Codex Pricing 2026: Free, $8 Go, $20 Plus, $100 Pro, Business Plans, Credits, Usage Limits

TL;DR

Codex pricing is your ChatGPT plan: Free $0, Go $8/mo, Plus $20/mo, Pro $100/mo (5x) or $200/mo (20x), Business $20/user/mo annual, Enterprise custom. There is no separate Codex subscription. Plus gets 15-150 GPT-6 Sol messages per 5-hour window. An API key skips the limits and bills GPT-6 Sol at $2/$10 per million tokens.

Usage is metered in credits, and every credit rate equals the model's API price at 4 cents per credit. OpenAI publishes local message allowances per 5-hour period for each model: on Plus, 5-45 on GPT-6 Astra, 15-150 on GPT-6 Sol, and 350-3,000 on GPT-6 Luna. Weekly limits sit on top with no published numbers. Plus has the 5-hour window again since August 25, 2026. Pro does not, for now. New sign-ups to Pro $200 have been paused since September 10. A multi-file refactor costs about 90 cents on GPT-6 Sol, 5 cents on Luna, and $4.50 on Astra.

Codex Plans and Prices

Codex ships inside ChatGPT Free, Go, Plus, Pro, Business, and Enterprise. Six tiers, plus an API key that needs no plan at all:

Plus and Pro users who reach a limit can buy additional credits. Business, Edu, and Enterprise workspaces on flexible pricing buy workspace credits. Free and Go cannot buy credits at all, which makes Plus the real floor for anyone who hits a limit mid-task.

Codex Plus: What $20 Per Month Buys (Plus vs Pro)

There is no product called Codex Plus. The tier is ChatGPT Plus at $20/month, and Codex is included in it. OpenAI's own pitch for the tier is that it will "power a few focused coding sessions each week." That is accurate.

• Local messages per 5 hours: 5-45 on GPT-6 Astra, 15-150 on GPT-6 Sol, 350-3,000 on GPT-6 Luna. Local messages and cloud chats share one allowance.
• Legacy models: 10-100 on GPT-5.6 Sol, 25-200 on GPT-5.6 Terra, 250-2,000 on GPT-5.6 Luna. GPT-5.5 leaves Codex on October 14, 2026.
• Credit top-ups: available. Plus is the cheapest plan that can pay to keep going instead of waiting for a reset.
• 5-hour window: active since August 25, 2026, with a weekly limit on top.

The ranges are wide because the limits are credit-based underneath. A one-line fix sits near the top of the range. An agentic run that reads thirty files and writes ten sits near the bottom. Plus vs Pro comes down to two things: Pro multiplies the allowance, and Pro currently has no 5-hour window at all.

Codex Pro: The $100 and $200 Tiers

Pro starts at $100/month, and you pick a 5x or 20x multiplier. The published GPT-6 ranges are close to, but not exactly, 5x and 20x the Plus numbers:

• Pro 5x, $100/mo: 25-225 GPT-6 Astra local messages per 5 hours, 70-700 on GPT-6 Sol, 1,750-14,000 on GPT-6 Luna.
• Pro 20x, $200/mo: 100-900 on Astra, 300-3,000 on Sol, 7,000-56,000 on Luna.

$100 buys 5 Plus-units of allowance. $200 buys 20. Twice the price for four times the capacity. The catch in September 2026 is availability. On September 10, OpenAI paused new sign-ups and upgrades to Pro $200 from Free, Go, Plus, or Pro $100. Existing Pro $200 subscriptions keep renewing. Pro $100 is unaffected. Per OpenAI's help center, anyone who had Pro $200 on September 10, or lost it in the 30 days before, can return to it once within 30 days of their access ending. If you are on Pro $200 now, think twice before downgrading.

