Keep your options open when buying AI
Give your best people room to use better tools. Keep the commitments small enough to change course.
In office hours this week, a business owner asked whether it was time to upgrade to an enterprise AI account. The company already had a team plan and a few premium seats it could assign where needed. More people were using AI, and the costs were getting harder to understand.
My advice was to give the people doing useful work a company card, a monthly budget, and permission to choose their tools. They should be able to use a better model when it arrives, without waiting for the whole company to change plans.1
Our starting assumption is that AI models and tools will keep changing quickly. A purchasing decision should leave room for the best choice to change within months.
Give people a card, a budget, and room to choose
Start with the people already finding useful ways to apply AI. On the call, I suggested roughly 10% of the company as a starting point. They might work in accounting, operations, or sales.1
For people who build with AI all day, I would skip the $100 tier and buy the $200 plan. The larger allowance is the more economical choice for heavy use. More important, I want our strongest people using AI often enough to become competent quickly. A usage cap makes them stop and ask whether a task is worth spending their remaining allowance on.
For our best forward-deployed engineers, the people building tools inside a client's business, I want room to experiment and keep working through problems. The $200 plans provide much more of that room, though they still have limits. Other employees should work within smaller budgets and show useful results before getting more capacity.
Give each person an approved monthly budget and a company card. Within that amount, let them add, replace, or cancel monthly subscriptions without a new purchase request each time. A $400 budget could cover two $200 plans, or one heavy-use plan plus money to try a newly released tool. The budget is a ceiling, not a spending target.
Write down which company data those accounts may use. For example, permit public research and drafts from approved company documents; keep payroll files, patient records, and client material restricted by contract out until the specific service is approved. Require a receipt and a short note naming the tool and what it is being used for.
Review subscriptions once a month and move money from unused tools to better ones. A $20 plan is a reasonable starting point for everyone who does not yet need heavy usage.
What the different plans buy
OpenAI makes ChatGPT and Codex. Anthropic makes Claude and Claude Code. Individual plans buy access for one person. Company plans add a shared workspace and administration; higher tiers can also change usage and features.2
| Plan type | OpenAI | Anthropic |
|---|---|---|
| $20 individual | ChatGPT Plus $20/month. The base allowance used below. | Claude Pro $20/month or $200/year. The base session allowance used below. |
| $100 individual | ChatGPT Pro 5× $100/month. 5× Plus usage. | Claude Max 5× $100/month. 5× Pro session usage. |
| $200 individual | ChatGPT Pro 20× $200/month for existing subscribers. 20× Plus usage. New purchases currently paused. | Claude Max 20× $200/month. 20× Pro session usage. |
| Company workspace | ChatGPT Business Standard: $25/month. Premium: $125/month. Premium has 5× Standard usage for Work/Codex and no five-hour limit. Two seats minimum. | Claude Team Standard: $25/month, 1.25× Pro session usage. Premium: $125/month, 6.25× Pro session usage. Two members minimum. |
| Enterprise | ChatGPT Enterprise Terms and pricing by agreement. Relevant for requirements such as regulated workspaces and centralized retention policies. | Claude Enterprise Terms and pricing depend on the offer. Adds controls such as domain claims and custom retention periods. |
My choice as of September 17, 2026
If I could choose one plan today, it would be ChatGPT Pro at $200 for use with Codex. My current assessment is that OpenAI has the stronger lineup, especially for everyday work: GPT-5.6 Sol is substantially better than Claude Opus 5. Claude Fable is competitive with GPT-6 Astra at the top end. When that changes, I want to be able to switch.4
Company plans also pay for administration. At both providers, annual Standard and Premium rates are $20 and $100 per seat per month, with a year-long commitment. The individual $100 and $200 tiers are monthly.2
An API lets custom software call a model directly. Its billing is separate. A subscription's advertised usage multiplier is not a dollar balance you can spend on arbitrary API calls.6
Can a company use personal $200 accounts?
Yes, for work the company permits. We found no published rule requiring a business to switch plans once it reaches a certain number of employees.
Each person needs their own account. Both providers prohibit sharing one login across a team. The two-seat minimum on company plans tells you how those plans are sold; it does not force a two-person business to buy one.7
A company email and card do not turn an individual plan into a managed company account. They put the address and payment under company control. They do not give an administrator the account's chat history, centrally enforced privacy settings, or a reliable way to revoke every session.
Can the same work email have Team or Business and an individual plan?
Yes, generally. In Claude, select the personal account from the account menu, then use Settings → Billing → Upgrade to Max. It is billed separately from the Team seat. In ChatGPT, select the personal workspace and use Settings → My Plan. A Business seat can coexist with a personal Pro subscription; the current $200 purchase pause still applies. Keep the personal account separate if that is the arrangement you want.8
Does disabling the company email cut off access?
