AI↗AIOpenWorkshop🤖 AOW

Intelligence, in practice.

Public Preview · In Motion

The workshop is open.
The benches are
still being built.

A practical place for understanding, testing, building with, and evaluating advanced AI.

Explore the Workshop Get Workshop Notes

AOW Signal

Curated AI news that deserves your attention.

AOW Signal tracks important developments across AI. We publish selectively—some days there may be nothing worth interrupting you for.

Some days may have no major signal.

Google’s Gemini 4 artwork: a large white 4 on a blue gradient.
Gemini 4 Argon · Google DeepMind
Primary signal

Gemini 4 Argon

Google enters its next frontier generation

Google introduced Gemini 4 Argon as its new flagship frontier model, designed for complex, long-horizon work across software engineering, knowledge work, multimodal reasoning, and cybersecurity defense.

Its standout technical signal is a 1M-token output limit, described by Google as industry-leading—pointing toward AI systems that can sustain much longer workflows rather than simply answer individual prompts.

AOW take · Major signal

Gemini 4 deserves close attention, but AOW will separate launch claims from verified experience:

Signal now · Test when accessible · Rank after verification.

Go deeper · AOW Signal 0002 →

No major signal today.

AOW Live Field Note 0001

OpenAI DevDay 2026

2026.09.29 · Editorial update 2026.09.30

From Models to an Agent Platform

AOW’s first post-event reading: models matter, but the system around them increasingly determines what people can build and trust.

01 →WHAT WAS ANNOUNCED
02 →WHAT ACTUALLY MATTERS
03 →WHAT AOW WILL TEST
04WHAT REMAINS UNCERTAIN

01 / WHAT WAS ANNOUNCED

Confirmed against official sources · availability snapshot: 2026.09.30. These are product descriptions, not AOW test results.

  • The model: GPT-6.1 Sol. OpenAI describes improved coding, computer use, and professional work with a lower-cost balance of capability. Performance claims come from OpenAI’s evaluations. [2]Available · API and ChatGPT Work/Codex on Plus, Pro, Business, Enterprise, and Edu. Not yet in Chat.
  • The building system: Codex, Codex Harness, and Agents API. Codex cloud work can continue across desktop, web, and mobile, using reusable environments with separate task workspaces. Agents API exposes the managed Codex harness—the execution system around the model—for sessions, tools, context compaction, recovery, and multi-agent work. [4] [3]Codex cloud · paid-plan/workspace availability. Agents API · beta API [10]; computer use in Codex/Work is limited to Pro 500 and Enterprise. [1]
  • Ongoing work and shared context: Dots and ChatGPT Space. Dots is OpenAI’s name for its always-on agent direction. Space brings shared knowledge and collaborative Pages into a workspace for people and AI. [9] [5] [1]Dots · gradual rollout to Pro and Business Premium in eligible markets; Enterprise, Edu, and Healthcare beta requires admin enablement. Space · Pro, Business, Enterprise on desktop/web; mobile reading/sharing available, creation/editing coming later.
  • External capabilities: Plugins. Extensions add sidebar apps, conversation panels, and file viewers/editors. Improved discovery and submission expand how applications reach people inside ChatGPT. [6] [1]Availability varies by surface · Free/Go web extensions are coming soon; composer mentions are desktop-only.
  • Decisions and privacy infrastructure. Decisions API constrains answers to predefined choices for classification, routing, or an agent’s next action. OpenAI Private Intelligence includes Private Safety Processing for automated safety review and a forthcoming Private Inference preview. [1] [7]Decisions API · limited preview, wider release planned. Private Safety Processing · approved Zero Data Retention organizations and required setup. Private Inference · preview planned for fall, not a generally available capability in this snapshot.
  • Distribution: OpenAI Marketplace. Eligible enterprise customers can apply part of an existing OpenAI commitment to approved partner products. Eligibility depends on agreements and program terms; customers contract with and pay the partner. This is an enterprise purchasing program, not a universal consumer app store. [8]Enterprise-focused · eligibility and approval required. AOW’s future curated external ecosystem has a different purpose.

02 / WHAT ACTUALLY MATTERS

The system around the model becomes part of the product.

AOW interpretation

The most consequential story was the connection between these layers. A capable model needs tools and a working environment; an agent needs context and continuity; useful applications need a way to reach people. The question expands from “Which model is best?” to “Which complete system helps us do useful, verifiable work?”

ModelsHarness / toolsAgentsWorkspace / contextPlugins / external capabilitiesDistribution

Read through MODELS · BUILDING · DISTRIBUTION, DevDay suggests AI products are increasingly becoming ecosystems. OpenAI presented substantial parts of that system; it has not demonstrated that the transition is complete, equally available, or reliable in every setting.

