Codex pricing 2026: is Codex free, what the CLI costs

Codex pricing

Codex has no price of its own. It is an entitlement on a ChatGPT plan, plus a second, entirely separate way to pay — an API key — that most buyers never evaluate.

Plan	Price	What the plan gives you in Codex
ChatGPT Free	$0	GPT-6 Luna at Standard speed in the desktop app, “subject to rollout”
ChatGPT Go	$8/month	The same single line as Free — Luna, desktop, Standard speed
ChatGPT Plus	$20/month	Codex on web, CLI, IDE extension and iOS; cloud features; GPT-6.1 Sol and GPT-6 Luna
ChatGPT Pro	$100, $200 or $500/month	No five-hour cap; Ultrafast on the $500 tier only
ChatGPT Business	$20/seat annual, $25 monthly	2+ seats; Standard Business usage matches Plus for most features
ChatGPT Enterprise / Edu	Contact sales	Priority processing, SCIM, EKM, RBAC, residency controls
API key	Per token, no subscription	CLI, SDK and IDE extension only — no cloud features at all

Prices verified 18 October 2026 against OpenAI’s Codex pricing documentation. Note for anyone with an old bookmark: developers.openai.com/codex now 308-redirects to learn.chatgpt.com/docs, and the plan cards live on the Codex page rather than on OpenAI’s consumer pricing page. The pricing tracker carries the dated change log.

What the Codex CLI costs

Nothing, to install. npm install -g @openai/codex, brew install --cask codex, or the shell one-liner from chatgpt.com/codex/install.sh. The cost arrives at the login prompt, and the choice there is the single most consequential pricing decision on this page:

Sign in with ChatGPT (codex login) draws on your subscription’s included usage. No per-token billing, and you get the cloud half of the product — GitHub code review, PR delegation, Slack and Linear.

Sign in with an API key (printenv OPENAI_API_KEY | codex login --with-api-key) bills through your OpenAI Platform account at standard API rates. In OpenAI’s words: “When you sign in with an API key, Codex uses standard API pricing instead of included ChatGPT plan credits.” You lose every cloud feature — the docs mark GitHub delegation, code review, Slack, Linear and cloud chats unavailable on an API key. What you gain is no rate limit, no five-hour window, and the ability to run Codex in CI/CD, which is what OpenAI recommends the mode for.

The two are not either/or. Both the desktop app and the CLI support either method, and the docs note that any user can “run extra local chats using an API key” once a plan limit is hit — so the normal professional pattern is a Plus seat for interactive work with an API key configured for automation.

One cost that sits outside both modes: if you point the Codex harness at a third-party provider rather than OpenAI, that provider’s server-side tool calls bill on their own meter. DeepSeek starts charging $0.75 per 1,000 search and fetch calls on 31 October 2026 for endpoints Codex already calls for free today, and for a search-heavy agent those calls can exceed the token cost of the task.

What you actually get on a Plus seat

OpenAI publishes estimates rather than hard caps, measured in local messages per five-hour window. These are the only usage numbers it quantifies, and they cover Plus and Standard Business only:

Model	Messages per 5 hours (Plus / Standard Business)
GPT-6 Astra	5–45
GPT-6.1 Sol	15–160
GPT-6 Sol	15–150
GPT-6 Luna	350–3,000

Three caveats OpenAI prints alongside them: local messages and cloud chats share one allowance; weekly limits may also apply on top of the five-hour window; and cloud tasks can consume more than a local message does. Pro plans currently have no five-hour limit at all, and Enterprise or Edu seats on flexible pricing have no fixed rate limits. Check /status inside a CLI session or the dashboard at chatgpt.com/codex/settings/usage rather than trusting the ranges — a 10x spread between the low and high estimate is the vendor telling you the number depends entirely on task size.

The credit is worth exactly four cents

This is the arithmetic no vendor page does, and it settles whether a subscription is a discount or a prepayment.

