Codex Pricing in 2026: ChatGPT Plans, Usage Limits, Credits and API-Key Costs

OpenAI's Help Center says Codex is included across ChatGPT plans, including Free and Go, with usage limits that vary by plan. Codex Cloud (tasks that run in OpenAI's sandbox instead of on your machine) is available to eligible Plus, all Pro tiers, Business, Enterprise, Healthcare and Education accounts, and is not included with Free or Go.

Pro is three plans, not one. The ChatGPT pricing page lists Pro 100, Pro 200 and Pro 500, and each carries monthly API credits of $100, $200 or $500 respectively. The footnote says activation and eligibility requirements apply, the first grant is prorated, unused credits expire at the end of the billing cycle, and the benefit can change.

Business and Enterprise. Business, Enterprise and Edu are billed through a separate credit-based rate card, and workspaces on flexible pricing can buy more workspace credits instead of hitting a hard stop. Enterprise and Edu without flexible pricing get the same per-seat limits as Plus for most features. We do not quote Business seat prices here; see chatgpt.com/pricing for your region.

Shared allowance. OpenAI notes that ChatGPT Work and Codex share usage: Work inside ChatGPT uses the same pricing, credits and limits as Codex, so a long Work session leaves less room for coding the same day.

Codex limits are rolling windows, not a monthly quota. OpenAI publishes ranges for local messages per five-hour period on Plus and Standard Business seats, and stresses that cloud tasks may use more, that weekly limits may also apply, and that the numbers are estimates rather than fixed caps:

Why the range is so wide. OpenAI says model choice, context, reasoning, tool use, retrieval and caching all change what a message costs, so prompt length alone is not a reliable estimate. A small script may use a fraction of the allowance; a long session that holds a big context uses much more per message.

Speed modes multiply usage. Fast mode draws included subscription usage 2.5x faster than Standard (2x on purchased credits). GPT-6 Astra Ultrafast and GPT-6.1 Sol Ultrafast draw 8x on included usage and 6x on credits. Image generation uses included limits about 3 to 5x faster on average.

When you hit the limit mid-task. OpenAI says the agent can finish the turn already in progress, subject to fair use. After that, Plus and Pro users can buy credits; anyone can switch to an API key and keep working locally at standard API rates.

Where to see your remaining limit. The usage dashboard at chatgpt.com/codex/settings/usage shows current limits and reset times, and /status inside a Codex CLI session shows what is left.

Credits are the pay-as-you-go unit for ChatGPT plans. Your included usage is spent first; after that, eligible usage draws from your credit balance, and one balance covers Codex, ChatGPT Work, Word, Excel and PowerPoint. Plus and Pro users buy them from Settings > Usage in ChatGPT or Usage & Billing in the Codex app, and automatic reload is available on some accounts.

Codex credit usage is token-based. These are OpenAI's Standard-speed rates, in credits per 1M tokens, from the Codex pricing page and the ChatGPT rate card:

What a task costs in credits. OpenAI's own example: a typical GPT-5.6 Sol task may use 2 to 15 credits. There is no separate cache-write charge on Codex credit billing, and GPT-6.1 Sol drawn from a subscription has no long-context surcharge in Codex.

What OpenAI does not print. None of the pages we read gives a single dollar price per credit; the Codex page says credit purchase prices and discounts depend on your plan or agreement. Read the price at checkout before you turn on automatic reload, and note that a balance can go negative if concurrent tasks finish after it hits zero.

GPT-5.6 Sol promotion. OpenAI says GPT-5.6 Sol's promotional pricing runs at least through November 21, 2026, and applies to usage paid with purchased credits; included plan usage and the five-hour and weekly limits are unchanged.

Signing in to Codex with an API key switches the meter to OpenAI API pricing. The Codex pricing page lists what you get: Codex in the CLI, SDK or IDE extension, model availability that follows your key, and no cloud-based features (no Codex cloud, GitHub code review or Slack). OpenAI recommends this route for automation in shared environments such as CI.