MODELS

GPT-6.1 Sol

BUILDING

Codex · Agents API · Harness
Space · Plugins

DISTRIBUTION

Plugins / app ecosystem
Marketplace (enterprise)

OpenAI DevDay 2026 · conceptual editorial map, not an exhaustive product list or a claim of equal availability.

03 / WHAT AOW WILL TEST

Future evaluation priorities · these tests have not been completed.

  1. Agents API / Codex Harness.

    Study the harness, environment, memory and context, MCP/tools/skills, compaction, and multi-agent behavior. Through the Glassbox lens: can important work remain inspectable, reviewable, source-aware, and human-directed?

  2. Codex mobile / cloud workflow.

    Test whether a real building task can move usefully between desktop, tablet/mobile, and cloud execution—with clear access, review, and handoff.

  3. ChatGPT Space.

    Evaluate continuity, shared context, and collaboration as a persistent human+AI workspace, beyond storing files.

  4. Plugins.

    Evaluate discovery, permissions, transparency, and usefulness. Does an integration improve the workflow enough to justify its access and complexity?

  5. Codex Game Studio / playful Coding DIY.

    Reserve a future experiment: can someone who is not primarily a gamer turn a personal game idea into something genuinely playable? Usefulness, curiosity, experimentation, and FUN all belong at this bench.

    Under verification · Game Studio’s official product scope and general availability are not established in this edition. Exploration remains a future plan.

  6. Dots.

    Study persistent and proactive work: permissions, memory, long-running responsibility, human handoff, notifications, failure recovery, and privacy.

04 / WHAT REMAINS UNCERTAIN

A compelling direction still needs ordinary-world evidence.

  • Reliability and recovery. How do persistent agents behave over long tasks, missed events, interruptions, and partial failures? Can people reliably stop, redirect, and recover the work?
  • Permissions and privacy. Are access boundaries understandable and enforceable across tools, memory, workspaces, and background activity?
  • Availability and economics. Which plans, regions, and clients actually receive each capability—and what will sustained model, tool, and cloud usage cost?
  • Governance and auditability. How will Marketplace eligibility, plugin quality control, and decision-system records stand up to independent scrutiny?
  • Demonstrations and autonomy. Which keynote demonstrations show mature products, and which show early direction? How much human review remains necessary as autonomy increases?

AOW will report what it tested, the conditions, and the failures alongside the results. Announcements are evidence of a direction; they are not proof of dependable outcomes.

Official sources

[1] DevDay recap[2] GPT-6.1 Sol[3] Agents API / Harness[4] Codex Cloud[5] Dots[6] Plugin extensions[7] Private Safety Processing[8] Marketplace scope[9] Space[10] Agents API beta setup

Reviewed 2026.09.30 · official product descriptions are separated from AOW interpretation and future tests. Codex Game Studio’s scope and the two exact keynote phrases remain under verification.

First post-event editorial version · updates are human-reviewed; this page does not update automatically.

AOW Signal 0002

Gemini 4 Argon

Google enters its next frontier generation

Major signal · Under evaluation. AOW’s practical testing is still to come.

01 →WHAT WAS ANNOUNCED
02 →WHAT ACTUALLY MATTERS
03 →WHAT AOW WILL TEST
04WHAT REMAINS UNCERTAIN

01 / WHAT WAS ANNOUNCED

Google announced Gemini 4 Argon on September 30 as its new frontier flagship. Google positions it for complex, long-horizon workflows: software engineering, enterprise knowledge work, multimodal reasoning, and cybersecurity defense. These capability and performance descriptions are Google’s claims, not AOW-tested results. [1] [2]

The standout technical announcement: a 1M-token output limit. Google describes this as industry-leading, increased from 64K. This is an output limit, not simply an input context window. More headroom could support sustained reasoning and longer work; the limit alone does not establish useful, coherent results at that length. [1]

Availability · Phased access. Google reports an initial rollout to trusted cyber defenders through Fairwind. Broader release is planned, starting with paid API customers and Google AI Ultra subscribers. General availability is not established by this announcement. [1]

02 / WHAT ACTUALLY MATTERS

From answering a prompt to carrying a goal through.

AOW interpretation: Argon is interesting beyond benchmark scores. Its long-horizon positioning and output headroom reflect a wider movement from Prompt → Answer toward work that continues through several stages:

GoalLong-running workToolsVerificationResult

For ordinary people, the practical question is whether an AI system can keep the brief intact, use the right tools, recover from mistakes, and deliver something that survives checking. Longer work makes reliability, permissions, and human review more consequential.

This connects to AOW’s Glassbox workflow: ASK → PLAN → WORK → VERIFY → REVIEW → RESULT. We want important work to remain inspectable: sources, actions, checks, uncertainty, and human decisions.

