Codex pricing follows ChatGPT plans: Free and Go ($8 a month) include limited access, Plus at $20 a month includes Codex on the web, in the CLI and in the IDE extension, and Pro comes in three tiers at $100, $200 and $500 a month (OpenAI, 14 November 2026). Business seats cost $20 per user a month billed annually. On an API key there is no plan, only per-token charges.
OpenAI uses the name Codex for a family of surfaces: a command-line agent, an IDE extension, a desktop and mobile app, and a cloud agent that works on tasks in a hosted environment and can review pull requests on GitHub. One allowance covers all of them, and OpenAI’s pricing page opens with the reminder that ChatGPT Work and Codex share the same pricing, credits and usage limits. There are two ways to pay: sign in with a ChatGPT plan and work within its limits, or sign in with an API key and pay for every token.
Figures on this page come from OpenAI’s Codex pricing documentation, its API price list and four Help Center articles, all consulted on 14 November 2026. The company-wide question of what a ChatGPT workspace costs per seat, and what the top plan adds in controls, is handled in our ChatGPT seat-cost comparison; here the subject is the coding agent.
OpenAI Codex pricing by plan
| Plan (OpenAI, US dollars, verified 14 November 2026) | Price | Who it is for | Codex included | Limits and admin features |
|---|---|---|---|---|
| Free | $0 | Exploring quick tasks | GPT-6 Luna at Standard speed in the desktop app, subject to rollout | Limited; no admin features |
| Go | $8 a month | Lightweight tasks | Same as Free | Limited; no admin features |
| Plus | $20 a month | A few focused sessions each week | Web, CLI, IDE extension and iOS; automatic code review and Slack integration; GPT-6.1 Sol and GPT-6 Luna | Five-hour limit; weekly limits may apply; credits available |
| Pro 100, Pro 200, Pro 500 | $100, $200 or $500 a month, monthly billing only | Daily, long-running use | Everything in Plus | No five-hour limit at present; Ultrafast mode on Pro 500 only |
| Business, Standard seat | $20 per user a month billed annually; $25 billed monthly; two seats minimum | Small and growing teams | Codex and ChatGPT on desktop and mobile; larger virtual machines for cloud chats | Per-seat limits as on Plus; workspace credits; SAML SSO, MFA, no training on business data by default |
| Business, Premium seat | $100 per user a month billed annually; $125 billed monthly | Heavier users in the same workspace | As Standard, with higher limits | No five-hour limit; seat types can be mixed |
| Enterprise and Edu | Contact sales | Whole organisations | Everything in Business, with priority request processing | SCIM, EKM, RBAC, user analytics, audit logs through the Compliance API, data retention and residency controls |
| API key | No plan; pay per token | Automation and shared environments such as CI | CLI, SDK and IDE extension only; no cloud features | Limits and controls of your API organisation |
Two notes on the model names. GPT-5.5 was retired from Codex for ChatGPT sign-ins on 14 October 2026, so guides that still recommend it are out of date. And for Enterprise and Edu workspaces, GPT-6.1 Sol is off until an administrator enables it.
Is Codex free?
Only at the edges. The Free and Go plans list Codex with one model, GPT-6 Luna, at Standard speed, in the desktop app, and OpenAI adds “subject to rollout”. Its plan comparison calls this “limited Codex access”. Image generation inside Codex is not available on Free.
The full set of surfaces begins with Plus. If you sign in to the CLI with an API key instead, nothing is free: there is no included allowance, and each request is charged at API rates. Eligible Free and Go accounts can buy credits to go past their limits, but OpenAI says that option is available only to a limited group.
