OpenAI Codex CLI launched in April 2025 as an open-source terminal agent for coding. It has grown into one of the most popular AI coding tools on the market. On September 26, 2026, its GitHub repository had 126,581 stars and its npm package recorded 83.0 million downloads across the prior 30 complete API days. OpenAI reported more than 5 million weekly users in June and 20 million active users in August, without publishing a time window for the newer figure.
In this article, you'll find the latest numbers, data, and facts about Codex CLI. From GitHub stars and npm downloads to benchmarks and pricing to technical specifications and enterprise adoption.
- 20M active Codex users (OpenAI, August 21, 2026, without a published activity window), after 5M+ weekly users in June
- 126,581 GitHub stars, 19,760 forks, and 83.0M npm downloads across the complete 30-day period from August 26 to September 24, 2026
- Rewritten from TypeScript to Rust in June 2025. 1,150+ total releases, Apache 2.0 license
- GPT-5.5 held the previous Terminal-Bench 2.0 record; GPT-5.6 Sol, generally available since July 9, 2026 and initially the default in OpenAI Codex, leads on Terminal-Bench 2.1 and SWE-Bench Pro among GA models
- OpenAI valued at $852B (March 2026), $40B+ ARR (August 2026), $180B total funding raised
1. What Is OpenAI Codex CLI?
Codex CLI is OpenAI's open-source AI coding agent for the terminal. You describe what you want in natural language. Codex writes code, creates files, runs commands, and handles version control. It runs directly in your terminal, sandboxed for safety.
The project launched on April 16, 2025, and has evolved rapidly. Here are the key milestones:
These user figures only make sense together with their definitions. OpenAI published a weekly figure on June 2. Tibo Sottiaux reported 20 million active users on August 21, but without an activity window. The 20 million figure therefore cannot be labeled as WAU or used to calculate a directly comparable growth rate from June.
Over 1,150 releases in a little more than a year and a half through September 26, 2026 works out to a bit more than two updates per day. Few open-source AI projects move at that pace.
2. GitHub Stars and npm Downloads
Codex CLI is fully open source under the Apache 2.0 license. Here are the key repository metrics:
- 126,581 GitHub stars (September 26, 2026)
- 19,760 forks (September 26, 2026)
- 83.0M npm downloads (complete 30-day period, August 26 to September 24, 2026)
- 543.1M npm downloads (previous 12 months to September 24, 2026)
- 1,150+ total releases
- ~4.4 million GitHub release downloads
- Language: 96.4% Rust (since June 2025 rewrite)
2.1. npm Download Growth
The npm download numbers show strong growth. After 41.8 million in May and 46.2 million in June, Codex reached 60.2 million in July and 74.3 million downloads in August, the latest complete calendar month. I do not use a September monthly figure yet because the available 64.4 million covers only September 1 to 25:
The jump from February to March 2026 stands out. By August, monthly downloads had reached 74.3 million, more than 900 times the 82,000 downloads in the launch month. The 83.0 million for the last 30 complete API days uses a different window, so it is not another calendar-month point.
2.2. GitHub Stars Comparison
Codex CLI has strong numbers, but the competition has moved fast. Here is how the three big AI coding CLIs compare on stars (note: the Gemini CLI has since been retired):
Claude Code leads with 148,179 stars. Codex CLI is second with 126,581, about 17% behind, followed by Gemini CLI at 107,163. GitHub stars do not measure usage through ChatGPT subscriptions or the desktop app, both important distribution paths for Codex.
3. User Milestones and Reach
OpenAI has published different user metrics for Codex. On June 2, 2026, it reported more than 5 million weekly users, around 20% of whom were knowledge workers. On August 21, Tibo Sottiaux of OpenAI reported 20 million active Codex users. The newer figure has no published activity window, so it cannot be read as WAU or as a fourfold increase from June.
The sources are the April announcement, OpenAI's June post, and the August 21 update. They are useful milestones, but not one consistent time series because their definitions differ.
The 20 million active-user report still signals how far Codex has grown beyond a conventional terminal audience. For context, ChatGPT overall has more than 1 billion weekly active users, per OpenAI CMO Colin Fleming (Adweek, August 31, 2026), and more than 50 million paying subscribers. Codex remains a smaller but rapidly growing part of the OpenAI ecosystem.