Pro also buys GPT-6 Pro in ChatGPT Chat. GPT-6 Pro is a ChatGPT name for GPT-6 Astra in Pro mode. There is no API model called GPT-6 Pro. It has weekly caps, even on the top plan: 200 per week on Pro $200 and 50 per week on Pro $100, shared with GPT-5.6 Sol Pro. Business Premium seats get 50 per week and Business Standard seats get 15 per month. Plus has no GPT-6 Pro in Chat, only Astra in Codex and Work.

Codex Usage Limits: 5-Hour Window and Weekly Cap

Codex usage is governed by two ceilings. A rolling 5-hour window (where it applies) and a weekly limit. OpenAI publishes estimated local messages per 5-hour period for each model and states that "weekly limits may also apply" with no numbers. The current GPT-6 tables:

GPT-5.6 models are still listed on Plus at 10-100 (Sol), 25-200 (Terra), and 250-2,000 (Luna). GPT-6 Sol gets 50% more messages than GPT-5.6 Sol on the same plan and burns half the credits per token. GPT-5.4 and GPT-5.4 mini left Codex for ChatGPT sign-in on August 31, 2026, and GPT-5.3-Codex-Spark was removed on September 14.

Where the 5-hour window applies today

OpenAI suspended the 5-hour window for Plus, Business, and Pro on July 12, 2026. On August 24, Thibault Sottiaux, who leads Codex at OpenAI, said it would return for Plus the next day. His reasoning, as reported by 9to5Mac: the window smooths compute load so the weekly allowance can stay generous, and casual Plus users were "accidentally" burning a whole week in one session. He also said the 5-hour limit stays off for Pro $100 and Pro $200 "for the upcoming months."

Banked resets and the September reset wave

OpenAI now issues banked Codex resets. A banked reset refreshes both the 5-hour and weekly windows when you choose to spend it, and it moves your weekly reset date. Plus, Pro, and Business accounts got banked resets during the GPT-6 Astra rollout on September 3 and 4, and Sottiaux announced a global reset for paid plans on September 7. If you have a banked reset, spend it at the start of a heavy week, not the end.

Users report quota draining faster than the tables suggest

The published ranges assume average sessions. The openai/codex issue tracker shows what happens at the edges. A Plus user on GPT-6 Astra Medium reported (#42987) the full 5-hour allowance gone in two turns: 100% to 42% in under five minutes, then to 0% in one or two more, on a 165-line diff in a Unity repo. Another Plus user logged (#44233) a weekly quota that dropped from 21% remaining to 0% while idle, with the weekly reset date jumping from September 15 to September 12. A meta tracker (#41220) collects dozens of similar accounting reports and is still open.

The arithmetic explains most of it. Astra burns 250 credits per million input tokens against Sol's 50. On Plus, one Astra message can cost as much as five Sol messages, which is why the Plus Astra range tops out at 45. Default to GPT-6 Sol and switch to Astra for the one hard step. Codex CLI v0.156.0, released September 22, 2026, turns /usage into an analytics dashboard with token totals and plugin and skill activity, so you can see which sessions ate the window. /status has shown estimated thread credits and cost since v0.148.0.

Codex Credits: Burn Rates Per Model

Credits are the unit Codex meters usage in. OpenAI publishes the burn rate per million tokens at Standard speed for every model:

Codex credits to USD: one credit equals 4 cents. OpenAI's pricing page does not print a per-credit dollar price, but every row divides cleanly into the API price list at $0.04. GPT-6 Sol costs $2.00 per million input tokens at the API and 50 credits in Codex. Astra costs $10.00 and 250 credits. Luna costs $0.10 and 2.5 credits. GPT-5.6 Sol costs $4.00 and 100 credits. Put the other way, 25 credits buy one API dollar. Credits are API token pricing in different units, with one difference: Codex credits carry no cache-write charge.

