Verdict: OpenAI's Codex update shipped on July 29, 2026 (CLI 0.146.0, plus the Codex desktop app joining ChatGPT on July 9) is the first release where the headline isn't a smarter model — it's a smarter workbench. Named sessions, pinned threads, forkable history, side conversations you can switch between without closing, and inline pull-request reviews turn Codex from a one-shot code generator into something closer to a managed junior engineer: a tool that holds onto your project across days, lets you review its diffs like a boss, and keeps the main task running while you ask a side question. For anyone using AI agents for real multi-day work — building a site, running a content system, maintaining an SDK — this is the update that makes Codex genuinely usable as a long-running partner rather than a chat you have to re-explain every morning.
Last verified: 2026-07-29
- CLI 0.146.0 shipped July 29, 2026: named sessions, pinned threads, side conversations, thread forking with paginated history, Agent Plugins marketplace support (Amazon Bedrock, Claude Code).
- Codex joined the ChatGPT desktop app on July 9, 2026: inline Markdown/code editing, PR review sidebar, multi-repo projects, Computer Use with GPT-5.6.
- Codex pricing: Free ($0, limited), Plus ($20/mo, 15–90 GPT-5.6 Sol messages / 5h), Pro 5x ($100/mo), Pro 20x ($200/mo), Business ($20/user/mo).
- Model lineup: GPT-5.6 (Sol/Terra/Luna variants) is generally available as of July 9, 2026.
- Pricing and limits change often — last checked 2026-07-29.
What's actually new in the OpenAI Codex July 2026 update?
The headline change is a workflow change, not a model change. Codex's releases across July 2026 — CLI 0.146.0 on July 29 and the Codex desktop-app merger on July 9 — concentrate on four practical problems that anyone running real multi-day projects hits: losing context, juggling multiple threads, reviewing work without switching tools, and letting a side question interrupt the main job.
1. Long-running context that survives across days
Before this cycle, a long Codex thread would drift: it "forgot" earlier decisions, or you'd have to re-paste the spec on every follow-up. The fix in 0.146.0 isn't a model upgrade — it's session persistence. You can name a new session with /new or /clear, pin the important threads so they stay reachable, and switch between side conversations without closing them (OpenAI Codex CLI 0.146.0 release notes, 2026-07-29). Threads also gained forking with paginated history — you can branch a conversation, explore a different fix, and keep the original thread intact, or fork into a temporary thread that doesn't clutter your sidebar. Submitted messages, final responses, approval settings, and even failed-turn errors are now preserved across interruptions, replays, imports, and forks.
What this means in practice: your link-building plan, your content brief, your refactor — it stays where you left it. You come back the next day, pin the thread, and Codex still knows what "the cockpit component" refers to. For anyone designing AI agent workflows that hold together for more than one turn, this is the persistence layer that earlier Codex versions lacked.
2. A sidebar that's a real project dashboard
The Codex sidebar now organizes multiple projects, tasks, reviews, and conversations so each has its own place. The July 9 desktop-app release added Multi-repository projects and a workspace file-tab side panel with drag-and-drop tab reordering (OpenAI Codex changelog, July 2026). Translation: instead of one chat that has to hold "the keyword research thing" and "the backlinks thing" and "the content briefs thing," each thread is its own object. You jump between them in the sidebar without scrolling back through history to find the one that had the right context.
3. Inline reviews where you approve, reject, or send back
This is the update that turns Codex from a chat box into a managed employee. You can now:
- Edit Markdown and code directly in the app, use inline annotations, and ask Codex to revise selected content.
- Review GitHub pull requests in a sidebar — reviewer feedback sits alongside the diff, and you can send inline review feedback, inspect proposed patches, and edit, accept, or reject them without leaving the app.
- Use PR Chat to ask Codex about changes in context (OpenAI "Codex joins the ChatGPT desktop app" announcement, 2026-07-09).
This is the "manager" pattern: Codex drafts, you approve or send back. For builders who have already been treating AI agents like managed team members rather than chatbots, this is Codex finally catching up to the workflow you wanted.
4. Side conversations that don't interrupt the main task
Before 0.146.0, asking Codex a quick tactical question — "why this keyword?", "what about this title tag?" — meant either interrupting the long-running task or opening a fresh context. The new side conversations let you open a side chat, brainstorm, and keep the main task running in the background. You can switch between side conversations without closing them (Codex CLI 0.146.0 release notes).
It's a small change with a big convenience payoff. You stop losing your place in the main thread every time you have a tangent.
5. Agent Plugins, marketplaces, and the desktop merger
Two structural changes matter beyond the UI tweaks:
- Agent Plugins are now first-class: manifests, workspace plugin publishing, and plugin marketplaces for Amazon Bedrock and Claude Code (Codex CLI 0.146.0). This means a team can package an internal workflow — a set of MCP servers, custom skills, app integrations — and share it across members, rather than everyone rebuilding the same Codex configuration. For any team treating the agent layer as infrastructure rather than a chat app, plugins are the packaging primitive that makes Codex shareable.