OpenAI quotes Codex overage in “credits” and declines to publish a dollar value: “Credit purchase prices and applicable discounts depend on your plan or agreement.” But the same documentation prices one service in both units. Desktop voice costs $0.05 per minute against a usage budget, or 1.25 credits per minute on a credit-based workspace. That fixes a credit at $0.04 — and once you have that number, every model rate converts:

Model	Credits per 1M (in / cached / out)	At $0.04/credit	API list price
GPT-6 Astra	250 / 25 / 1,250	$10.00 / $1.00 / $50.00	$10.00 / $1.00 / $50.00
GPT-6.1 Sol	50 / 2.5 / 250	$2.00 / $0.10 / $10.00	$2.00 / $0.10 / $10.00
GPT-6 Sol	50 / 5 / 250	$2.00 / $0.20 / $10.00	$2.00 / $0.20 / $10.00
GPT-6 Luna	2.5 / 0.25 / 12.5	$0.10 / $0.01 / $0.50	$0.10 / $0.01 / $0.50

Twelve figures, twelve exact matches. Two independent checks agree: Ultrafast costs 6x in purchased credits and GPT-6.1 Sol Ultrafast lists at $12/$60 against a standard $2/$10 — 6x. And the one-time grant attached to the Pro 200 change was described as 62,500 credits worth about $2,500, which is $0.04 a credit again.

So overage credits carry no subscription discount and no markup. A Codex seat is a prepaid block of tokens at list price, and the only saving a subscription delivers is the allowance you never exhaust — which is exactly why the five-hour estimates above matter more than the headline price. Buying a bigger plan to get a better rate does not work; there is no better rate.

One genuine asymmetry in the subscription’s favour: Codex credit billing has no separate cache-write charge, while the API bills cache writes at $2.50 per million tokens on GPT-6.1 Sol and $12.50 on Astra. For agent work that rebuilds large prefixes, that is the subscription’s only structural edge — and it is a real one.

Paying for speed

Speed mode	Cost against included usage	Cost in purchased credits
Fast	2.5x	2x
GPT-6 Astra Ultrafast	8x	6x
GPT-6.1 Sol Ultrafast	8x	6x

“Priority processing” was renamed Fast mode on 30 July 2026, so older write-ups describing a priority tier are describing this. Ultrafast is the one worth understanding before you buy up the ladder: it is a surcharge the subscription absorbs rather than a free extra. Pro 500 buys roughly 300 output tokens per second, around 8x standard — but it drains included limits at 8x, so 25x of headroom behaves like about 3x if you leave it on. Worth it when a developer is blocked and watching; wasteful for background work. On the API the same mode is $12/$60 against the standard $2/$10. Image generation consumes included limits roughly 3–5x faster than text, and desktop voice bills $0.05 a minute.

The Pro ladder, and one date this month

A caveat on sourcing, because it affects the numbers most buyers want. OpenAI’s Codex documentation does not express Pro allowances as multiples of Plus — it says only that “Pro plans currently have no five-hour limit.” The 5x / 10x / 25x framing comes from OpenAI’s ChatGPT plan help articles, which have been returning 403 to automated checks since mid-October; the figures below are as last verified on 14 October 2026 and are flagged for re-verification rather than re-confirmed today.

On that basis: included usage on Pro 200 falls from 20x to 10x of Plus on 30 October 2026 with no price change. Subscribers who bought before 29 September 2026 keep the old 20x allowance through 29 October plus a one-time grant of 62,500 credits (about $2,500) expiring 31 December 2026. If you hold one of those seats it is the only plan in the lineup priced below the going rate of $20 per 1x of Plus usage — and that ends this month.

Against Claude. Claude Pro at $20/month includes Claude Code, and Claude Team Standard at $25/seat ($20 annual) includes it too, so the entry price is identical on both sides for individuals and teams alike. The team decision is not about price.