Four consequences follow from the table:

• GPT-6 Luna burns a twentieth of GPT-6 Sol, and GPT-6 Astra burns 5x Sol. That is why Plus gets 350-3,000 Luna messages against 5-45 Astra messages.
• Output burns 5x the input rate on GPT-6 models and 6x on GPT-5.6 Terra and Luna. Test generation and scaffolding drain a window faster than debugging and code review.
• Cached input burns a tenth of fresh input. Long sessions in one repo benefit. Jumping between repos resets the advantage.
• Fast mode (called priority processing until July 30, 2026) burns 2.5x the Standard credit rate on GPT-6 models. On the API, Fast mode costs 2x Standard, so GPT-6 Astra Fast is $20/$100, and both service_tier: "priority" and "fast" still work.

What a Codex Session Costs

The credit arithmetic is exact, so per-task cost falls out of token volume. Input tokens times 50 plus output tokens times 250, per million, at 4 cents per credit, gives the GPT-6 Sol figure:

Token volumes are estimates. The arithmetic is not. The same refactor that cost $2.40 on GPT-5.6 Sol at its July price costs 90 cents on GPT-6 Sol. Model choice is still the largest lever on a Codex bill: Astra costs 100x Luna for the same tokens.

This also explains why two developers on the same plan report different experiences. One running small fixes on Sol gets 150 messages per window. One running 300k-token agentic refactors on Astra gets 5.

Codex API Pricing

Running Codex with an API key, or calling the models directly, bills per token with no subscription. Standard rates from OpenAI's API pricing page as of September 22, 2026:

GPT-6 Sol and Luna cost half of GPT-5.6 Sol and Luna on input. Luna output drops from $1.20 to $0.50. GPT-5.6 Sol's $4/$20 rate is promotional and runs at least through November 21, 2026. Cache writes bill at 1.25x input on the 5.6 and 6 generations ($2.50 per million on GPT-6 Sol). Long-context requests bill at a higher tier: $4/$15 on GPT-6 Sol and $20/$75 on Astra. Fast mode on gpt-5.3-codex costs $3.50/$28.00, double Standard.

Codex CLI Pricing

Codex CLI is free, open-source software. The latest release is v0.156.0 (September 22, 2026). You pay through one of two channels:

• ChatGPT sign-in: included on Plus, Pro, Business, and Enterprise. Free and Go do not include Codex in the CLI. CLI usage draws from the same allowance as the desktop app, web, and IDE, so a heavy CLI session and a heavy app session compete for one budget. Run /usage (an analytics dashboard since v0.156.0) or /status to see token totals and estimated credits.
• API key: per-token billing at the rates above. No 5-hour window and no weekly cap. Cost scales linearly with usage.

The CLI also supports custom model providers through config.toml (base_url, env_key, wire_api), so it can point at any Responses-compatible endpoint. Setup is in our Codex provider configuration guide. One removal to know: codex mcp-server and the standalone codex-mcp-server binary were removed on September 5, 2026. Integrations now go through the app server, which is still experimental. codex mcp, which connects Codex to external MCP servers, still works.

Business and Enterprise

Business Standard seats are $20/user/month on annual billing, or $25/user/month monthly. The per-seat Codex allowance matches Plus: 15-150 GPT-6 Sol local messages per 5 hours, 350-3,000 on Luna.

Premium seats launched in August 2026 at $100/user/month annual or $125 monthly. They carry 5x the Standard usage and no 5-hour limit, so only the weekly reset applies. A workspace can mix Standard and Premium seats and reassign them, which is the cheapest way to give two heavy Codex users on a ten-person team Pro-level headroom. Business, Edu, and Enterprise workspaces on flexible pricing can buy workspace credits. Enterprise and Edu on flexible pricing have no fixed rate limits; usage draws on credits.

Is Codex Free? Free and Go Plan Limits

Yes. Codex is included in ChatGPT Free at $0/month. Free and Go both get GPT-6 Luna at Standard speed in the ChatGPT desktop app, subject to rollout. Neither plan includes Codex in the CLI, so the CLI starts at Plus.