It can block a new sign-in. It does not reliably end existing access. Claude's email login requires a fresh link sent to that mailbox. Disabling the Google identity used for Google sign-in blocks that route too. But a person already signed in may still have an active session. ChatGPT can also use a separate password or passkey, so mailbox access alone is not a dependable control.9
Before someone leaves, have them save company work to the shared folder, revoke their AI sessions and coding-tool authorizations, and cancel subscriptions the company will stop funding. IT should also revoke their access to connected company services and managed devices. If IT must revoke access centrally without the employee's cooperation, use a company workspace with managed sign-in.10
Turn off training before using company material
On an individual plan, make these settings part of setup. Opting out of training does not delete the work you have saved. Company plans exclude business data from model training by default, but also retain work under their applicable policies.11
Settings for ChatGPT, Codex, and Claude
- ChatGPT and Codex: In ChatGPT, open Settings → Data Controls and turn off Improve the model for everyone. This covers new personal-plan Codex tasks too. Separately, open Codex Settings and turn off training on full environments; the ChatGPT toggle does not change that setting.12
- Claude and Claude Code: In the personal Claude account, open Settings → Privacy and turn off Help Improve our AI models. It covers chats and coding sessions using that account.13
What stays stored? Saved chats remain until deleted. OpenAI's Codex-chat policy schedules deleted chats for removal within 30 days, with exceptions. Anthropic also states 30 days after chat deletion; opting into training can put data in its training systems for up to five years. Open the note for safety, feedback, and legal exceptions.14
Patient records need a different setup. An ordinary Pro or Max subscription does not provide a Business Associate Agreement for HIPAA-covered work. OpenAI lists ChatGPT for Healthcare and Enterprise with Regulated Workspace among its eligible products. For Claude, the Primary Owner must activate HIPAA compliance in a HIPAA-ready Enterprise organization and accept the BAA. Claude Code has additional zero-retention requirements, and Cowork is excluded.15
An illustrative 20-person budget
For work approved for individual accounts. Monthly subscriptions throughout.
18 people on a $20 plan
2 operators with a $400 monthly budget each
Total monthly budget, before tax
Each operator's $400 budget can cover $200 Claude Max and $200 ChatGPT Pro, if they already have Pro or new sign-ups reopen. Until then, a new Codex subscriber can use $100 Pro as a temporary fallback, bringing actual company spending to $960. If all 20 people also need company administration, 20 monthly Standard seats cost $500; adding the two $400 operator budgets makes a $1,300 ceiling.2
Keep the work in files you control
The other point from the call was practical: files and folders make it easier to use a different tool without starting over. For people working in Claude Code or Codex, keep reusable instructions, source documents, and finished work in company-controlled working folders. Let either tool use the same approved material.
Company work/
Marketing/
Instructions.txt
Source documents/
Drafts/
Finished work/
Give each tool instructions to read that material. Put important decisions and working preferences there, rather than leaving the only copy in one app's memory or chat history. Tool permissions, connections, and app-specific projects may still need separate setup. A shared folder makes the work more portable without promising that every account feature will transfer.1
Keep the purchasing commitment small
A monthly business plan can preserve flexibility. If you need central billing and access controls, pay for them monthly. Avoid treating an annual discount as the automatic next step.
At current Standard rates, 20 company seats cost $500 a month. Annual billing costs $4,800 for the year, compared with $6,000 across twelve monthly payments. Paying an extra $100 a month keeps the next eleven months uncommitted. Removing a user does not necessarily remove a paid seat from an existing term.16
Before accepting credits or an enterprise offer, ask for the cost of leaving after three months in writing. Include any remaining seat payments, minimum spending, and notice deadline. A $10,000 credit attached to a $30,000 required spend still leaves $20,000 to pay, even if another tool becomes better.
If a client requires an enterprise agreement, buy the seats needed for that work and keep a separate monthly budget for other approved tools. One contract should not use up the company's entire AI budget.
Sources & notes
Tap a source number in the article for details. Provider documentation checked September 17, 2026.
The recommendation from office hours
These recommendations draw on Sam Gaddis's September 17, 2026 WSC Applied AI for Executives office hours. The discussion favored capable operators, permission to use different providers, short purchasing commitments, and company-controlled files. The roughly 10% starting point is Sam's suggested approach, not a measured adoption benchmark. The $400 allowance and 20-person budget are illustrations.
↩︎ 1 ↩︎ 2 ↩︎ 3Plan prices and company seats
Prices are U.S. web list prices before tax, checked September 17, 2026. Individual plans cover one user. Company Standard and Premium seats are $25 and $125 on monthly billing, or $20 and $100 per month with annual billing. Both company plans start at two seats. OpenAI's Premium multiplier applies to Work/Codex. Claude's multipliers describe session usage; weekly limits also apply.