Google describes internal monitoring of Argon’s reasoning and actions to detect misalignment. [2] Internal model/agent monitoring and user-visible inspectability with provenance are related directions, but different things. Google’s monitoring is not equivalent to AOW Glassbox, nor a promise that users can inspect the complete process.

03 / WHAT AOW WILL TEST

Future evaluation priorities · when meaningful access becomes available. These tests have not been completed.

  • Long-horizon reliability: maintaining a goal over many steps, noticing drift, and recovering from failure.
  • Coding / software engineering: usable changes, reproducible builds, meaningful tests, and reviewable results.
  • Multimodal and knowledge work: understanding images, video, and documents; checking sources and factual accuracy.
  • Tool use / agentic workflows: where available, inspect permissions, actions, handoffs, and evidence behind the result.
  • Very long outputs: consistency, coherence, contradictions, and whether the extra length helps a real task.
  • Practical experience: total cost, latency, usability, failure recovery, and how much of the work a person can inspect.

04 / WHAT REMAINS UNCERTAIN

Launch benchmarks and vendor demonstrations are insufficient for an AOW ranking decision. Google’s methodology combines its own evaluations with provider-reported numbers and public leaderboard results; thinking settings, harnesses, and task conditions matter. [3]

  • Broader availability and the capabilities accessible in each product.
  • Practical reliability outside curated benchmarks, including recovery when a task goes wrong.
  • Coherence over very long outputs, and whether the 1M-token limit helps genuine workflows.
  • Real-world cost and latency across a complete task.
  • Comparable, repeatable performance against GPT-6 Astra / Sol, Claude, and other frontier systems.

Major industry signal · vendor benchmark claims · AOW-tested practical performance · final editorial ranking are distinct stages. This launch does not change AOW’s existing model framework or ranking decisions. Quality Talks!

Signal now · Test when accessible · Rank after verification.

Official sources

[1] Google · launch & rollout[2] DeepMind · capabilities & safeguards[3] Evaluation methodology

Source check · 2026.10.01. Google’s launch claims are attributed above; AOW interpretation and future tests are labeled separately. No independent AOW performance verdict yet. Signal artwork: Google DeepMind.

01 / THE OPEN WORKSHOP

A little orientation.
A lot of possibility.

AIOpenWorkshop helps ordinary people understand what advanced AI can actually do—and put that understanding to use. Bring a question, a project, or simply your curiosity.

More than an AI-news feed or a conventional training center, AOW connects useful guidance with hands-on experiments and honest evaluation. We look for meaningful progress, and make room for what is still uncertain.

GuideCuratorWorkshopVerifierField Map
02 / THE MAKER BENCH

Coding DIY in the AI Era

Building · permanent area
From Needs to Fits.

Make the little software
you wish existed.

A study helper, a photo workflow, a research tool, or a game for one. AI can make highly personal software more accessible to students, teachers, photographers, historians, researchers, small-business owners, and hobbyists.

Start with a real need. Experiment, learn, and check what you build. And leave room for fun—making things is part of the reward.

Post-DevDay exploration: a personal game idea, made playable. The Field Note reserves a Codex Game Studio experiment; its official scope remains under verification.

Small needs worth exploring

My Screenshot

A familiar action, easier to reach. What would a capture tool built around your routine look like?

Project seed · not a released app

When a useful button moves

A reverse-translation control changes place. Could a personal tool preserve the workflow you rely on?

Design question · experiment to come

Something just for you

A tiny game, a reading companion, an unusual glossary. It can be useful, playful, or both.

Open invitation · start with your own need
03 / THE TOOL SHELF

AI models & tools

A changing field map. Choose by task and evidence, with room for independent labs and open-source communities.

Base models

Start with a task. Then choose a model.

Compare accuracy, clarity, effort, and the work you can verify.

Building · evaluation notes to follow
Agents

From instructions to action.

Inspect permissions, handoffs, failures, and the final result.

Building · evaluation notes to follow
Coding

Small tools. Real needs.

Try a bounded brief and check the software yourself.

Building · evaluation notes to follow
Image

Write a visual brief.

Specify subject, composition, light, and intended use.

Building · evaluation notes to follow
Video

Think in individual shots.

Check continuity, camera motion, and unwanted elements.

Building · evaluation notes to follow
Voice

Listen beyond the demo.

Consider intelligibility, language, latency, and consent.

Building · evaluation notes to follow
Research

Follow the evidence.

Check citations, source quality, and missing context.

Building · evaluation notes to follow
Creation

Move from idea to deliverable.

Review usefulness, editability, and the finished work.

Building · evaluation notes to follow

Start with a task. Then choose a model. Give two tools the same brief. Compare accuracy, clarity, time, and how much editing you need.