Recommendation. Already on ChatGPT Plus? You already have Codex — start there and watch /status. Hitting Plus limits weekly? Pro 100. Because credits buy tokens at list price, climbing the ladder buys headroom rather than a discount: skip Pro 200 unless you run multi-agent workflows all day, and buy Pro 500 only if you specifically want Ultrafast latency. Automating Codex in CI? Add an API key to the seat you already have rather than upgrading it.

What Codex is in 2026

Codex is an autonomous coding agent that runs in OpenAI’s cloud, reachable from four surfaces: the macOS app (launched 2 February 2026), the IDE extension for VS Code, Visual Studio and JetBrains, Codex Web, and the Codex CLI. OpenAI claims the CLI gets roughly 4x the token efficiency of Claude Code, meaning a $20 API budget buys about what $80 does on Claude’s API. The exact ratio resists independent verification, but the direction is consistent across community reports and it is the most under-sold thing about the product.

Around those sit multi-agent v2, image inputs, a Slack integration (@Codex in a thread opens a pull request), 90+ plugins, SSH to remote devboxes, and the Codex SDK. Chronicle, an on-device screen-capture memory feature for macOS, is Pro-only. Two procurement routes matter to enterprise buyers because they change who invoices you: Amazon Bedrock, GA since 1 June 2026, counts Codex usage against existing AWS commitments, and Oracle Universal Credits does the same on Oracle paper.

For scale: 5M+ weekly Codex developers as of June 2026, up from 3M in early April, disclosed alongside the acquisition of Ona, formerly Gitpod, whose persistent cloud sandboxes were bought to stretch agent runs from hours toward days.

The model lineup, and the dates that bite

Codex has run on the GPT-6 generation since 22 September 2026. The default is GPT-6.1 Sol (gpt-6.1-sol), which replaced GPT-6 Sol seven days after it shipped.

	GPT-6.1 Sol	GPT-6 Luna
Input / output per 1M tokens	$2.00 / $10.00	$0.10 / $0.50
Cached input	$0.10	$0.01
Cache write	$2.50	$0.125
Context	1,050,000 tokens	~1M tokens
Above 272K tokens	2x input, 1.5x output	2x input, 1.5x output
Knowledge cutoff	30 April 2026	30 April 2026

Rates verified 18 October 2026 against OpenAI’s pricing docs; the API cost calculator carries the full table. Batch and Flex tiers run at exactly half these rates, and data-residency endpoints add 10% for models released on or after 5 March 2026.

Three things that are easy to get wrong:

The cached rate is the one that matters. GPT-6.1 Sol held the $2/$10 headline from GPT-6 Sol but halved cached input from $0.20 to $0.10. Codex re-sends a large, stable repository prefix on every turn, so cache behaviour dominates the bill far more than the headline rate does.

GPT-6 Sol is not deprecated. A pinned gpt-6-sol configuration keeps working with no shutdown date — at double the cached-input rate, on a model OpenAI has stopped treating as current. The models it has deprecated on the Codex side are GPT-5.3-Codex, GPT-5.1 and GPT-5.4-Nano, announced 1 October 2026 with a firm 1 April 2027 shutdown; the model retirement tracker has the dates.

GPT-5.5 left the subscription surface on 14 October 2026 at noon PDT — ChatGPT, ChatGPT Work and Codex on every plan (what the retirement does and does not cover). The API is not affected, and GPT-5.5 keeps working in Codex sessions authenticated with an API key — a product retirement rather than a model shutdown, and a concrete reason some teams keep a key on the seat. Replacements for Codex work are GPT-6.1 Sol or GPT-6 Astra.

Two API notes: on Chat Completions gpt-6-luna supports function calling only with reasoning_effort set to none, so tool-using agents belong on the Responses API; and when you read launch benchmarks, check the effort column first, because headline scores are measured at xhigh and max while the documented default is medium.