4. npm Downloads Comparison
Monthly npm downloads are a useful proxy for developer adoption. Here is how the three CLI tools compare over the last 30 days:
Over the period from August 26 to September 24, 2026, Codex leads with 83.0 million npm downloads, versus 57.0 million for Claude Code. Gemini CLI records 1.4 million. npm counts package requests, not individual people, and it does not capture Codex use through ChatGPT or the desktop app.
npm downloads don't capture the full picture for Codex. Many users access Codex through the desktop app or through ChatGPT's web interface (the Codex Web agent), neither of which shows up in npm stats.
5. VS Code Extension
Both Codex and Claude Code offer VS Code extensions. These Marketplace figures are from September 26, 2026:
Claude Code leads on extension installs, 26.4 million versus 14.8 million for Codex, and rating, 3.7 versus 3.1 out of 5. The Claude Code extension and the Codex extension cover only one access path, not total use of either product.
The "Most Loved" category in the Pragmatic Engineer Survey paints an even clearer picture. Codex does not appear in the top three:
Codex doesn't appear in the "Most Loved" top three. Developers use it, but they don't love it the way they love Claude Code or Cursor. With the newer models and the desktop app, that could change, but for now the satisfaction gap is real and worth tracking.
6. Coding Benchmarks
OpenAI has released several specialized models for Codex. The benchmark progression shows steady improvement:
GPT-5.5 sets the current GA record at 82.7% on Terminal-Bench 2.0, released April 23, 2026. Terminal-Bench tests how well an AI agent handles real-world terminal tasks end to end. That was the strongest GA result on the 2.0 version of the benchmark; on the newer Terminal-Bench 2.1, GPT-5.6 Sol has since raised the bar to 88.8%.
GPT-5.4 trades a small amount of Terminal-Bench performance (75.1% vs. 77.3% from GPT-5.3-Codex) for a massive context window increase to 1 million tokens. GPT-5.5 delivers both the larger context and a new benchmark record.
Since July 9, 2026, GPT-5.6 Sol has been generally available and was introduced as the default coding model in OpenAI Codex. GPT-6 Astra, Sol, and Luna are now available there too. According to reporting by Axios and Bloomberg, the GA rollout started one day after the Trump administration lifted restrictions on July 8, 2026 (a characterization the White House disputes), and covers ChatGPT (Plus/Pro/Business/Enterprise), ChatGPT Work, Codex, and the API. Sol posts 88.8% on Terminal-Bench 2.1 and 64.6% on SWE-Bench Pro, still trailing Claude Mythos 5 and Fable 5 on the harder benchmark. On the Artificial Analysis Coding Agent Index, Sol scores 80, versus Fable 5's 77.2. The larger Sol Ultra variant scores 91.9% on Terminal-Bench. Pro and Enterprise plans additionally get Ultra mode with four parallel sub-agents. Also new: Programmatic Tool Calling with model-generated JavaScript running inside a V8 sandbox. Source: openai.com/index/gpt-5-6/.
7. Pricing and Costs
Codex CLI is included in all ChatGPT subscription plans. Unlike Claude Code (which requires a separate Claude subscription), Codex comes bundled with ChatGPT:
7.1. API Pricing
For API access, OpenAI offers several model options. The GPT-5.6 Sol and GPT-6 Astra prices below apply to requests with up to 272,000 input tokens; longer requests use higher rates.
GPT-5 has a lower input price but a higher output price than codex-mini-latest, as the table above shows. For code generation tasks where output tends to be longer than input, codex-mini-latest may be the more economical choice. The newer GPT-5.6 family spans a wider range, from budget-friendly Luna up to the flagship Sol model, which costs significantly more for stronger coding performance.
7.2. Price Comparison with Competitors
Codex and Claude Code both start at $20/month. OpenAI also offers Pro at $100 per month. Existing Pro $200 subscriptions continue, but new Pro $200 sign-ups are currently paused.
At the entry level, both Codex (via ChatGPT Plus) and Claude Code (via Claude Pro) cost $20/month. The value proposition differs: ChatGPT Plus includes Codex as one feature among many, alongside GPT-6 Sol/Luna, GPT Image 2, browsing, and file uploads. Claude Pro is more focused on coding and long-form tasks.
8. Technical Specifications
The technical overview:
The sandbox architecture is worth highlighting. Codex uses Seatbelt on macOS and Landlock + seccomp on Linux to isolate executed code. That means Codex can run commands in your terminal without risking damage to your system. It is a design choice that prioritizes safety over convenience, and it is one of the areas where Codex genuinely differentiates itself.
