Codex Pricing and Usage Limits (July 2026): Free, $20 Plus, $100 Pro, Business

TL;DR

Codex pricing bundles into every ChatGPT plan: Free ($0), Go ($8/mo), Plus ($20/mo), Pro from $100/mo with 5x or 20x rate limits, Business $20/user/mo ($25 monthly), and Enterprise/Edu custom. There is no standalone Codex subscription. Codex usage is capped two ways: a 5-hour rolling window and a weekly cap that stacks on top. On Plus, the 5-hour window is 15-90 messages on GPT-5.6 Sol, 20-110 on Terra, and 50-280 on Luna; Pro multiplies that by 5 or 20.

API-key usage skips both caps and bills per token: gpt-5.6-luna is $1.00/M input and $6.00/M output, Terra is $2.50/$15, Sol is $5/$30. Inside Codex the same tokens meter as credits at roughly 4 cents each. A typical session lands at $0.50-$2.00 in API-equivalent cost.

Codex usage on ChatGPT Plus ($20/mo): 15-90 messages per rolling 5-hour window on GPT-5.6 Sol, 20-110 on Terra, 50-280 on Luna, plus a separate weekly cap. Pro 5x ($100) multiplies the window limits by 5, Pro 20x ($200) by 20. The ranges are wide because limits are credit-based underneath: message cost scales with tokens consumed, and start-of-session context loading counts against the allowance. Credits track API token prices at roughly 4 cents each.

Every Codex Plan

Codex is included in ChatGPT Free, Go, Plus, Pro, Business, and Enterprise. The six tiers as of June 2026:

The other path is an API key with no subscription at all: usage bills at standard API token rates (covered in API Pricing below). You lose the cloud features and integrations, and keep raw model access.

The $100 Pro Tier, Explained

The change that generated the most Reddit threads: Pro no longer means $200. Pro starts at $100/month and you pick a rate-limit multiplier at checkout.

• Pro 5x, $100/mo: 5x the Plus limits. On GPT-5.6 Sol that is 75-450 messages per 5-hour window (Plus is 15-90). On Luna it is 250-1,400 (Plus is 50-280).
• Pro 20x, $200/mo: 20x the Plus limits. 300-1,800 GPT-5.6 Sol messages or 1,000-5,600 Luna messages per 5-hour window.

Per-unit capacity pricing is identical across the two: $100 buys 5 Plus-units of limit, $200 buys 20, so the 20x option is twice the price for four times the capacity. If you hit Pro 5x limits even occasionally, the 20x upgrade is the cheaper marginal capacity.

Codex Usage Limits Explained

Codex enforces two limits at once. A 5-hour rolling window caps messages per model, and a weekly cap resets 7 days after your first message of the week. The two are tracked independently, so a long refactor can drain the 5-hour window while the weekly counter keeps ticking. The published per-window message allowances on GPT-5.6:

The ranges look vague until you see what drives them: limits are enforced in credits, and message cost varies with token consumption. A one-line fix sits at the bottom of a plan's range; an agentic run that reads thirty files and writes ten sits at the top. Cheaper models get more messages because they burn fewer credits per token: Luna at 25 credits/M input versus Sol at 125.

When both windows are exhausted you either wait for the rolling reset (5 hours after the usage that consumed it, or the weekly reset) or buy additional credits from Codex Settings > Usage > Credits, available on Plus and Pro. A June 11, 2026 changelog added “rate-limit reset banking,” a one-time free reset granted at launch for Plus and Pro; it appears in the CLI at the cap or via codex /reset on CLI v0.135+.

How the caps feel in practice

The published 5-hour framing draws steady complaints because real-repo sessions exhaust it long before 5 hours. On the OpenAI developer forum, one Plus user documented the gap in detail:

The mechanism behind the complaint is real: because context-loading at the start of a session counts against the credit allowance, a large repo can spend a meaningful share of a window before the first edit. Community reports note the weekly cap on a $20 Plus plan can drain in about three hours if a single long refactor is triggered. The forward fix most heavy users land on is the same one OpenAI documents, an API key, which removes both the 5-hour and weekly caps and bills per token instead.

Credit Burn Rates Per Model

OpenAI publishes exact credit burn rates per million tokens. This is the table that lets you predict costs instead of guessing:

Cross-reference these against API prices and the conversion falls out: GPT-5.6 Sol costs $5.00/M input at the API and 125 credits in Codex, so 1 credit equals 4 cents. The same ratio holds for every row (Terra: $2.50/M and 62.5 credits; Luna: $1.00/M and 25 credits). Credits are API token pricing in different units.

Practical consequences:

• Switching from Sol to Terra in Codex halves your burn rate for the same token volume. Luna burns a fifth of Sol on input, which is why it gets 50-280 Plus messages per window versus Sol's 15-90.
• Output is 6x the input rate on every model. Tasks that generate a lot of code (test generation, scaffolding) burn credits much faster than tasks that mostly read (debugging, code review).
• Cached input is 10x cheaper than fresh input. Long sessions in one repo benefit; jumping between repos resets the cache advantage. Cache writes on 5.6 bill at 1.25x the uncached input rate.

