How to pay for Claude Code and Codex: plans, seats or the API

I use Claude Code and Codex in Conductor. For someone doing most of their coding in Claude and using Codex alongside it, my starting recommendation after DevDay is Claude Max 20x at $200 and ChatGPT Pro 100 at $100. If you rarely hit the limits of a cheaper plan, keep it. There is no saving in buying capacity you won't use.

The difficult part is comparing what those prices buy. A subscription includes a usage allowance; the API bills tokens. Neither a plan's advertised multiplier nor a model's token price tells you how many useful tasks it will finish. Choose the models that work for you, then use your limit history and any extra charges to choose the plan.

I'd choose according to how the agents fit into the work:

How you use agents	What I'd pay for
A few times a week	Compare metered API use with a $20 plan, including whether you want the chat app
Most days	A $20 plan; compare top-ups with an upgrade if limits repeatedly interrupt work
Heavy use, with Claude doing most of the work	Claude Max 20x plus ChatGPT Pro 100, then reassess from actual usage
Buying for a team	Standard seats to start; Premium where limits or extra charges justify it
A company needing enterprise controls	Compare the contract, seat fees and metered usage together
Products and shared bots	API billing
Team CI	Approved business automation or API billing

Prices are US web prices before tax, checked on 17 October 2026. Annual prices below are monthly equivalents, paid for a year upfront. The update includes Claude's monthly API credits and a correction to Sonnet 5.5's cached-input price and worked example.

When a subscription is better value#

A subscription is attractive when the work it covers would cost more through the API. It can also include product features, such as hosted tasks, that an API key does not buy. ccusage can estimate the API value of recorded usage, but that figure is not a bill or a guaranteed saving. Check that it recognises your models and cache tokens, and include any credits you bought on top of the subscription.

Run that comparison over a representative week, including quiet days. Then check when the allowance ran out. A plan that covers your monthly volume can still interrupt work if too much of it lands in one five-hour window. Conversely, a large API-equivalent total does not make an expensive plan worthwhile if a cheaper one already covers the same work.

For occasional use, API billing may be cheaper than a subscription. Check which controls actually stop spending: a budget alert alone is not a hard cap. For products serving other people, use API billing. Team automation needs its own check: use an approved business automation feature or API credentials, with an owner and a budget. Anthropic's legal and compliance page distinguishes individual Claude Code and Agent SDK use from routing other users' requests through consumer subscriptions.

What the plan multipliers actually buy#

Plan	Price a month	Advertised allowance multiple
Claude Pro	$20, or $17 billed yearly	1x
Claude Max 5x	$100	5x
Claude Max 20x	$200	20x
ChatGPT Plus	$20	1x
ChatGPT Pro 100	$100	5x at announcement
ChatGPT Pro 200	$200	10x at announcement
ChatGPT Pro 500	$500	25x at announcement

Both companies price their higher tiers as multiples of the $20 plan. Anthropic's Max plans include 5 or 20 times Pro's usage in each five-hour session, shared between the Claude apps and Claude Code. OpenAI's September announcement put Pro 100 at 5x Plus and the new Pro 200 at 10x; its DevDay recap gives Pro 500 as 25x. The current Pro help page confirms the prices and tier order without publishing those ratios, so treat them as the announced comparison, not a guaranteed task count.

Against monthly Claude Pro, Max 20x has the strongest advertised session allowance per dollar among Claude's personal plans: ten times the price for up to twenty times the session usage. Max 5x scales price and advertised usage together. On those announced ratios, OpenAI's Pro tiers work out at about $20 per multiple of Plus. Annual Claude Pro is also cheaper than paying monthly.

Those are comparisons within each provider. Claude's 20x and ChatGPT's 10x have different baselines, model costs and limits. They do not mean Claude delivers twice as much work. Claude's advertised session multiples also do not establish the same ratios for weekly capacity.

Compare the price of an upgrade with the extra usage it would actually replace. Moving from Claude Max 5x to Max 20x costs another $100 a month. If you regularly spend more than that on ordinary extra usage, and the larger plan would cover it, upgrading is worth considering. The same price difference separates ChatGPT Pro 100 and Pro 200. Occasional busy weeks may still be cheaper to cover with credits.

Claude's usage credits use API rates, with bundle discounts available. OpenAI has a separate credit rate card: purchase prices depend on the plan or agreement, and Codex credit billing has no separate cache-write charge. Use the price shown in your account instead of converting a promotional credit grant into a universal exchange rate.

Check model access as well as the allowance before upgrading. Limits can be specific to a model or feature. Claude Max has both five-hour and weekly limits; OpenAI's Codex pricing page currently lists no five-hour limit for Pro. A weekly cap determines sustained capacity, while a five-hour cap can stop a busy afternoon with room left in the weekly meter.

Claude's included API credits#

Max and Team now include monthly credits for the Claude API: $100 on Max 5x and $200 on Max 20x. Team contributes $20 per Standard seat or $100 per Premium seat to a shared pool capped at $500 a month.