Codex limits per plan: the five-hour window and weekly caps
OpenAI is unusual in publishing message estimates. For Plus and for Standard Business seats, its pricing page gives these ranges of local messages per five-hour period:
| Model (OpenAI estimates, verified 14 November 2026) | Local messages per five hours, Plus and Standard Business |
|---|---|
| GPT-6 Luna | 350 to 3,000 |
| GPT-6.1 Terra | 35 to 285 |
| GPT-6.1 Sol | 15 to 160 |
| GPT-6 Sol | 15 to 150 |
| GPT-6 Astra | 5 to 45 |
The ranges are wide for a reason. OpenAI states that they are not fixed message limits, and that usage depends on the model, the context processed, reasoning, tool use, retrieval and caching, so prompt length alone is a poor guide. Further rules from the same page:
- Weekly limits may also apply on top of the five-hour window.
- Local messages and cloud chats draw on one allowance, and cloud tasks may use more of it than local messages.
- Pro plans currently have no five-hour limit.
- Speed modes multiply consumption. Fast mode uses included limits at 2.5 times the Standard rate. Ultrafast, for GPT-6 Astra and GPT-6.1 Sol, uses them at 8 times.
- Image generation uses included limits three to five times faster than a similar turn without it.
- Code review through GitHub is counted as its own category; reviews run locally count toward general limits.
- A turn in progress is allowed to finish when you hit a limit, subject to fair-use limits.
Current limits and reset times are in the usage dashboard, and /status shows the remaining allowance during a CLI session. OpenAI Support will not reset a limit.
Codex Pro vs Plus
| What changes (OpenAI, verified 14 November 2026) | Plus | Pro 100 | Pro 200 | Pro 500 |
|---|---|---|---|---|
| Monthly price | $20 | $100 | $200 | $500 |
| Annual billing | Not stated | No | No | No |
| Five-hour limit | Yes | None at present | None at present | None at present |
| Included Work and Codex usage | Baseline | Higher than Plus | More than Pro 100 | Highest of the three |
| Ultrafast mode | No | No | No | Yes |
| Credits after the allowance | Yes | Yes | Yes | Yes |
OpenAI does not put numbers on the Pro allowances. Its Help Center says only that Pro 200 includes more usage than Pro 100 and that Pro 500 offers the most, and that a larger allowance “does not guarantee twice as many completed tasks”. The practical difference between Plus and Pro is the five-hour window: a Plus user on GPT-6.1 Sol can exhaust 15 to 160 messages in an afternoon and must then wait or pay, while a Pro user is bounded by the longer-period allowance.
Ultrafast is expensive even where it is included. It burns the included allowance at 8 times the Standard rate, and once the allowance is gone it is charged at 6 times the Standard credit rate. Buying credits on Pro 100 or Pro 200 does not add it.
Codex CLI pricing with an API key
Signing in to the CLI, SDK or IDE extension with an API key removes plans from the picture. OpenAI’s Codex page says model availability then follows the API models open to your key, that cloud features such as GitHub code review and Slack are not available, and that usage is billed at API prices.
| Model (OpenAI API list, Standard speed, verified 14 November 2026) | Input, per million tokens | Cached input | Output |
|---|---|---|---|
| gpt-6-astra | $10.00 | $1.00 | $50.00 |
| gpt-6.1-sol | $2.00 | $0.10 | $10.00 |
| gpt-6.1-terra | $1.00 | $0.05 | $5.00 |
| gpt-6-luna | $0.10 | $0.01 | $0.50 |
Regional processing endpoints carry a 10% uplift on recent models. OpenAI warns against using API prices to estimate what a subscription includes; the two are separate.
To see how much the model choice matters, take one agent task as an illustration: 2 million input tokens, of which 1.5 million are cache hits, and 100,000 output tokens.
- GPT-6.1 Sol: 0.5 × $2 + 1.5 × $0.10 + 0.1 × $10 = $2.15.
- GPT-6 Luna: 0.5 × $0.10 + 1.5 × $0.01 + 0.1 × $0.50 = $0.115.
- GPT-6 Astra: 0.5 × $10 + 1.5 × $1 + 0.1 × $50 = $11.50.
The same work costs a hundred times more on the largest model than on the smallest. The token counts are ours, chosen for round numbers; measure your own with a pilot.
