What new features ship with Codex-Maxxing?

Three concrete changes: (1) durable threads — long-lived Codex sessions that persist memory, repository awareness, and tool state across days and weeks; (2) parallel agent fan-out — running multiple Codex agents simultaneously in isolated git worktrees, each owning a different feature, bug, test suite, or audit; (3) 62-app plugin architecture covering GitHub, Slack, Linear, Jira, and dozens of others, plus a 'computer use' capability from GPT-5.4 onward that lets Codex operate browsers and desktop apps directly. The Codex desktop app is now on both macOS and Windows.

What models power Codex in June 2026?

Codex runs on GPT-5.5 as the primary model as of April 2026, with GPT-5.4 still available as the native computer-use model for tasks where Codex needs to operate a browser or desktop UI directly. The full GPT-5.5-Cyber model (released June 23, 2026) is also accessible via Codex for security-specific work — vulnerability discovery, patch generation, and the broader Daybreak / Patch the Planet workflows.

Is Codex-Maxxing just a rebrand?

Partly, but the substantive product change is real. The 'Codex-Maxxing' framing names a strategy that was happening anyway — Codex evolving from autocomplete to agent to durable workspace. What's new in the June 22, 2026 white paper is the explicit recognition that Codex's most valuable use case is multi-week, multi-agent, multi-tool project ownership — not single-prompt code completion. The accompanying product changes (durable threads, parallel fan-out, broader plugin coverage) are substantive and not available in most competitor stacks at the same depth.

Quick Answer

What Is Codex-Maxxing? OpenAI's Long-Running Work (Jun 2026)

Q: What is Codex-Maxxing?

Codex-Maxxing is the name OpenAI gave to its June 2026 white paper and product strategy treating Codex not as a one-shot code-completion or single-prompt coding agent, but as a persistent, durable workspace for long-running project work. The core idea is 'durable threads' — Codex sessions that accumulate context, memory, repository state, plugin connections, and conversation history over weeks or months, rather than resetting per prompt. Codex-Maxxing also expands Codex from a developer-only tool to analysts, marketers, designers, investors, and sales — non-developer users are reportedly growing 3x faster than developers inside Codex.

Published: June 24, 2026

What Is Codex-Maxxing? OpenAI’s Long-Running Work Strategy

On June 22, 2026, OpenAI published a white paper titled “Codex-Maxxing for Long-Running Work” that recasts Codex from a code-completion or short-task agent into a persistent, durable workspace for multi-week, multi-tool, multi-agent project work. It’s part product strategy, part terminology, and part response to Claude Code’s growing share of the agentic-coding category. Here’s what changed and how to use it.

Last verified: June 24, 2026.

TL;DR

Aspect	Codex-Maxxing in June 2026
Core idea	Codex as a durable workspace, not a one-shot agent
Key feature	Durable threads — long-lived sessions that persist memory, repo state, tool state
Concurrency	Parallel agent fan-out in isolated git worktrees
Models	GPT-5.5 (primary), GPT-5.4 (computer use), GPT-5.5-Cyber (security)
Plugins	62 third-party app integrations
Target users	Expanded from developers to analysts, marketers, designers, investors, sales
Non-dev growth rate	Roughly 3x faster than developer growth inside Codex

What the white paper actually says

OpenAI’s “Codex-Maxxing for Long-Running Work” PDF (cdn.openai.com/pdf/8a9f00cf-d379-4e20-b06f-dd7ba5196a11/OAI_WhitePaper_Codex-maxxing26.pdf) makes four claims:

The unit of work in Codex has shifted from “completion” to “project.” Most valuable Codex usage today is multi-day, multi-file, multi-tool — not single-prompt.
Memory has to be durable to make that work. A coding agent that forgets the last week of repository decisions is not useful. Durable threads make memory persistent.
Parallelism is the productivity unlock. The right Codex pattern is N agents fanning out across N work-streams (features, bug triage, test writing, security review), each in its own git worktree, recombining at PR time.
The Codex user base is no longer just developers. Analysts, marketers, designers, investors, and sales reps are using Codex for tasks like building dashboards, drafting analyses, generating internal tools, and running data exploration. That cohort is growing roughly 3x faster than the developer cohort.