These are Standard rates (short context) from OpenAI's API pricing page. Batch and Flex are half of Standard (gpt-6.1-sol $1.00 / $5.00), Fast is double ($4.00 / $20.00), and regional data-residency endpoints carry a 10% uplift for models released on or after March 5, 2026.

Plan or API key? A subscription is a flat ceiling: good when a person codes interactively and the five-hour windows rarely bite. An API key is a meter: good for scheduled or CI jobs, for bursts that would exhaust a window, and for teams that want spend per project. OpenAI itself warns not to use API token prices to estimate how many tasks a plan includes, because subscription usage is metered differently.

Model retirements differ by sign-in. GPT-5.5 was retired from Codex with ChatGPT sign-in on October 14, 2026, and GPT-5.4 and GPT-5.4 mini are no longer available there, but OpenAI notes that API access and Codex sessions on your own API key are not affected.

People searching for "openai api credits" or "chatgpt api credits" usually mean one of three things, and they are not interchangeable:

ChatGPT usage credits. The credits described above. They extend Codex, Work and Office add-ins after your plan limit and are separate from your ChatGPT wallet (a US-dollar balance that can hold gift-card funds). They do not pay for API calls.

Monthly API credits on Pro. Pro 100, Pro 200 and Pro 500 include $100, $200 or $500 of monthly API credits "to build your own apps and agents", per the ChatGPT pricing page. Unused amounts expire at the end of the billing cycle. If you are on Pro, this is the balance an API-key Codex session or your own agent would draw from first.

API billing. Everything else on an API key is billed at the per-token prices in the previous section, plus tool fees where you use them (OpenAI lists web search at $10.00 per 1,000 calls plus search content tokens, and hosted containers from $0.03 per 20-minute session).

The practical rule. Interactive Codex on a plan spends ChatGPT usage and then ChatGPT credits. Anything you script against the API, including codex exec on an API key and an agent that reads tweets, spends API credits. Our OpenAI API key guide covers creating and scoping the key.

A common Codex or OpenAI agent job is "read what people posted about X today and summarize it". Two bills add up: the X data and the model tokens. We measured both on real posts instead of guessing: 1,000 recent posts from @OpenAIDevs and @OpenAI, tokenized with o200k_base, the encoding OpenAI's current models use.

The tweet object is heavy. A tweet as the API returns it (text, author block, entities, media, counts) has a median of 2,304 tokens (p90 3,905, max 8,590). The same tweet compacted to one line (handle, time, likes, reposts, views, text) has a median of 77 tokens (p90 108). A full 20-tweet page is 50,058 tokens raw versus 1,584 compacted, about 31.6x fewer.

The estimate. We priced a run as 1,500 instruction tokens plus the tweets in, 800 tokens out, at Standard uncached API rates. These are assumptions you should replace with your own prompt; the per-tweet token counts are measured.

Takeaway. Compacting tweets before they reach the model matters more than which plan you are on: on gpt-6.1-sol it turns a $0.51 run into a $0.028 run. With compacted input, the model costs about twice the data; with raw JSON, it costs over 30x the data.

Projected over a month. One 100-tweet run a day costs about $0.45 of X data plus $0.83 on gpt-6.1-sol (compacted), or $0.04 on gpt-6-luna. Hourly runs (720 a month) cost $10.80 of data plus $19.99 on gpt-6.1-sol; pasting raw JSON instead would push the model share to $370.14.

Inside Codex on a plan. If the agent runs as a Codex task on a ChatGPT plan, the same token volume draws on your five-hour window instead of dollars, and OpenAI's own tip applies: every MCP server you enable adds context to each message, so disable the ones a job does not need. The X data cost is the same either way.

To read X posts as structured data (text, author, counts), an agent needs an X data source. Two priced options:

Per 1,000 tweets that is $5 on X's pay-per-usage rate card (docs.x.com/x-api/getting-started/pricing) against $0.15, about 33x. X also gives $20 of credits on a first eligible card and matches the first auto-recharge up to $50.