Framework updated
2026.09.28
Try a workshop exercise

Turn notes into a useful result

“Help me turn these notes into a practical project. Identify missing information, make a short plan, then create the first usable deliverable. Explain how I can check the result.”

Check: Open the deliverable, try its main function, and verify factual claims.

Review a proposal

“Review this proposal for a reader who is new to the subject. Separate evidence from assumptions, identify the three biggest gaps, and produce a clearer revision.”

Check: Compare the revision with the original so important meaning is preserved.

Write a visual brief

“Create a square editorial image of a ceramic coffee cup on a dark table, lit by soft window light. Keep the upper third clear for a headline. No text or logos.”

Check: Shapes, unwanted text, composition. Refine one decision at a time.

Plan one video shot

“One continuous shot inside a quiet café at sunrise. The camera slowly moves toward a cup as steam rises. Warm window light. No cuts, no text, no dialogue.”

Check: Object continuity, camera motion, and unintended elements.

From the earlier workshop: a website brief

“Make a website for my university study group. Include this week's reading, a glossary, and three practice questions with expandable answers. Use plain English and make it comfortable to read on a phone.”

Brief → Refine → Review. Replace example content with your own material, try every link, read every claim, and confirm who can access it. Never assume automatic updates or shared data unless those features were built. AOW itself is built with Sites.

Editorial continuity & review status

The practical exercises above are retained from the September 13 workshop. They are suggested starting points, not measured model comparisons.

The earlier edition’s named model selections and time-sensitive release watch are awaiting a fresh source check. They are not presented here as current rankings or verified availability. Future entries will show their source, review date, evidence status, and evaluation limits.

04 / AGENTS & GLASSBOXIn Motion

See how the work happens.

Important AI work should be inspectable where possible: the brief, the plan, the actions, the checks, and the human decisions that shape the result.

ASKPLANWORKVERIFYREVIEWRESULT

Glassbox is AOW’s developing approach to transparent execution. Evidence, uncertainty, and failures belong in view.

Post-DevDay priority: study Agents API / Codex Harness through the Glassbox lens. New infrastructure creates opportunities for inspection; it is not the same as AOW’s review method.

Explore the Glassbox public demo ↗

Candidate awaiting review · deeper guidance is coming.

05 / AI HARDWAREBuilding

A new field of interfaces.

A place to observe and document AI-oriented hardware, emerging computing categories, and how people interact with them.

ComputersPhonesWearablesAudioRoboticsDedicated AI DevicesInterfaces

Category framework only. Product records and verified specifications are still to come. This area documents the field; it is not a shopping ranking.

06 / HOW THE WORKSHOP WORKS

Keep learning. Keep checking.

AI changes. Our notes should change with it. Each update should make clear what we know, how we know it, and what needs another look.

01 →WATCH

Notice meaningful changes.

02 →UNDERSTAND

Find context and sources.

03 →TEST

Try a bounded task.

04 →VERIFY

Check the actual result.

05 →SHARE

Show evidence and limits.

06UPDATE

Revisit as things change.

Confirmed facts

Supported by a traceable source, with scope and date.

Reported information

Attributed to its source; not independently established.

Experiments

What we tried, under which conditions, and what happened.

Interpretation

Our reading of the evidence, clearly labeled as judgment.

Uncertainty

Expected, rumored, unconfirmed, or under evaluation—said plainly.

A LETTER FROM THE WORKSHOP

Get Workshop Notes

Important launches, meaningful AI changes, and new AOW experiments. No noise.

07 / CURRENT WORKSHOP STATUS

An open door. Work in progress.

Status reviewed · 2026.09.30

OPEN

AOW purpose, principles, starter exercises, and the public Founding Preview.

IN MOTION

AOW Signal, Live Field Note 0001, Glassbox, and post-DevDay evaluation.

BUILDING

Coding DIY, tool evaluation, AI Hardware, and Workshop Notes signup.

COMING NEXT

Agents / Harness, Codex workflow, Space, Plugins, and Game Studio exploration.

08 / OUR PRINCIPLES

What stays constant.

  • Practical over performative
  • Evidence over hype
  • Transparent about uncertainty
  • Human review where it matters
  • Vendor-neutral
  • Useful to ordinary people
  • Open exploration
  • Quality Talks!

OpenAI DevDay is our founding-preview occasion. Our field of view includes OpenAI, Anthropic, Google, xAI, Meta, independent labs, open-source communities, hardware companies, and researchers—according to relevance and quality.

Independent and not affiliated with model providers. No hidden ads or soft advertising. Future sponsorships or affiliate relationships will be explicitly disclosed. AI-assisted, human-directed; review remains part of the work.