New subscribers become eligible after seven days and must link a Console organisation to claim them. Credits expire each billing cycle. They cover eligible API and Agent SDK work, including self-run claude -p with the linked organisation's API key, but exclude interactive Claude Code and extra subscription usage. Runs launched by the Claude Code GitHub Action, IDE or desktop app are also excluded.

If you already pay for eligible API work, count the credit you will actually use when comparing plans. It could offset a substantial bill. If you only use interactive Claude Code, it does not extend your allowance.

If you pay per token, read the cached column#

These are standard-speed API prices per million tokens for a selection of coding models, from Anthropic's pricing docs and OpenAI's pricing docs.

Model	Input	Cached input	Output
Claude Fable 5.1	$10	$0.25	$50
Claude Opus 5.5	$4	$0.20	$20
Claude Sonnet 5.5	$2	$0.10	$10
GPT-6 Astra	$10	$1	$50
GPT-6.1 Sol	$2	$0.10	$10

For the models listed here, API cache writes cost 1.25 times ordinary input; Anthropic's one-hour writes cost twice ordinary input. Long-context pricing can change the bill too. OpenAI's GPT-6 Astra and GPT-6.1 Sol rates rise above 272,000 input tokens, while the listed Claude models retain their standard per-token rates across the context window.

For long agent sessions, cached input can be a substantial part of the bill. A typical agent call includes the accumulated context: its instructions, the files it has read, the output of every command and its own earlier replies. With prompt caching, the part the model has already seen is charged at the cached rate, reducing the cost of repeatedly processing that context. Anthropic's Opus 5.5 announcement says cache reads "make up the majority of agentic and coding work costs".

For a simple rate comparison, assume 10 million cached input tokens, 200,000 cache-write tokens at the 1.25x rate and 50,000 billable output tokens, including any billed reasoning. Assume no cache expiry, retries, tool fees or long-context surcharge. These are fixed quantities chosen to show the arithmetic, not a measured task or a forecast of what each model will consume.

Model	Reads	Writes	Output	Total
Opus 5.5	$2	$1	$1	$4
Sonnet 5.5	$1	$0.50	$0.50	$2
GPT-6.1 Sol	$1	$0.50	$0.50	$2
Fable 5.1	$2.50	$2.50	$2.50	$7.50
GPT-6 Astra	$10	$2.50	$2.50	$15

In this example, Opus costs twice as much as Sonnet, Sonnet and Sol tie, and Astra costs twice as much as Fable. Those results follow from holding token counts fixed. They do not establish the cheapest model for a real task.

Real measurements can reverse that ranking. On 17 October, Artificial Analysis reported a weighted cost per Intelligence Index task of $3.26 for GPT-6 Astra at max effort and $5.98 for Opus 5.5 at max effort with default fallback. Opus scored higher on that index, 58 against 53. This is a mixed evaluation suite, not a forecast of your coding bill, but it shows why token prices alone are insufficient.

Try a few representative tasks with the models you are considering. Count retries, time spent steering and whether the result passes review. A cheaper model that needs repeated correction can cost more to finish; a more expensive model per token can finish with fewer tokens. GPT-6.1 Sol deserves that comparison alongside Opus, rather than treating every Codex task as an Astra task.

Make your allowance last#

Model choice, reasoning effort and unnecessary retries all affect cost. Keeping reusable context cached helps too. Claude Code's prompt caching guide is worth reading in full. These are the parts I'd pass on:

• Avoid needless model switches in long Claude Code sessions. Each model has its own cache, but a switch can still be worth the cost when another model is better suited to the work.
• Leave fast mode off unless the time saved justifies the cost. In Claude Code, enabling it at the start avoids rebuilding a long context at fast-mode rates, and it uses paid credits rather than your allowance. Codex Fast consumes included usage at 2.5x; Astra Ultrafast at 8x.
• Start fresh for unrelated tasks. Compaction rebuilds the cache, while Claude Code's /rewind can reuse a still-warm prefix. Neither saving tokens nor preserving a cache is worth keeping a conversation that has become confused.
• For API keys or usage credits, Claude Code normally uses a five-minute cache. Setting promptCacheTtl to 1h costs more to write but can avoid rewriting context after a break. Subscription usage normally gets an hour for the main conversation.
• Keep what every request carries small. OpenAI's advice for Codex is to keep AGENTS.md short and disable MCP servers you aren't using, because each one adds context to every message. The same goes for CLAUDE.md and MCP servers in Claude Code.
• Check the meters every week. In Claude Code, /usage shows your plan's bars, what used them (skills, subagents and MCP servers), and a flag when cache misses or long context account for 10% or more of recent use. In Codex, /status shows your remaining limits.

Running more than one plan#

If one provider's limits interrupt work, a plan with the other can add useful capacity and access to a second model family. Try it on real tasks before committing to two large subscriptions. Moving a task across also takes a handover, so it helps to leave clear notes and committed work.