The Rust rewrite is another notable decision. Moving from TypeScript to Rust in June 2025 significantly reduced startup time and memory usage. At 96.4% Rust, Codex is one of the few AI tools built primarily in a systems language.
8.1. Context Window Comparison
Even the standard Codex CLI ships with 256K tokens, ahead of Claude Code's 200K. With GPT-5.4, GPT-5.5, and GPT-5.6, Codex jumps to 1 million tokens, matching the Antigravity CLI:
With GPT-5.4, GPT-5.5, and GPT-5.6, Codex offers one of the largest context windows in the AI coding space at 1 million tokens. That is roughly 750,000 words, or about 2,500 pages. Large enough to fit most entire codebases in a single context window.
9. OpenAI Financials
Codex doesn't exist in isolation. It is part of OpenAI's broader ecosystem. Understanding the parent company's financials provides context for where Codex fits:
- OpenAI total revenue: $40B+ ARR (Bloomberg, August 13, 2026), ~$3.3B per month
- ChatGPT: 1B+ weekly active users (end of August 2026), 50M+ paying subscribers
- Valuation: $852B (March 31, 2026)
Codex is included in all ChatGPT subscriptions, so its revenue is not broken out separately. That said, Codex drives subscription upgrades: many developers sign up for ChatGPT Plus or Pro specifically for the coding agent.
From $6B ARR at the start of 2025 to more than $40B by August 2026, Bloomberg reported, more than double the run rate from the end of 2025. Anthropic's run rate reached more than $65B by the end of July 2026 according to Bloomberg, and Google's Gemini product line generates roughly $5B.
OpenAI's valuation, total funding, and annualized revenue, shown in the chart above, are staggering numbers even by AI industry standards. Anthropic's May 2026 Series H valued it at $965 billion, surpassing OpenAI in private market valuation.
9.1. Latest Funding Round
OpenAI's $122 billion round in March 2026 was the largest venture capital deal in history. The investor breakdown:
Amazon alone invested $50 billion. Nvidia and SoftBank contributed $30 billion each. For comparison, Anthropic's largest round was $30 billion total. OpenAI raised 4x that in a single deal.
9.2. AI Company Valuations
Despite its massive funding round, OpenAI is no longer the most highly valued AI startup. Anthropic overtook it in May 2026:
Anthropic reached a $965 billion valuation after its Series H in May 2026, surpassing OpenAI, as the chart above shows. Both companies filed for IPOs confidentially in June 2026 (TechCrunch). xAI is no longer valued as a standalone company: it merged with SpaceX in February 2026, and the combined entity is valued at $1.25 trillion, with xAI's implied share at roughly $250 billion.
Bloomberg and other outlets reported on September 15 and 16, 2026 that OpenAI is in talks for a new funding round at a valuation of up to $1.5 trillion, citing growing Codex demand and the new GPT-6 models among the reasons. No decision has been made yet. Reporting attributed to The New York Times also points to OpenAI's IPO, originally expected this fall, slipping to 2027.
9.3. Revenue Comparison
Where does OpenAI stand relative to the competition?
Anthropic now leads on run-rate revenue, reaching more than $65 billion ARR by the end of July 2026 according to Bloomberg, as the chart above shows. That's up from roughly $9 billion at the end of 2025. OpenAI's annualized revenue topped $40 billion by August 2026, Bloomberg reported. Both companies confidentially filed for IPOs in June 2026 (TechCrunch), setting up a head-to-head race that will be defined in part by their AI coding tools.
10. Enterprise Adoption
The enterprise numbers are OpenAI's strongest argument for Codex:
- 90% of Fortune 100 companies use OpenAI products (including Codex)
- 6x enterprise growth since January 2026
- 1B+ weekly active users across ChatGPT (50M+ paying)
- ~4,500 employees (Financial Times, March 2026), with a stated plan to reach 8,000 by the end of 2026
10.1. Named Enterprise Customers
10.2. Codex Security
In March 2026, OpenAI launched Codex Security in preview. The early numbers are significant:
- 1.2 million commits scanned
- 10,561 high-severity findings
That is roughly 1 high-severity security issue per 114 commits. Codex Security automates vulnerability detection in the CI/CD pipeline, a direct competitor to Snyk and SonarQube but powered by LLMs. Enterprise customers care deeply about this. Security tooling is often the deciding factor for large organizations choosing an AI coding platform.