Two caveats. OpenAI publishes no message counts for Free or Go, and neither plan can buy credits, so hitting a limit means waiting for the reset. Go at $8/month is described as suitable for lightweight coding tasks. Codex Go vs Plus comes down to this: Plus at $20/month adds the CLI, web, IDE extension, and iOS, GPT-6 Sol and Astra, a published 15-150 Sol messages per 5 hours, and credit top-ups mid-task.

There is no separate student price for Codex. Students use the same Free, Go, or Plus plans. Edu pricing covers institutions.

Codex vs Claude Code Pricing

As of September 2026 the two ladders match almost price for price, at both the subscription and the API level:

At matching sticker prices the decision comes down to capability per dollar and how each vendor meters. The cheapest model on either side is GPT-6 Luna at $0.10/$0.50. Anthropic has no model priced near it; its cheapest current model is Haiku 4.5 at $1/$5. Both vendors moved on September 22: OpenAI shipped GPT-6 Sol and Luna, and Anthropic shipped Claude Opus 5.5 and said it is raising 5-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. Full breakdowns are in Claude Code pricing, Codex vs Claude Code, and Claude vs ChatGPT.

Codex Subscription vs API Key

A Codex subscription buys a flat monthly bill and a hard ceiling. An API key buys no ceiling and a variable bill. The crossover depends on the model.

• On GPT-6 Sol: sessions cost $0.15 to $0.90, so Plus ($20/mo) breaks even at roughly 22 to 133 sessions a month and Pro 5x ($100/mo) at roughly 110 to 670.
• On GPT-6 Luna: sessions cost under a cent to 5 cents. Plus does not break even until several hundred sessions a month.
• On GPT-6 Astra: sessions cost $0.75 to $4.50. Plus breaks even at 4 to 27 sessions, which is why Astra users feel the subscription limits first.

The GPT-6 price cut moved the crossover. The API route fits developers who batch work into occasional deep sessions, teams building products on the models, and anyone running Luna for high-volume simple work. What you give up is cloud features, integrations, and the flat bill. What you gain is no 5-hour window and no weekly cap.

Which Codex Plan to Pick

Free or Go ($0-$8/mo)

You want to evaluate Codex or use it a few times a week. Both run GPT-6 Luna in the desktop app only, with no CLI. Neither can buy credits when a limit lands.

Plus ($20/mo)

You code regularly but not in all-day agentic sprints. 15-150 GPT-6 Sol messages per 5 hours covers most individual workflows. Expect the 5-hour window to bite on long sessions. Credit top-ups handle the occasional overage.

Pro 5x ($100/mo)

You hit Plus limits weekly. 70-700 Sol messages per 5 hours, and no 5-hour window for now. This is the only Pro tier open to new subscribers as of September 22, 2026.

Pro 20x ($200/mo)

You run parallel agents or lean on Astra. 300-3,000 Sol or 100-900 Astra messages per 5 hours. New sign-ups are paused. If you already have it, keep it.

Business or Enterprise

Your team needs admin controls and shared billing. Standard seats at $20/user/month annual match Plus. Premium seats at $100 give heavy users 5x and drop the 5-hour window. Enterprise adds custom contracts and credit-based usage.

API key

You run fewer than 20-odd Sol sessions a month, you want Luna at $0.10/$0.50 for bulk work, or you need programmatic access with no window limits.

Pricing History

Skip the Usage Cap: Pay-Per-Token Open Models

If the recurring pain is the cap and not the price, the fix is a per-token API with no 5-hour window and no weekly cap. Morph serves open-weight coding models on custom kernels through an OpenAI-compatible endpoint at https://api.morphllm.com/v1. You pay for the tokens you use, and nothing throttles a long refactor mid-session.

Each of these swaps into an existing Codex CLI or any OpenAI-SDK client by pointing base_url at the endpoint above and sending a Morph API key (setup in our Codex provider configuration guide). The tradeoff is plain. You give up OpenAI's cloud features, managed integrations, and the GPT-6 weights. You keep uncapped throughput and per-token billing. See the model lineup, full pricing, and the best open-source LLMs for coding.

Frequently Asked Questions