What’s new in the product

Durable threads

A Codex thread now persists:

Codebase context (which repos, which branches, which open PRs)
Tool / plugin state (which GitHub workflows you’ve connected, which Slack channels, which Linear projects)
Conversation history (what you decided last week, what you tried, what worked)
User preferences (style guide, test conventions, deployment patterns)

This is paired with a “dreaming-based” memory system OpenAI rolled out for ChatGPT earlier in 2026, which automatically updates factual recall and adherence to your preferences without explicit user input.

Parallel agent fan-out

The Codex desktop app (macOS February 2026, Windows shipped after) supports running multiple Codex agents at once, each in its own isolated git worktree:

One agent on a new feature branch
Another triaging open bugs
A third writing tests for a recently-merged module
A fourth running a security audit (Codex Security plugin powered by GPT-5.5-Cyber)

Each agent has its own durable thread. Work converges at PR review time.

62-app plugin architecture

GitHub, Slack, Linear, Jira, Asana, Notion, Figma, Datadog, Sentry, PagerDuty, and dozens more. The Codex Security plugin (Daybreak) connects to GPT-5.5-Cyber for vulnerability discovery and patch generation. The plugin architecture is positioned as the surface where Codex starts touching non-developer workflows — marketing, sales, ops, finance.

Computer use

From GPT-5.4 onward, Codex can operate browsers and desktop UIs directly — clicking, typing, navigating. This is what makes Codex useful for non-developers. A marketer can ask Codex to update fields in a CMS or pull data from a SaaS dashboard, and Codex can do it the way a person would, without an API integration.

Why OpenAI is positioning it this way now

Three reasons converge:

1. Claude Code is winning the agentic-coding narrative

Anthropic’s Claude Code is the obvious competitor. Claude Code shipped Auto Dream (cross-session memory) in May 2026 and Artifacts (live shareable work artifacts) in mid-June. Codex-Maxxing is partly OpenAI’s narrative response — a way to talk about the same durable-workspace direction in OpenAI vocabulary.

2. Non-developer users are the volume

OpenAI has said publicly that non-developer use of Codex is growing roughly 3x the rate of developer use. That changes the product story from “a tool for coders” to “a tool for any knowledge worker who can describe what they want.” Codex-Maxxing names that pivot.

3. The June 22, 2026 Samsung deal

The same week OpenAI published Codex-Maxxing, it announced its largest enterprise ChatGPT deployment ever — Samsung Electronics deploying ChatGPT Enterprise and Codex to all Korean employees plus all Device eXperience (DX) division employees worldwide. Codex-Maxxing is the framing that makes Codex saleable to a non-developer enterprise workforce.

How to actually use Codex-Maxxing today

If you’re a developer

Enable durable threads on the project you’re spending the most time in.
Connect plugins — GitHub, Slack, Linear / Jira, the test runner of your choice.
Try the fan-out pattern — open three Codex agents in parallel git worktrees and give them adjacent work-streams. Triage one issue queue while another writes tests while a third reviews a new branch.
Pilot Codex Security — point the Codex Security plugin (GPT-5.5-Cyber) at a small repo and review the findings. Don’t merge fixes blindly.

If you’re not a developer

Pick one recurring task — a weekly dashboard, a recurring data pull, a routine analysis.
Start a Codex thread for it — let durable memory accumulate over a few iterations.
Use computer-use plugin connections to bridge to whichever SaaS or internal tool the workflow touches.

What this is, and what it isn’t

What it is

A real product strategy with substantive feature backing — durable threads, parallel fan-out, broader plugin coverage, computer-use
A response to Claude Code’s narrative momentum
A framing for OpenAI’s expansion into non-developer knowledge work

What it isn’t

A new model. GPT-5.5 / GPT-5.4 / GPT-5.5-Cyber were already shipping.
A break with prior Codex strategy. Codex was already trending this way.
A guaranteed win over Claude Code. Both stacks are now converging on durable-workspace strategies; the differentiation is real but increasingly narrow.

What to watch from here

Claude Code’s response — does Anthropic match parallel fan-out tooling?
Plugin ecosystem growth — 62 apps today, but the long-tail is what determines reach
Codex computer-use reliability — how often does it correctly do non-developer tasks end-to-end?
Codex Security adoption — how many security teams actually deploy GPT-5.5-Cyber via Codex in production?
OpenAI’s pricing posture — will durable threads remain inside the Pro tier or move to higher tiers as memory storage costs scale?