← Back to research
·9 min read·company

Codex

Codex is OpenAI's coding agent platform with desktop apps (macOS + Windows), cloud sandboxes, CLI, IDE extensions, and Slack/Linear integrations — powered by the GPT-5.5 / GPT-5.4 model lineup.

Key takeaways

  • Desktop app now runs on macOS and Windows (March 2026) as a command center for multi-agent workflows with built-in worktrees and cloud environments
  • Adoption is massive — 4M+ weekly developers by April 2026 and over $1B in annualized revenue, with non-developers now ~20% of users
  • Same agent across desktop app, IDE extensions, CLI, web, Chrome extension, and mobile — plus Slack, Linear, and GitHub integrations

FAQ

What is OpenAI Codex?

Codex is OpenAI's coding agent platform available as a desktop app (macOS and Windows), CLI, IDE extension, Chrome extension, web interface, and Slack/Linear/GitHub integrations.

How much does Codex cost?

Included with ChatGPT Free, Go ($8/mo), Plus ($20/mo), Pro (from $100/mo), Business, Edu, and Enterprise plans. API-key users pay per token; subscribers can buy extra credits.

What models power Codex?

As of June 2026 the lineup is GPT-5.5 (frontier), GPT-5.4, and GPT-5.4 mini. Earlier codex-specific models (GPT-5.3-Codex, codex-1) have been sunset for subscribers.

Does Codex compete with Tembo?

Partial overlap — both orchestrate parallel agents. Codex is OpenAI-only but now ships Linear and Slack integrations; Tembo is agent-agnostic with Jira, signed commits, and BYOK.

Executive Summary

Codex is OpenAI's comprehensive coding agent platform, offering a unified experience across desktop app (macOS and Windows), web interface, CLI, IDE extensions, Chrome extension, and mobile. It targets developers who want parallel agent execution with visual orchestration, powered by the GPT-5.5 / GPT-5.4 model lineup.[1][2] By April 2026 Codex had surpassed 4 million weekly developers and over $1 billion in annualized revenue, and its Skills, Automations, and plugin features push it beyond code writing toward general autonomous work.[3][4]

AttributeValue
CompanyOpenAI
Founded2015
Funding$11.3B+
Employees~3,000
HeadquartersSan Francisco, CA

Product Overview

Codex is OpenAI's coding agent platform, now available across multiple surfaces:[5][6]

All surfaces connect via your ChatGPT account — start in the app, continue in your editor, finish in terminal.[5][7][6]

Codex launched powered by codex-1 (an o3 variant optimized for software engineering)[8] but the model lineup has turned over rapidly. As of June 2026, Codex runs on GPT-5.5 (frontier, released April 23, 2026), GPT-5.4, and GPT-5.4 mini; the codex-suffixed models (GPT-5.3-Codex, GPT-5.3-Codex-Spark) have been sunset for ChatGPT subscribers.[1][2][9]

Key Capabilities

CapabilityDescription
Multi-Agent OrchestrationRun parallel agents across projects with visual monitoring
SkillsExtend to code understanding, prototyping, and documentation aligned with team standards
AutomationsUnprompted background work: issue triage, alert monitoring, CI/CD
Computer UseIn-app browser and computer-use capabilities (April 2026)[1]
Sites PluginCreate and deploy hosted web projects, in preview (June 2026)[1]
Cross-Surface SyncSame agent across desktop app, IDE, CLI, web, and mobile
Code AnalysisReview for bugs, logic errors, and edge cases with root cause analysis

Product Surfaces / Editions

SurfaceDescriptionAvailability
Codex AppDesktop command center for multi-agent workflows — macOS (Feb 2026), Windows via PowerShell (Mar 2026)[1]GA
Codex WebCloud-based at chatgpt.com/codex with parallel sandboxesGA
Codex CLIOpen-source terminal agent (Apache 2.0)GA
IDE ExtensionsVS Code, Cursor, Windsurf integrationGA
Chrome ExtensionBackground operation across browser tabs (May 2026)[1]GA
MobileVia ChatGPT app, remote connection to Mac hosts (May 2026)[1]GA
IntegrationsSlack, Linear, GitHub (code review, GitHub Action), SDK[6]GA

Technical Architecture

Cloud architecture:[8]

  • Isolated container per task
  • Internet access configurable (now enabled)
  • Verifiable evidence via terminal logs and test outputs
  • AGENTS.md files guide agent behavior

Key Technical Details

AspectDetail
DeploymentCloud (sandboxes) + Local (CLI, desktop app)
Model(s)GPT-5.5, GPT-5.4, GPT-5.4 mini (as of June 2026)[2]
IntegrationsVS Code, Cursor, Windsurf, GitHub, Slack, Linear
Open SourceCLI only (Apache 2.0) — 90.3k GitHub stars, 13.3k forks as of June 2026[10]

CLI installation:[10]

npm install -g @openai/codex
# or
brew install --cask codex

Adoption

MetricValue (as of June 2026)
Weekly developers4M+ (April 2026), up from 3M two weeks prior[3]
Non-developer share~20% of ~5M weekly users, growing 3x faster than engineers[3]
Annualized revenue$1B+ (company claim)[4]
Enterprise growth6x since January 2026 within ChatGPT Business/Enterprise (company claim)[11]
GitHub stars (CLI)90,298[10]

OpenAI cites Virgin Atlantic (test coverage), Ramp (code review), Notion (feature building), Cisco (large-repo reasoning), and Rakuten (incident response) as enterprise users — all company-attributed claims.[11]