What a Codex Session Actually Costs

A typical coding session costs $0.50-$2.00 at API rates. Using the burn table above, you can derive per-task costs on GPT-5.6 Sol from token volume:

Token volumes are estimates; the credit arithmetic is exact (input tokens x 125 + output tokens x 750, per million, at 4 cents per credit) for Sol. Run the same task on Luna and the cost drops five-fold on input. Output-heavy work like test generation costs disproportionately more because output burns 750 credits/M versus 125 for input on Sol.

This also explains why two developers on the same plan report wildly different experiences. One running small fixes gets 80 messages per window. One running 300k-token agentic refactors gets 15.

Codex API Pricing

Running Codex with an API key (or calling the models directly) bills per token with no subscription. Current OpenAI API rates:

GPT-5.6 reached general availability on July 9, 2026 after a preview that opened June 25. Sol is priced identically to the previous-gen GPT-5.5 flagship ($5/$30); Terra is the value default at $2.50/$15; Luna is the high-volume tier at $1/$6. Cached-input reads keep the 90% discount, and on 5.6-generation models cache writes bill at 1.25x the uncached input rate.

Codex CLI Pricing

Codex CLI itself is free software. You pay through one of two channels:

• ChatGPT sign-in: included on every plan, even Free ($0/mo) and Go ($8/mo). CLI usage draws from the same 5-hour-window limits as Codex web and IDE, so a heavy CLI session and a heavy web session compete for the same allowance.
• API key: per-token billing at the rates above. No window limits, no cloud features. Costs scale linearly with usage.

The CLI also supports custom model providers via config.toml (base_url, env_key, wire_api), so you can point it at any Responses-compatible endpoint. Setup details in our Codex provider configuration guide.

Business and Enterprise

Business is $20/user/month on annual billing ($25 monthly). Its Codex allowance matches the Plus 5-hour window per seat. Earlier in 2026, Business offered pay-as-you-go Codex seats that billed for actual token usage, but as of June 24, 2026, new pay-as-you-go seats are no longer available (existing seats are unaffected).

What Business and Enterprise add over individual plans: cloud features, GitHub, Slack, and Linear integrations, and admin controls. Enterprise and Edu are custom contracts on top of the same model.

Is Codex Free? What Remains After the Promo

Yes, with caveats. Codex is included in ChatGPT Free at $0/month, including CLI access via ChatGPT sign-in. What changed: the promotional period that boosted free-tier limits has ended, so Free now carries the lowest limits of any plan. There is no separate student price for Codex; students use the same Free, Go, or Plus plans (Enterprise/Edu covers institutions).

If Free limits block you, the upgrade ladder is Go at $8/month, then Plus at $20/month. Plus is the first tier with credit top-ups when you hit a window limit.

Codex vs Claude Code Pricing

The two ladders now mirror each other almost exactly:

At matching sticker prices, the decision moves to capability per dollar. On the API side, GPT-5.6 Sol at $5/$30 is the priciest of the three OpenAI tiers; Terra at $2.50/$15 undercuts Claude Sonnet 4.6 on input while matching it on output; Luna at $1/$6 is the cheapest option across either vendor for high-volume work. Full breakdowns in Claude Code pricing, Codex vs Claude Code, and Claude vs ChatGPT.

When the API Beats a Subscription

At $0.50-$2.00 per session, an API key beats Plus ($20/mo) below 10-40 sessions per month, and beats Pro 5x ($100/mo) below 50-200 sessions. Short sessions push the breakeven up; long agentic sessions pull it down.

The API route fits developers who batch work into occasional deep sessions, teams building products on top of the models, and anyone who wants gpt-5.4-mini at $0.75/$4.50 per million tokens for high-volume simple tasks. What you give up: cloud features, GitHub/Slack/Linear integrations, and the predictable flat monthly bill. Daily users come out ahead on a subscription because window allowances cover far more tokens than the same dollars buy at API rates.

Which Plan Should You Pick

Free or Go ($0-8/mo)

You want to evaluate Codex CLI or use it a few times a week. Free works for evaluation; Go at $8/month is the cheapest plan with usable headroom.

Plus ($20/mo)

You code regularly but not in all-day agentic sprints. 15-90 GPT-5.6 Sol messages per 5-hour window covers most individual workflows, and credit top-ups handle the occasional overage without a plan change.

Pro 5x ($100/mo)

You hit Plus limits weekly. 75-450 GPT-5.6 Sol messages per window (or 250-1,400 on Luna). This is the tier for daily multi-file agentic work.

Pro 20x ($200/mo)

You run parallel agents or sustained refactors. 300-1,800 Sol messages per window. Twice the price of Pro 5x for four times the capacity, so heavy users get better unit economics here.

Business / Enterprise

Your team needs integrations and admin controls. Business is $20/user/month annual; Enterprise adds custom contracts.

API key

Under 10-40 sessions per month, or you need programmatic access. gpt-5.6-luna at $1/$6 per million tokens, no window limits. For uncapped open-weight coding models, DeepSeek V4 Flash on Morph is $0.139/$0.278 per million tokens.

Pricing History

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

If the recurring pain is the cap, not the price, the fix is a per-token API with no 5-hour or weekly window. 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.

DeepSeek V4 Flash at $0.139/$0.278 per million tokens undercuts every Codex model, including Luna at $1/$6, by roughly 7x on input. It 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 honest: you give up OpenAI's cloud features, managed integrations, and the specific GPT-5.6 weights. What you keep is uncapped throughput and per-token billing. See the model lineup, full pricing, and the best open-source LLMs for coding.

Frequently Asked Questions