For separate Claude accounts, Claude Code supports distinct configuration directories through CLAUDE_CONFIG_DIR. Its authentication docs describe aliases such as claude-work. cswap adds saved logins and a usage dashboard. Its README warns against running /logout before switching, because that may revoke the saved login. Use separate accounts when you need separate billing or workspaces; measure demand before buying another subscription.

For Codex, Orca is another option to investigate for managing accounts alongside parallel agents. Account switching and session behaviour vary by tool and version, so check how it handles existing sessions before relying on it for a workday. Keep work and personal credentials separate, and review any tool that stores them.

Anthropic's Thariq Shihipar said in February that having multiple Max accounts is allowed under its terms, while reselling tokens is not. That is a useful clarification if you're considering a second Claude subscription, though it does not explicitly approve automated rotation when an allowance runs out.

OpenAI also supports multiple accounts with different subscription plans. Its terms prohibit circumventing rate limits, so I wouldn't treat a second account as permission to get around a Codex cap. Account-switching tools make it possible to move between accounts; they don't settle what the provider permits.

What I'd pick after DevDay#

The $300 combination makes sense when Claude is your main coding tool and Codex gets enough work to justify more than Plus. Max 20x offers a better advertised session allowance per dollar than Max 5x; Pro 100 gives you room to assess Codex without starting at $200 or $500. If Codex is your main tool, put the larger budget there instead.

Pro 200 and Pro 500 buy more capacity. The announced ratios imply the same price per multiple of Plus. That can be worth paying for if it avoids interruptions or repeated top-ups. Pro 500 also adds Ultrafast for Astra and Sol. Its 8x allowance consumption makes it expensive to use throughout the day, but faster responses can be valuable when you are waiting on the agent. Judge that by time saved on completed work, not tokens per second alone.

Check /status in Codex and its usage dashboard over a few representative weeks. If Pro 100 keeps interrupting useful work, compare Pro 200 with what you spent on credits. If it has plenty left, consider whether Plus would be enough.

Existing Pro 200 subscribers should check their transition email before changing plans. Eligible accounts keep the old allowance through 29 October while the subscription remains active. Confirm whether the credit grant survives a downgrade or cancellation, and whether you expect to use it before expiry. Its headline $2,500 value is not cash saved: retaining a $200 plan instead of a $100 one for November and December costs another $200. Keep it only if the usable credits and extra capacity justify that cost. Moving from $200 to $100 saves $1,200 over a full year.

Buying for a team#

Seat	Monthly	Billed yearly	Advertised allowance multiple
Claude Team Standard	$25	$20	1.25x
Claude Team Premium	$125	$100	6.25x
ChatGPT Business Standard	$25	$20	1x
ChatGPT Business Premium	$125	$100	5x

Using monthly Pro and Plus as the baselines, the advertised multiples put annual Claude Team seats at 25% more allowance per dollar and annual ChatGPT Business seats at the same rate. These comparisons do not include the value of workspace administration or contractual data protections, and they do not guarantee equivalent weekly capacity.

Start with the seat that fits each person's usage, then upgrade people whose limits or extra charges justify it. Daily use alone does not tell you whether they need Premium. Model access can matter: Fable runs on paid usage credits for Claude Team Standard, while Premium includes it within a cap of half the seat's weekly allowance. Annual billing reduces the seat price, but monthly billing can be sensible while you learn how many seats and which tiers you need.

Claude Team is for 2 to 150 people, and every seat includes Claude Code. ChatGPT Business needs at least two seats, and its Premium seats have no five-hour limit. When someone reaches their seat's limit, Claude Team owners can buy usage credits at API rates, and ChatGPT Business members draw on a shared pool of credits the workspace has bought.

For new usage-based Claude Enterprise plans, the seat fee buys access and usage is billed separately. Self-serve plans require at least 20 seats. There is no included token allowance; admins can set organisation and individual spending limits. Existing seat-based contracts can differ. ChatGPT Enterprise is priced by contract, with credit-based or token-based billing. Its token-based rate card lists the API prices and leaves discounts to each agreement. Uneven usage alone does not establish that Enterprise is cheaper. Compare the full contract with Team, including seat fees, metered usage and the controls your organisation needs. Claude Team's 150-person ceiling is another reason to consider Enterprise.

For company code, use an account and configuration your organisation has approved. Business plans provide different contractual protections and administration from consumer accounts; turning off training is only one part of that decision. Anthropic lets Free, Pro and Max users choose whether their chats and Claude Code sessions train its models, and OpenAI may train on personal ChatGPT accounts unless you opt out. Neither trains on business plans or the API by default. If someone on your team uses a personal plan for work, check that setting together, and whether the company is happy with it.

Recheck after a model or plan change. Use ccusage for an API-cost estimate, the provider's meter for remaining allowance, and your invoices for actual spending. If your work gets done without interruptions and the cheaper plan covers it, keep the cheaper plan.