How TwitterAPI.io bills. Tweet endpoints such as /twitter/tweet/advanced_search and /twitter/user/last_tweets cost 15 credits per tweet returned ($1 = 100,000 credits), with a minimum of 15 credits per request: a full 20-tweet page is $0.003, and a search that finds nothing costs the $0.00015 minimum. That floor matters for agents that poll: checking every 60 seconds for a month is 43,200 checks, $6.48 if every one comes back empty; every 5 minutes, $1.30. You pay only for the data returned, no X account is needed for reads, and purchased credits never expire.

Plans and rate limits. Pay as you go, or subscribe (Starter $29, Builder $99, Pro $199, Scale $499, Business $1,499 a month; details on the pricing page). Accounts run at 0.2 requests per second before the first payment and 20 once paid. New accounts that pass the automatic bot check get up to $1 of free credits (at most 100,000 credits, no card, one per person or network, expiring after 31 days).

If the agent should post. Writes are a separate product: one /twitter/user_login_v2 login ($0.10, with a proxy you supply) returns login_cookies, then each /twitter/create_tweet_v2 post is 60 credits ($0.0006), billed per HTTP 200 call. The call acts as that account, so X can rate-limit, lock or suspend it; keep a human pace, and expect the session to need a re-login now and then.

Question. How many tokens does a real page of tweets occupy when handed to an OpenAI model, and therefore what does an X-reading agent run cost? As a side question: how much of OpenAI's own posting is about Codex, and how often those posts mention pricing or limits?

Accounts and method. Two public organization accounts, @OpenAIDevs and @OpenAI, read with /twitter/user/last_tweets (newest first, replies excluded), 25 pages each: 50 requests, 1,000 posts, $0.15 of data. Codex posts are those whose text, quoted text or card mentions "codex"; pricing posts are Codex posts mentioning prices, plans, credits, limits or costs. Tokens were counted with tiktoken o200k_base.

Window and sample. The most recent 500 posts of each account at collection time, collected in 2026: about 104 days for @OpenAIDevs (4.8 posts a day) and about 532 days for @OpenAI (0.9 a day).

What it shows. Codex is close to a third of @OpenAIDevs' timeline, and Codex originals get slightly higher median views and likes than the rest on both accounts (@OpenAI: 6,026 vs 4,619 median likes). Only 36 of 229 Codex posts touch pricing or limits, so check the pricing page itself rather than the timeline for price changes.

What it cannot show. Token counts depend on the tokenizer and on what you put in each line; a prompt with long instructions or tool schemas shifts the totals. Org-account posts are longer and richer in media than an average search result, so a keyword search may run lighter. The model prices are estimates from list prices, not invoices; we did not call an OpenAI model in this study.

Privacy. Aggregates only. Both accounts are organizations; reposted and quoted authors were not analyzed, raw pages stayed off the repository, and the committed sample keeps only de-identified counts.

Compact before you prompt. Send one line per tweet, not the API object; on our sample that is about 31.6x fewer input tokens.

Pick the model per step. Classification and filtering run fine on gpt-6-luna ($0.10 / $0.50 per 1M tokens); keep gpt-6.1-sol or gpt-5.3-codex for the summary or the code change.

Fetch only new tweets. Use a since_time: cursor in /twitter/tweet/advanced_search so an hourly run pays for new posts plus the $0.00015 minimum when nothing changed.

Trim Codex context. Keep AGENTS.md short and enable only the MCP servers a job uses, as OpenAI's own usage tips recommend.

Use Batch for overnight digests. OpenAI prices Batch at half of Standard, which suits a nightly summary that does not need an instant answer.

Start building. Create a key on the dashboard, run the snippet below to measure your own prompt, and read the endpoint reference at docs.twitterapi.io. The Codex MCP guide shows how to give Codex the same X data as tools.