Codex credits: what happens after the limit
When a plan’s allowance is used up, credits keep Codex running. Included usage is always consumed first.
| Model (credits per million tokens at Standard speed, verified 14 November 2026) | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra | 250 | 25 | 1,250 |
| GPT-6.1 Sol | 50 | 2.5 | 250 |
| GPT-6.1 Terra | 25 | 1.25 | 125 |
| GPT-6 Sol | 50 | 5 | 250 |
| GPT-6 Luna | 2.5 | 0.25 | 12.5 |
How credits behave, according to the Help Center:
- Who can buy them. Plus and Pro users, from Settings, Usage in ChatGPT or Usage and Billing in the Codex app. Business workspace owners buy a pool that all members share once their seat limits are reached.
- What they cost. OpenAI says purchase prices and discounts depend on your plan or agreement, and the pages we read give rates in credits, not dollars. Each credit rate is 25 times the API dollar price for the same model (50 credits against $2 for GPT-6.1 Sol input), so the illustration task above comes to 53.75 credits.
- Speed modes. With purchased credits, Fast mode costs twice the Standard rate and Ultrafast six times.
- Expiry. Credits last 12 months from purchase and do not roll over; credits bought in Japan after 13 November 2026 expire after 180 days.
- Automatic reload. Optional, with a minimum balance, a target balance and a monthly cap. A saved instruction such as “ask me before using credits” does not stop credits being spent.
- Refunds. Non-refundable unless law or contract requires it.
OpenAI’s own example is that a typical GPT-5.6 Sol task may use 2 to 15 credits.
Codex on Business and Enterprise workspaces
A Business workspace needs at least two seats, Standard or Premium in any mix. Standard seats have the same published five-hour ranges as Plus; Premium seats have higher limits and no five-hour window. Credits bought by the workspace owner sit in one pool, and when no credits are left a member who has run out is blocked until the owner adds more. Workspaces that held a Codex-only seat before 24 June 2026 keep that legacy seat type, which carries Codex usage only and draws on workspace credits from first use.
Enterprise and Edu are sold by quote, and OpenAI publishes no price. What it does publish is the mechanics. With credit-based flexible pricing there are no fixed rate limits: usage scales with a shared credit pool bought at contract level, and the workspace overage limit starts at “No limit” until an owner sets one. Workspaces without flexible pricing have the same per-seat limits as Plus for most features. Some Enterprise agreements are billed in US dollars per token at contracted rates instead of credits. The seat caps and controls that separate the two workspace plans are in the seat-cost comparison linked in the introduction.
Codex cost for 10, 50 and 200 developers
List prices over 12 months, before tax, credits and any negotiated discount.
| Route (OpenAI list, verified 14 November 2026) | Per developer per year | 10 developers | 50 developers | 200 developers |
|---|---|---|---|---|
| Business, Standard seats, billed annually | 12 × $20 = $240 | $2,400 | $12,000 | $48,000 |
| Business, Standard seats, billed monthly | 12 × $25 = $300 | $3,000 | $15,000 | $60,000 |
| Business, Premium seats, billed annually | 12 × $100 = $1,200 | $12,000 | $60,000 | $240,000 |
| Business, four Standard seats to one Premium, annual | Mixed | 8 × $240 + 2 × $1,200 = $4,320 | 40 × $240 + 10 × $1,200 = $21,600 | 160 × $240 + 40 × $1,200 = $86,400 |
| Personal Plus plans, expensed | 12 × $20 = $240 | $2,400 | $12,000 | $48,000 |
| Personal Pro 200 plans, expensed | 12 × $200 = $2,400 | $24,000 | $120,000 | $480,000 |
| Enterprise | Quote | Quote | Quote | Quote |
| API keys | No seat fee | Tokens only | Tokens only | Tokens only |
Reading the table:
- Plus and a Standard Business seat on annual billing cost the same $240 a year. The Business seat adds SAML SSO, a no-training default and a shared credit pool, so expensing personal plans saves nothing and loses the controls.
