Is Codex free? What OpenAI's Free plan does not say

What Codex costs

Codex is one agent with several front ends: a CLI, an IDE extension, a desktop app and a cloud version. This post is about the CLI. Its README states that the repository "is licensed under the Apache-2.0 License", and it signs in two ways: with a ChatGPT account, spending the plan's allowance, or with an OpenAI API key, paying per token.

The README recommends signing in "to use Codex as part of your Plus, Pro, Business, Edu, or Enterprise plan". OpenAI's Codex pricing page goes further down the ladder and adds two cheaper tiers below Plus. Everything below comes from that page and the README, both re-read on 29 September 2026.

In order, the descriptions form a ladder: quick tasks, lightweight tasks, a few focused sessions a week, then a plan sized to your usage. Free and Go name GPT-6 Luna in the desktop app, and the CLI is first named on the Plus card. Plus is the plan most people mean when they say Codex, and OpenAI describes it as enough for "a few focused coding sessions each week".

What the free plan does not say

OpenAI publishes usage estimates for Codex, and its table starts at Plus. There is no column for Free or for Go. The Free plan is priced at $0 and linked from the page, and its allowance is stated nowhere on it. Pro has no column either, for the opposite reason: "Pro plans currently have no five-hour limit."

The published estimates are counted in local messages per five-hour period, and OpenAI attaches warnings to them. Above the table: "These estimates are not fixed message limits". Below it: "Local messages and cloud chats share your plan's usage allowance. Weekly limits may also apply."

A message is what you send. What it costs depends on what the agent does next, and OpenAI says so on the same page: "Tasks that look similar can consume different amounts of your allowance." It adds that "prompt length alone isn't a reliable estimate." That is why the ranges run to more than tenfold: an allowance counted in anything other than tokens cannot tell you how much work it holds, as our post on request counts shows.

The page is also quiet on what happens when the Free plan runs out. It says that Plus and Pro users "can purchase additional credits to continue working", and that "All users may also run extra local chats using an API key". For a free account, then, the route past the limit is the API, paid per token. Codex's free plan fits the pattern our post on reset windows found: paid plans document their limits better than free tiers do.

What Clixad does differently

Clixad does the same job in the terminal, editing files and running commands, and it asks before either unless you switch that off. The difference is the bill.

Tokens are the unit. Every model call is metered from its real token count at a published rate, 100,000 credits to one US dollar of list value. The price of each model is public before you pick it, and a small job costs a small amount.

Five cheap models are free every day. GPT-5 Nano, Gemini 2.5 Flash-Lite, Claude Haiku 5.5, DeepSeek V4 Flash and Gemini 3.5 Flash-Lite cost nothing up to 60,000 credits of usage per account per day. That allowance sits inside a daily budget shared by everybody, so on a busy day the free turns can run out sooner. Free usage adds and deducts nothing. Past either limit those models cost credits again until the next day.

The rest of the catalogue runs on credits. Thirteen models from five vendors sit on one balance, and /model switches between them mid-session. OpenAI's two are GPT-5 Nano at the cheap end and GPT-6 Sol, which the Plus card names. A new account starts with 25,000 credits. After that you earn credits by completing an advertiser action on an offerwall page in a browser tab, or you buy them; the smallest pack is $3 for 280,000 credits.

Most of OpenAI's newest models are missing. GPT-6 Sol is in our catalogue, which was last read from OpenRouter on 27 September 2026. GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra, the other three models in OpenAI's Plus estimates, are not.

Earning is uneven. Credits arrive on completion. The most common outcome of starting a survey is a screenout: a few questions in, the survey decides you are not the respondent it wants and ends with nothing credited, or with a small bonus. From our own ledger, completed offers on our CPX wall have credited from 9,000 to 114,000 credits on completion, with a median of 51,000: 26 completions recorded in our ledger between 27 July 2026 and 12 September 2026, as of 14 September 2026. The ledger does not record how long an offer took.

The session shows ads. While a model call runs, one line above the input box can show a sponsored message marked Ad, and up to two more places on screen can show one marked the same way. Nothing you type takes part in choosing them, and CLIXAD_SPONSOR=0 stops every request for them being made at all. Codex shows no advertising, and that counts in its favour.

Your instruction file carries over. Codex reads project instructions from AGENTS.md, which its pricing page mentions in its advice on making limits last. Clixad reads the same file, next to its own CLIXAD.md, so a repository set up for Codex needs no new one.

One agent task, priced

A plan price leaves out what you get, so this blog prices one fixed workload: a single instruction that an agent turns into twelve model calls, starting at 12,000 input tokens and resending the whole conversation every round. The shape comes from the tier-limits post.

GPT-6 Sol sits outside the free allowance. The 25,000-credit signup bonus does not cover a task on it, and a full day at the earning cap buys 4.

The free part is the top two rows: 19 whole tasks a day on GPT-5 Nano and 10 on Gemini 2.5 Flash-Lite, with no earning at all. Next to what Plus estimates for GPT-6 Luna, that is small.

Who Clixad fits

Someone who codes in bursts. Credits do not expire, so a quiet week costs nothing and the balance is still there the week after. A subscription bills the quiet week at the price of the busy one.

Someone who wants a price before the task. Codex's free plan publishes no allowance and its paid plans publish ranges. Our table gives one number per model for a defined task, and the wallet is checked before a request starts.

Someone who wants several vendors on one balance. Claude, GPT, Gemini, DeepSeek and Kimi models sit behind the same /model picker, priced per token. Trying Claude Sonnet 5.5 on a problem GPT-6 Sol fumbled costs that call and nothing more.

Someone without a card to put on file. The account is a GitHub login, and neither the free allowance nor the offerwall asks for payment details.

Who should pay for Codex

Anyone for whom Codex is the daily tool. Plus costs $20 a month, its card lists "GPT-6 Sol and GPT-6 Luna", and OpenAI estimates 350 to 3,000 local messages every five hours on GPT-6 Luna. Pro starts at $100 a month and has no five-hour limit at all.

We cannot convert those messages into our tasks, and OpenAI warns against converting them into anything. The gap is wide enough that it hardly matters. The bottom of OpenAI's Plus estimate for GPT-6 Sol is 15 messages in five hours, and our whole daily earning cap covers 4 twelve-call tasks on the same GPT-6 Sol. The units differ, and the gap is still too wide.

Buying credits does not change the picture for full-time use: bought credits pay for tokens at provider cost with a 1.5 markup, which the funding post sets out, where a subscription is one flat price.

Codex also does things we do not. Its Plus plan includes "Cloud-based integrations like automatic code review and Slack integration", and Clixad has neither. Most of OpenAI's newest models are in Codex and not here. If those matter to how you work, stay with Codex and pay for Plus or Pro.

Other free agents exist, and they are built differently from ours. FreeBuff grants a daily allowance of hours on its own model picker. If your work leans on Anthropic's models, the Claude Code comparison runs the same numbers against Anthropic's plans.

FAQ