What Codex does best

Codex and Claude Code are not competing on the same axis. Claude Code’s bet is interactive pair programming — you watch it work and redirect as it goes. Codex’s bet is asynchronous delegation: describe a task, walk away, review a pull request. Cursor sits at the interactive extreme, Codex at the asynchronous one, Claude Code in between. That single difference drives everything below.

Asynchronous task completion — the describe-walk-away-review loop is genuinely transformative for tests, documentation and API migrations, where watching is a waste of attention.

Token efficiency — the 4x claim against Claude Code is the feature most users do not appreciate until they see an API bill.

Speed — tasks that take Claude Code 3–5 minutes often finish in 1–2.

Real-time steering — interrupt a running agent with new context without losing state.

Multi-agent v2 — community reports put Codex ahead of Claude Code’s Agent Teams on stuck states and merge quality.

Image inputs and Slack — screenshot to working implementation, and @Codex in a thread removes the friction between “someone suggested a change” and “a PR exists.”

Where Codex falls short (vs Claude Code)

Less interactive by design. A feature when you want delegation, a limitation when you want to pair.

Verbose without prompting. The default bias is elaborate scaffolding over minimal solutions. Coachable, but you have to ask.

Weaker on sprawling codebases. Claude Code is more reliable for “understand this whole repository before changing anything.”

Narrower ecosystem. Claude Code’s MCP support reaches thousands of external tools; the Codex SDK’s ecosystem is smaller.

Explanation quality. Claude Code explains what it did and why; Codex delivers completed work — less useful when you want to learn from it.

A prompt-cache defect that costs real money. An issue filed 5 September 2026 reports that Codex changes reasoning effort by mutating request-level configuration rather than appending the configuration_update item OpenAI shipped with GPT-6 Astra on 3 September — which invalidates the prompt cache. On Astra that is expensive in a specific way: cache reads bill at $1.00/MTok and cache writes at $12.50/MTok, so each unnecessary rebuild of a large prefix costs 12.5x what preserving it does. OpenClaw merged support for the primitive on 4 September. Check your cached-read versus cache-write token counts before assuming the cache saving applies to you — and note this is the one place a subscription seat protects you, since credit billing has no cache-write charge.

Security and provisioning: three things to settle before rollout

None of these are reasons not to buy Codex. All three are configuration decisions a buyer, not the vendor, has to make.

Pin CLI 0.95.0 or later. Codex CLI was escaped in two separate 2026 disclosures, both through the structural gap that hit every major coding agent this year. Pillar Security found the “safe” command allowlist trusted the git show command name without validating its arguments, fixed in v0.95.0. Earlier, Cymulate showed apply_patch could create .codex/config.toml without approval when the directory did not already exist — the default state — letting notify commands execute outside the sandbox. Do not treat the allowlist as a security boundary.

Scope the GitHub connector to repositories, not organisations. Hacktron AI disclosed on 18 September 2026 that researchers reached OpenAI’s internal monorepo by taking over an employee’s account and simply prompting that account’s Codex to open a pull request, because the account held a live organisation-wide connection. No Codex vulnerability was involved. The lesson is provisioning: scope write access per repository, require human approval on agent-authored changes that the same credential cannot supply, and confirm that revoking a session also invalidates connected-app tokens. Full analysis.

Assume the harness is neutral. GreyNoise reports that the operator of an AI-orchestrated campaign against PaperCut NG/MF ran hundreds of agents on the Codex harness paired with a DeepSeek model, explicitly not OpenAI models, eventually touching 395 organisations in 48 countries. Nothing there indicates a Codex vulnerability — openai/codex is Apache-2.0 software anyone can point at anything. It establishes that the dangerous capability sits in the orchestration loop rather than the weights, so model-provider usage policies are bypassed by construction. Evaluate Codex on what it is permitted to reach in your environment. Full analysis.

