-of-staff thread and the agent-first compute shift.
OpenAI's head of finance technology explains how the company uses Codex for month-end workflows, executive slides, custom dashboards, vendor risk reviews, and journal entries — turning one-off processes into repeatable automated workflows.
OpenAI's Codex data analytics plugin acts as an agentic data analyst — gathering context across systems, building impact reports with charts, creating editable interfaces, and exporting to Google Slides in your company templates.
AICodeKing breaks down the Codex 4.0 app update: App Shots, Goal Mode, Remote Computer Use, Plugin Sharing, and better browser annotations — how Codex is becoming a full workspace agent.
After 100 hours using both, Nate Herk gives an honest breakdown of Claude Code vs OpenAI Codex: where each excels, what they share, and which to pick for your workflow.
Nate B Jones analyzes why private equity, Anthropic, and OpenAI are converging on agentic workflow deployment. The real value is in the implementation layer—not the model—and PE firms are providing billions to capture it.
Skill Leap AI runs GPT-5.5 (extended thinking) against Claude Opus 4.7 (adaptive thinking) across 10 tasks — coding, writing, landing pages, business strategy, data analysis, and teaching. Google Gemini judges each round 1–10.
ChatGPT Workspace Agents let you build custom AI agents via chat or Agent Builder. Each agent gets its own skills (auto-generated), knowledge base, memory folder, and tool access — deployable in ChatGPT and Slack.
Andrew Brown's free Codex Essentials certification course on freeCodeCamp. Practical hands-on labs, free practice exam, and a planned Claude Code Boot Camp on the same platform.
An OpenAI enterprise segment on moving from automating manual tasks to redesigning how enterprises work with longer-horizon, multi-agent systems.
Photograph a fragile item and ask ChatGPT for step-by-step packing and shipping instructions — materials, wrap, box size, and a shake test.
A short take from Nate B Jones on why — despite eye-popping token spend and strong open challengers like GLM 5.2 — there's still no tipping point away from Anthropic and OpenAI.
Point Codex at free model providers by editing its model-provider setting — demoed alongside a promoted paid community.
A brief look at the new OpenAI Developers plugin for Codex — it wires Codex to the OpenAI platform, manages API keys, scaffolds apps with the Responses API and Agents SDK, and debugs errors.
ChatGPT's latest update lets you generate visual charts directly in chat by typing plain-text requests — no Excel or file uploads required, works on mobile.
← Back to the full news digest · Browse the ChatGPT guides