- Codex joined the ChatGPT desktop app on macOS and Windows on July 9, 2026. Existing Codex app users keep their projects, settings, and workflows; they can make Codex the default view. The merger is significant because it puts Codex's code-aware loop inside the same app you already use for general ChatGPT work, rather than splitting your attention across two tools (OpenAI Codex changelog, July 2026).
How is Codex different from ChatGPT?
Codex is an agent. ChatGPT is a conversation. Both run on the same OpenAI account and the same frontier models (GPT-5.6 Sol, Terra, Luna as of July 2026), but they loop differently:
| ChatGPT | Codex | |
|---|---|---|
| Loop | Prompt in, response out — you copy code | Task in, PR/diff out — it edits your files and runs commands |
| Surfaces | Chat box, web, mobile | CLI, IDE extension (VS Code, Cursor, Windsurf), web, desktop app, iOS |
| Memory | One conversation at a time | Named sessions, pinned threads, forkable history |
| Work on your repo | No — you paste snippets | Yes — reads repo, edits files, runs tests, opens PRs |
| Async | No | Yes — cloud tasks run 1–30 min; queue multiple in parallel |
| Pricing | Included in every ChatGPT plan | Included in every ChatGPT plan (Free and up); usage metered by plan |
(Source: ChatGPT vs Codex comparison, morphllm.com, June 2026; OpenAI Codex product page)
The leverage difference is access. ChatGPT can talk about a code snippet you paste. Codex reads your actual repository, makes edits across files, runs your tests, and iterates on the result. The July 2026 update sharpens the agent side: persistent threads and review workflows make the "delegated engineer" loop real, not aspirational.
What does Codex cost in 2026?
Codex is included in every ChatGPT plan — you do not buy it separately. The limits are what differ (OpenAI Codex pricing, morphllm.com, July 2026; eesel AI pricing guide, 2026):
| Plan | Price | Codex access | Limit flavor |
|---|---|---|---|
| Free | $0 | Web, CLI (lowest limits) | Promo-era boosted limits have ended |
| Go | $8/mo | Web, CLI | Higher than Free; no cloud task delegation |
| Plus | $20/mo | Web, CLI, IDE, iOS | 15–90 GPT-5.6 Sol messages / 5h window; cloud tasks, code reviews |
| Pro 5x | $100/mo | Everything in Plus | 5× the Plus 5h window limits (added April 9, 2026) |
| Pro 20x | $200/mo | Everything in Plus | 20× the Plus 5h window limits |
| Business | $20/user/mo (annual) | Everything in Pro | Cloud features, GitHub/Slack/Linear integrations, admin controls |
| Enterprise / Edu | Custom | Everything in Business | Custom contracts, SSO, audit logs, data residency |
Two important mechanics:
- 5-hour rolling window. All Codex limits are measured per 5-hour rolling window (per model), not per month. A weekly cap also resets 7 days after your first message of the week. When both windows are exhausted, you wait for reset or buy credit top-ups (available on Plus and Pro).
- Token-based credits, not messages. Since April 2, 2026, Codex switched from per-message billing to token-credit billing. A typical small bug fix burns 10 credits ($0.40); a multi-file refactor burns 60 credits ($2.40). Credits map to API token prices at roughly 4 cents each (OpenAI Codex pricing).
For most people deciding whether to upgrade from Free: the $20/month Plus tier is the threshold where cloud task delegation, IDE extension, and code reviews unlock. Below that, Codex is essentially a CLI you can experiment with.
When should you use Codex vs ChatGPT?
Use ChatGPT for: brainstorming strategy, drafting copy, explaining a concept, sketching a plan, writing a content brief, generating a headline, reasoning about a decision. Anything where you want a thinking partner and the output is text you'll act on yourself.
Use Codex for: building a landing page from a spec, refactoring across files, fixing a failing test, migrating a library, generating an internal tool, running a multi-file task while you sleep, opening a PR. Anything where the output is work done in your repository.
The July 2026 update widens the Codex side. Persistent threads mean you can keep a content-system project running for weeks — not days. Reviews mean you stay in the approver seat. Side conversations mean you don't have to choose between "ask the tangent" and "keep the main job moving." If you've been routing Claude Opus 5 or other frontier agents into your stack by effort level and fast mode, Codex's 0.146.0 release is the OpenAI equivalent of effort-routing: it spends the same tokens, but on a better-managed work surface.
How do you actually use Codex for real multi-day work?