The harness is now a product of its own

In August 2026 OpenAI documented the Codex app-server as a stable integration target and repositioned the harness as a platform — widely misreported as an open-sourcing, since the repository has been Apache-2.0 since April 2025. The Agents API went to public beta on 10 September 2026, exposing that managed harness behind one API call with no fee of its own, billing instead for OpenAI-hosted sandboxes at $0.09–$5.76 per hour of wall-clock time with a non-configurable one-hour idle deletion. US-only during beta, and zero data retention is unsupported.

That is a third way to pay for Codex, and it splits what you are evaluating in two: the Codex product is the bundled seat, while the Codex harness is an independently purchasable building block billed by rented sandbox time rather than by tokens.

Codex vs Claude Code vs Cursor

Job	Best	Why
Async delegation (walk-away tasks)	Codex	Built for exactly this workflow
Interactive pair programming	Cursor	Tightest editor loop
Multi-file refactors with reasoning	Claude Code	Context handling and explanations
Cost-sensitive API use	Codex	The 4x efficiency claim
Team workflows in Slack	Codex	The Slack integration is unique
Image-to-code	Codex	More polished than the alternatives
Ecosystem depth	Claude Code	MCP reaches thousands of tools
Mission-critical maturity	Claude Code	More years of heavy production use

The longer-form version of the middle row is in our Cursor vs Claude Code comparison.

What the 2026 pro dev stack costs

The old answer — Cursor plus Claude Code at about $40 a month — stopped being the default when SpaceXAI folded Cursor into a Grok-wide subscription ladder on 12 October 2026 and Cursor Pro at $20 ceased to exist, with non-Grok models now billed separately at API rates. That strengthens Codex’s position by accident: the bundled ChatGPT Plus seat is the cheapest way to get a frontier-model coding agent with the model cost included.

• Full coverage, $70: ChatGPT Plus ($20) + Claude Pro ($20) + Cursor Premium ($30) — and the Cursor line now carries variable third-party model spend on top.
• The new default, $40: ChatGPT Plus ($20) + Claude Pro ($20). Two agents with different philosophies, no metered model billing.
• Budget, $10: GitHub Copilot Pro alone — autocomplete and chat, 1,500 AI credits. Less capable, cheapest credible option.

Start at ChatGPT Plus because Codex is included, then add Claude Pro when you notice yourself wanting to watch the work rather than review it.

Who should and shouldn’t buy it

Buy Codex if you are already on ChatGPT Plus; prefer “describe the task and walk away” to pair programming; drive coding models through the API where the token efficiency compounds; run a Slack-centric team; do image-to-code work; or coordinate multi-agent workflows.

Look elsewhere if you want live pair programming (Claude Code or Cursor); need reliable whole-repository context (Claude Code); are learning from the AI’s explanations (Claude Code); or need the longest production track record for mission-critical workflows (Claude Code).

The verdict

Codex is the biggest tool re-evaluation of 2026, because asynchronous delegation unlocks work that never got delegated before — nobody wants to watch an AI code for 20 minutes.

On price, the honest read is that the $20 Plus seat is excellent value precisely until you exhaust it, at which point Codex costs exactly what the API costs and the subscription stops helping. That makes the five-hour estimates, not the plan price, the number to evaluate.

For a sense of what the tool does when cost is not a constraint, OpenAI published its own internal telemetry: 3.1 agent-workdays per human workday across its research organisation, with the median researcher burning more than $600 a day at API prices and the 90th percentile past $7,000. Read the caveat OpenAI prints itself: more than half of successful four-to-eight-hour tasks still required at least one human intervention.

FAQ

Is Codex free?

Partly. There is a free rung — ChatGPT Free lists GPT-6 Luna at Standard speed in the desktop app, “subject to rollout” — but it is enough to evaluate Codex, not to work with it. The Codex CLI is genuinely free to install on any plan. What costs money is the model usage behind it, which either comes out of a paid ChatGPT plan’s included allowance or is billed per token against an API key. There is no free tier of the cloud features at all.