11. Survey Data and Market Position
Two major surveys provide context for Codex's market position:
11.1. Pragmatic Engineer Survey (February 2026)
The most comprehensive AI coding survey to date (~900 developers):
- Claude Code ranked #1 most-used and #1 most-loved AI coding tool
- 95% of developers use AI tools weekly
- Codex ranked behind Claude Code and GitHub Copilot in overall usage
11.2. JetBrains Survey (January 2026)
The JetBrains numbers (3% usage) look low, but they were collected in January 2026, before the desktop app launch and the GPT-5.3-Codex/GPT-5.4 releases. The user milestones published later in 2026 suggest that adoption has accelerated dramatically since then.
For complex coding tasks (multi-file refactoring, architecture decisions, large-scale debugging), Claude Code leads at 44%. Codex / ChatGPT sits at 19%. This is the segment where terminal agents have the most impact, and where Codex has the most room to grow.
12. Competitor Overview
Important context for this comparison: the Gemini CLI was retired on June 18, 2026 (its successor is the Antigravity CLI), but its repo numbers remain a useful baseline. And contrary to popular belief, Claude Code is not open source:
Each tool has clear strengths. Codex has the reported milestone of 20 million active users, a Rust codebase, strong Terminal-Bench results, and now leads the current npm period. Claude Code leads on GitHub stars, Marketplace installs, satisfaction, and complex-task usage. The 20 million Codex figure has no published activity window, so it cannot be compared as WAU with other products.
For a much deeper comparison, read our article on Claude Code vs. OpenAI Codex.
13. Interesting Facts and Records
To close out, here are the most notable facts about Codex CLI:
- Rust rewrite in 2 months: Codex was rewritten from TypeScript to Rust in June 2025, just two months after the initial open-source release. 96.4% of the codebase is now Rust.
- 1,150+ releases in a year and a half: That averages out to a bit more than two releases per day. Few open-source projects maintain this pace.
- From 82K to 74.3M npm downloads: Monthly npm downloads grew more than 900x from the April 2025 launch month to August 2026, the latest complete calendar month.
- 5M weekly users and 20M active users: OpenAI confirmed more than 5M weekly users in June. Tibo Sottiaux reported 20M active Codex users on August 21 without an activity window, so the two figures are not directly comparable.
- 82.7% Terminal-Bench 2.0: GPT-5.5 holds the highest Terminal-Bench 2.0 score among all GA AI coding models tested.
- GPT-5.6 Sol/Terra/Luna generally available since July 9, 2026: GA started on July 9, 2026, according to Axios/Bloomberg one day after government restrictions were lifted. Sol was introduced as the Codex default in July. Luna is the default in regular Free/Go chat, while Terra is available to Free and Go with limited access in ChatGPT Work and Codex. Pro and Enterprise plans additionally get Ultra mode with four parallel sub-agents. Sol hits 88.8% on Terminal-Bench 2.1 and 64.6% on SWE-Bench Pro. On the Artificial Analysis Coding Agent Index, Sol scores 80, edging out Fable 5's 77.2, and the larger Sol Ultra variant pushes the Terminal-Bench score even higher, to 91.9%.
- GPT-6 Sol and GPT-6 Luna since September 22, 2026: Both models are available in Codex for Plus, Pro, Business, Enterprise, and Edu users, each 50% cheaper than its GPT-5.6 predecessor. OpenAI positions Sol for complex, agentic coding tasks.
- $122 billion funding round: OpenAI's March 2026 round was the largest VC deal in history. Amazon invested $50 billion alone.
- Codex Security scanned 1.2 million commits: Found 10,561 high-severity findings in the preview period.
- 90% of Fortune 100: Nearly all Fortune 100 companies use OpenAI products, giving Codex a built-in enterprise distribution channel.
The bottom line: Codex CLI has established itself as a serious contender in the AI coding market. It now leads the current npm download period but still trails Claude Code on GitHub stars and satisfaction, and OpenAI's ecosystem advantages (ChatGPT integration, enterprise relationships, and massive funding) give it a strong foundation for growth.
If you want to compare Claude Code and Codex in detail, check out our Claude Code vs. OpenAI Codex comparison.