The pattern that works — distilled from the release notes and what we've seen with adjacent agentic setups — is the batch-and-review loop:
- Set up once. Install Codex:
npm i -g @openai/codex(CLI), or open the Codex view in the ChatGPT desktop app. Sign in with your ChatGPT account. Codex reads your repo and runs in a sandbox; local edits happen via the CLI or IDE extension, cloud tasks run in OpenAI's sandboxed container (OpenAI Codex CLI docs). - Batch the morning. Queue 3–5 well-scoped tasks before you start manual work: "fix the TypeScript error in the onboarding validation," "add error boundaries to the admin dashboard," "migrate the legacy auth middleware to the new session manager." Each becomes its own thread. Pin the important ones so they survive the 5h window.
- Use a side conversation for tangents. A new title-tag idea or a question about why Codex picked a keyword doesn't interrupt the main task. Open a side chat; ask; close it; the main thread keeps running.
- Review like a boss. When a task completes, open the review sidebar. Accept the diff, send it back with a comment, or edit inline. The PR-chat feature lets you ask Codex "why this approach?" in context, without leaving the app.
- Fork when the path splits. If a task has two plausible fixes, fork the thread, try both, keep the winner. Temporary forks don't clutter your sidebar.
This is the same delegated-work pattern that works with any serious agentic setup — the difference is Codex now holds the state, so you stop re-explaining.
What this means for you
If you're a developer or technical founder, the July 2026 Codex update is the one that makes Codex safe to use for the boring middle of real projects — multi-file refactors, SDK migrations, test backlogs — because the persistent-thread + review combination catches the failures before they hit main. Start with Free to learn the surface, then Plus for the IDE extension and cloud tasks.
If you're a marketer or content operator, the direct application isn't "write my content" — that's still ChatGPT's job. Codex's value is on the build side: turning a content spec into a working landing page, templating an SEO tracker as a small internal app, automating the publish-and-verify loop. The July update's review workflow is the part that keeps a human in the approver seat — which is the right line for content production, where originality and source-verified claims are the moat.
If you're deciding between Codex and Claude Code, both are now serious agents. Codex's edge is the desktop-app integration with ChatGPT and the plugin marketplace model; Claude Code's edge is local-first execution and tighter terminal control. For most teams the practical answer is both — ChatGPT Plus ($20/mo) already includes Codex, so the marginal cost of trying it is zero on top of a plan you probably already have.
FAQ
Q: What is OpenAI Codex in 2026?
A: OpenAI Codex is an autonomous coding agent built into ChatGPT. It reads your repository, edits files across multiple files, runs tests in a sandboxed environment, and proposes pull requests. As of July 2026 it ships as a CLI (@openai/codex), an IDE extension (VS Code, Cursor, Windsurf), a web surface at chatgpt.com/codex, and the ChatGPT desktop app on macOS and Windows.
Q: Is Codex free with ChatGPT Plus? A: Yes, but Plus ($20/mo, 15–90 GPT-5.6 Sol messages per 5-hour window, 10–60 cloud tasks per 5h) is the first tier with full cloud features. The Free plan includes limited Codex access with the lowest limits of any plan and no cloud task delegation; Go ($8/mo) adds CLI access without cloud tasks. See OpenAI's Codex plan limits and pricing.
Q: Do ChatGPT and Codex share the same usage limit? A: On Plus and Pro, yes — one combined 5-hour rolling window covers both. On Free and Go, Codex access is separate and limited. When the window is exhausted, Plus and Pro users can buy credit top-ups or wait for the reset. A weekly cap also applies.
Q: What model does Codex use in July 2026?
A: GPT-5.6 became generally available on July 9, 2026 across ChatGPT, Codex, and the API. GPT-5.6 ships as Sol, Terra, and Luna variants in ChatGPT and Codex. GPT-5.3-Codex remains available for explicit selection; ChatGPT Pro users can also access GPT-5.3-Codex-Spark as a research preview. You can switch models in the CLI with /model.
Q: Can Codex remember a project across days?
A: As of CLI 0.146.0 (July 29, 2026), yes — you can name sessions with /new, pin important threads, switch between side conversations without closing them, and fork threads with paginated history. Submitted messages, final responses, approval settings, and failed-turn errors are preserved across interruptions, replays, imports, and forks.
Q: How is Codex different from GitHub Copilot? A: Copilot is keystroke-level inline autocomplete inside your IDE; Codex is an autonomous agent that clones your repo, writes multi-file changes, runs tests, and opens a pull request. Copilot augments your editing; Codex replaces the editing session for well-scoped tasks. They operate at different levels of abstraction.
Q: Does Codex work without a ChatGPT subscription? A: You can run Codex CLI on an OpenAI API key with no subscription at all — usage bills at standard API token rates. You lose cloud features (GitHub integration, Slack, code review) and get delayed access to newer models. For most users the recommended path is to authenticate with your ChatGPT subscription.
Discussion
0 comments