- The seat price buys a limit, not a volume. A Standard seat is capped by the five-hour ranges above. Developers who run agents all day will either need Premium seats or spend credits.
- API keys suit jobs, not people. CI pipelines and scripted runs have no seat to attach to, and per-token billing is the honest measure for them.
The cost lines a Codex seat does not show
- Credits after the allowance, with automatic reload if someone switched it on.
- Premium model multipliers. GPT-6 Astra is five times the credit rate of GPT-6.1 Sol and a hundred times that of GPT-6 Luna.
- Speed modes. Fast at 2.5 times and Ultrafast at 8 times the included allowance.
- Seats that sit idle. A Premium seat held by a light user is $1,200 a year for limits nobody reaches.
- API fallback. OpenAI notes that any user can run extra local chats on an API key at standard rates, which is a second bill outside the workspace.
- Shared allowance. ChatGPT Work, and on personal plans ChatGPT for Excel, draw on the same limits as Codex.
- Voice in the desktop app, at $0.05 a minute from the Codex usage budget.
The same breakdown for neighbouring tools: Claude Code’s subscription and API routes, Cursor’s new plan structure and GitHub Copilot’s per-seat credits.
When running models in-house is cheaper than Codex, and when it is not
Codex can be configured to talk to a model server you run yourself, which takes OpenAI’s limits and token charges out of the loop for that work. The configuration is in our guide to running Codex against a local model; this section is only about the money.
A local model is likely to cost less when:
- Developers keep running out. If a team is buying Premium seats or reloading credits every month, the routine share of that work is a candidate for a model on your own hardware.
- The jobs are automated and repetitive. Batch refactors, test generation and CI checks are metered per token on an API key, and they do not need the largest model.
- Capable hardware is already in place and shared. The cost per request falls as more people use the same server.
- The repository cannot be sent to a cloud service, in which case price is not the deciding factor.
It is likely to cost more when:
- Usage fits inside a $240 seat. Light and moderate users are cheap to serve on a plan.
- You would buy hardware for a handful of developers. A server that idles most of the day never earns its cost back.
- The task needs the strongest model. Open-weight coding models differ from GPT-6.1 Sol and GPT-6 Astra in quality, and failed attempts are paid for in engineering time.
- You rely on the cloud agent. Hosted tasks, GitHub code review and the Slack integration are part of the service and have no local equivalent in the CLI.
No hardware figures appear here, by design: they change too fast and depend on your starting point. Work through the method in our TCO guide for on-premises AI platforms and the page on flat-rate versus token pricing, then pick a model from the open-weight coding model comparison. Other terminal agents and editors that work with local models are covered in our survey of on-prem code assistants.
How VDF AI fits
VDF AI does not run, resell or host Codex and has no integration with it. Where it is relevant is the step after an individual developer’s set-up: a company deciding where the model runs, which models are allowed, which repositories each team can reach, and what evidence remains for an audit.
For that, VDF AI offers VDF Code, a coding assistant deployed in your VPC, in your own data centre or fully air-gapped. It covers inline completions, chat inside the IDE, multi-file edit mode and pull request review, with a policy engine to allow or block models and an audit log of prompts and completions. On-premises licensing is per seat and does not meter model tokens. See the pricing page for how that is structured.
Sources
OpenAI’s documentation, API price list and Help Center, consulted on 14 November 2026.
- OpenAI: Codex pricing
- OpenAI: API pricing
- OpenAI: ChatGPT plans and pricing
- OpenAI Help Center: about ChatGPT Pro tiers
- OpenAI Help Center: using credits on personal plans
- OpenAI Help Center: flexible pricing for Enterprise, Edu and Business
- OpenAI Help Center: rate card for Business and Enterprise credit-based pricing
- OpenAI Help Center: ChatGPT Business overview
Budgeting for coding agents across a team? Look at VDF Code for a deployment on your own infrastructure, or get in touch to discuss your requirements.