Codex for Open Source (announced March 2026): eligible open-source maintainers get six months of ChatGPT Pro with Codex free, plus conditional access to Codex Security for core maintainers with write access.[12]


Strengths

  • Unified experience — Same agent across desktop app (macOS + Windows), IDE, CLI, web, Chrome extension, and mobile with synced context[1]
  • Multi-agent parallel — Built-in worktrees + cloud environments for concurrent work
  • Skills & Automations — Beyond code to full engineering workflows without prompting
  • Rapid model cadence — GPT-5.3-Codex, Spark, GPT-5.4, and GPT-5.5 all shipped between February and April 2026[1]
  • Massive adoption — 4M+ weekly developers, $1B+ annualized revenue[3][4]
  • Open source CLI — Apache 2.0 with 90k+ GitHub stars[10]
  • OSS program — Free ChatGPT Pro + Codex Security for open-source maintainers[12]

Cautions

  • No Linux desktop app — App covers macOS and Windows; Linux users are limited to CLI/IDE/web[1]
  • Forced model churn — GPT-5.2 and GPT-5.3-Codex were sunset for subscribers despite users preferring them over GPT-5.5 for some workloads[9]
  • Model behavior complaints — Users reported GPT-5.3-Codex as "lazy and irrationally terrified of breaking anything," plus false-positive safety refusals in newer models[13]
  • OpenAI lock-in — Only works with OpenAI models; no model choice, no BYOK
  • Integration gaps — Slack, Linear, and GitHub now covered, but no Jira; signed commits still missing for regulated industries

Pricing & Licensing

TierPriceIncludes
ChatGPT Free$0Limited Codex access
ChatGPT Go$8/moLimited Codex access
ChatGPT Plus$20/moe.g. 15–80 GPT-5.5 local messages per 5-hour window
ChatGPT ProFrom $100/mo5x or 20x rate-limit options (80–400 / 320–1,600 GPT-5.5 local messages per window)
ChatGPT BusinessPay-as-you-goAdmin controls
ChatGPT Enterprise / EduContact salesSSO, compliance

All figures from the Codex pricing page as of June 2026.[2]

Overage and API options: Subscribers can purchase additional credits priced per million tokens (GPT-5.5 costs 125 credits input / 750 output per million tokens; GPT-5.4 mini is ~85% cheaper); API-key users pay standard token-based API pricing with no fixed rate limits.[2]

Licensing model: Subscription (bundled with ChatGPT plans) + usage-based API

Hidden costs: Heavy usage requires Pro 20x or credit purchases; cloud tasks and code reviews draw from separate allowances[2]


Competitive Positioning

Direct Competitors

CompetitorDifferentiation
Claude CodeCodex has desktop apps with visual orchestration; Claude Code is terminal-first
ConductorBoth desktop orchestrators with worktrees; Codex is first-party from model provider
OpenCodeOpenCode is model-agnostic; Codex is OpenAI-only but more polished
CursorDifferent form factor (IDE vs orchestration layer); complementary
TemboTembo is agent-agnostic with Jira, signed commits, BYOK — enterprise focus

When to Choose Codex Over Alternatives

  • Choose Codex when: You want unified desktop app + IDE + CLI experience with OpenAI models
  • Choose Claude Code when: You prefer Anthropic models or need terminal-first simplicity
  • Choose Conductor when: You need model-agnostic orchestration with multiple agent types
  • Choose Tembo when: You need enterprise integrations (Jira, signed commits, BYOK)

Ideal Customer Profile

Best fit:

  • Developers already in the ChatGPT/OpenAI ecosystem
  • Teams wanting parallel agent execution with visual monitoring
  • Mac and Windows users who prefer native desktop apps over web/terminal
  • Non-developer knowledge workers adopting agentic workflows (now ~20% of users)[3]

Poor fit:

  • Enterprises requiring Jira integration or signed commits
  • Teams wanting model flexibility (Anthropic, Google, etc.)
  • Linux desktop-app users
  • Teams that need long-term model stability — OpenAI sunsets models on short notice[9]

Viability Assessment

FactorAssessment
Financial HealthStrong — $1B+ annualized Codex revenue within a ~$6B quarter[4]
Market PositionLeader — 4M+ weekly developers, fastest-growing coding agent[3]
Innovation PaceRapid — four model releases and five new surfaces shipped Feb–Jun 2026[1]
Community/EcosystemActive — 90k+ GitHub stars on the CLI, OSS maintainer program[10][12]
Long-term OutlookPositive — expanding beyond developers into general enterprise work

OpenAI has unmatched resources and market position. The main risks are model-churn frustration among power users and competition from Anthropic on code quality — but OpenAI's iteration speed is unmatched.


Bottom Line

Codex has become OpenAI's flagship agentic product: desktop apps on macOS and Windows, IDE extensions, CLI, web, Chrome extension, mobile, and Slack/Linear/GitHub integrations — all on one account, now serving over 4 million weekly developers with $1B+ in annualized revenue.

Skills, Automations, computer use, and the Sites plugin push Codex past code writing toward general autonomous work — non-developers are now its fastest-growing user segment. This is the "background agents" vision that enterprise tools like Tembo have been building toward.

Recommended for: Developers in the ChatGPT ecosystem who want parallel agents in a polished desktop app with cross-surface sync.

Not recommended for: Enterprises needing Jira integration, signed commits, BYOK, or multi-model flexibility — or teams that can't absorb OpenAI's aggressive model sunsets.

Outlook: The February prediction of Windows expansion has already shipped. Expect a Linux app, deeper enterprise plugins, and continued blurring of the line between coding agent and general work agent as non-developer adoption accelerates.


Research by Ry Walker Research • methodology