Is Codex free for ChatGPT Plus users?

Yes — Codex is included with ChatGPT Plus at $20/month at no extra charge, across web, the CLI, the IDE extension and iOS, plus the cloud features (GitHub code review, PR delegation, Slack, Linear). “Included” means an allowance, not unlimited: OpenAI estimates 15–160 GPT-6.1 Sol messages per five-hour window on a Plus seat, local and cloud usage share that pool, and weekly limits may apply on top. Go beyond it and you buy credits or attach an API key.

How much does the Codex CLI cost?

The CLI is free software. Billing depends on how you sign it in. codex login with a ChatGPT account draws on your subscription’s included usage, with no per-token charge. codex login --with-api-key bills standard API rates through your OpenAI Platform account — $2 per million input tokens and $10 per million output on GPT-6.1 Sol, with cached input at $0.10 and cache writes at $2.50. API-key mode has no rate limit and works in CI/CD, but loses every cloud feature.

Are Codex credits cheaper than paying API rates?

No. OpenAI does not publish a dollar value for a credit, but its own documentation prices desktop voice at both $0.05 per minute and 1.25 credits per minute, which fixes a credit at $0.04. At that rate every published Codex credit price converts exactly to the API list price — GPT-6.1 Sol at 50 / 2.5 / 250 credits per million tokens is $2.00 / $0.10 / $10.00. Overage credits carry no discount. The only structural saving is that credit billing has no separate cache-write charge.

Which ChatGPT plan should I buy for Codex?

Plus at $20 for almost everyone — it unlocks the full product and every surface. Move to Pro 100 only when you hit Plus limits weekly; Pro plans currently have no five-hour cap. Because credits buy tokens at list price, climbing the ladder buys headroom rather than a better rate, so Pro 200 is for all-day multi-agent work and Pro 500 is worth it only if you specifically want Ultrafast latency. Teams start at Business, $20/seat on annual billing with a two-seat minimum.

Can I use Codex without a ChatGPT subscription?

Yes, with an API key. The API Key option covers the CLI, the SDK and the IDE extension, bills per token at standard API rates, and follows whichever models your key can reach. It is the right choice for automation and CI/CD. What it cannot do is any cloud-based feature — GitHub delegation and code review, Slack, Linear and cloud chats are all marked unavailable on an API key, and Codex cloud requires a ChatGPT sign-in.

What changed

• 18 Oct 2026: Restructured answer-first for pricing. Added the plan table, the five-hour usage estimates and the credit arithmetic: a Codex credit is $0.04, so overage buys tokens at API list price. OpenAI’s Codex docs moved to learn.chatgpt.com/docs; a ChatGPT Go tier at $8/month appears on them. The 5x/10x/25x Pro allowance figures could not be re-verified — the help articles now 403.
• 14 Oct 2026: GPT-5.5 left Codex, ChatGPT and ChatGPT Work on every plan at noon PDT; the API keeps it, and the last GPT-4o / GPT-4.1 snapshots were deprecated for 14 January 2027 — see analysis.
• 13 Oct 2026: Stack costs rebuilt after Cursor’s ladder was replaced on 12 October.
• 11 Oct 2026: Claude Team Standard at $25/seat now includes Claude Code, so the team entry price is identical on both sides.
• 9 Oct 2026: Ultrafast priced for GPT-6.1 Sol at $12/$60, Pro 500 only, draining included Codex limits at 8x.
• 1 Oct 2026: GPT-5.3-Codex, GPT-5.1 and GPT-5.4-Nano deprecated with a firm 1 April 2027 shutdown.
• 30 Sep 2026: GPT-6.1 Sol became the Codex default seven days after GPT-6 Sol, halving cached input to $0.10.
• 19 Sep 2026: Hacktron showed a compromised Codex account with an org-wide GitHub connection is a production credential; scope the connector per repository.