The budget examples use monthly prices. Eighteen $20 plans plus two $300 tool selections cost $960. Giving those two people $400 ceilings makes the budget $1,160. The company-workspace alternative replaces the 18 individual base plans with 20 Standard seats: $500 + $800 = $1,300.
↩︎ 1 ↩︎ 2 ↩︎ 3How we calculated allowance per dollar
Both providers advertise 5× their respective base allowance at $100 and 20× at $200. The larger plan therefore offers four times the advertised allowance at twice the price: (20 ÷ 200) ÷ (5 ÷ 100) = 2. This is our calculation from the published multipliers, not a provider promise about completed work.
OpenAI compares Pro with Plus. Anthropic compares Max with Pro's session allowance. Claude also has weekly limits; model and feature limits can differ. These are not identical units across providers, guaranteed token counts, or a promise that every weekly limit scales fourfold. The higher tier is better value only when the added capacity is useful.
↩︎ 1Sam's current model preference
These are Sam Gaddis's model preferences as of September 17, 2026, for heavy individual use where a personal account is permitted. The links below identify the models and their product context; the comparative judgments are Sam's. Model-specific and other usage limits still apply.
↩︎ 1The temporary pause on $200 ChatGPT Pro
OpenAI paused new Pro 20× subscriptions and upgrades on September 10, 2026. Existing subscriptions continue. It has not given a reopening date in this notice.
A scheduled cancellation or downgrade can be undone before the billing cycle ends. After it takes effect, the $200 tier cannot be repurchased while the pause remains. Joining a Business workspace does not itself require cancelling the separate personal subscription.
↩︎ 1Subscriptions, coding agents, and API billing
Calling a provider's general API with an API key is billed separately from ChatGPT or Claude subscriptions. Signing into Codex or Claude Code with a supported subscription is a different billing route.
Anthropic's current Agent SDK notice says subscription-authenticated SDK, claude -p, and third-party app usage still draws on subscription limits. Its proposed separate monthly SDK credits were paused on June 15. Do not count the old credit table lower on that page as an available benefit.
↩︎ 1Individual accounts used for work
Neither set of published consumer terms reviewed for this note states that reaching a particular employee count forces a business onto a company plan. This is a finding from the reviewed terms, not an assurance that every type of work is permitted. Both prohibit sharing account credentials or providing another person access to your account.
Each employee needs their own login. Company policy, client obligations, and the provider's usage rules still apply. Paying for individual accounts does not create a shared administrative workspace.
↩︎ 1One work email, separate personal and company access
Claude documents keeping a personal Free, Pro, or Max account alongside a Team or Enterprise organization under the same email. Switch accounts from the profile menu. To upgrade the personal account, open its Settings → Billing → Upgrade to Max. This buys a separate subscription; it does not convert a Team seat into Max.
OpenAI documents separate personal and Business workspaces under one login. Choose the personal workspace before buying Pro. Business administrators cannot require personal-workspace merging through Business onboarding.
Managed domains are an exception. Claude Enterprise domain claims can require a work-email personal account to move into the organization. OpenAI Enterprise or Edu onboarding can also require migration. If the personal option is unavailable, check the company's domain policy with IT before purchasing.
↩︎ 1Why disabling email is not a complete access control
Claude supports email sign-in links and Google sign-in, with no dedicated Claude password. A disabled mailbox prevents receiving a new sign-in link. That does not prove existing Claude sessions have ended: its web sessions expire after 28 days of inactivity and are refreshed while active.
ChatGPT supports passwords and passkeys. A passkey can authenticate without an email verification code, subject to the account's sign-in configuration. Company single sign-on is a separate case controlled through the identity provider.
Our conclusion from these documented login methods: control of the email inbox is useful, but it is not equivalent to revoking the AI account's sessions. A company credit card controls payment, not authentication.
↩︎ 1Sessions and coding access need separate attention
Claude: on the web, Settings → Account → Log Out signs out browser, desktop, and mobile sessions. Claude Code tokens are separate: open Settings → Claude Code and use the trash icon for each authorization.
ChatGPT: Settings → Security → Active sessions provides session management and Log out all. Other sessions can take up to 30 minutes to end. That screen does not cover Codex CLI or third-party apps, so it is not proof that all coding access has been revoked. Single-sign-on accounts use their organization's identity controls.
For a company-funded individual account, record the subscriptions and connected services before the employee leaves. Have the employee revoke the applicable authorizations and hand over company files; IT should separately remove access to company drives, repositories, identity accounts, and managed devices.
↩︎ 1Company plans and training defaults
OpenAI's business products and Anthropic's commercial products exclude customer inputs and outputs from model training by default. Voluntarily submitted feedback or explicit opt-in programs can have different terms.
This policy concerns training. It does not mean chat history is never stored, that deleted data disappears immediately, or that every connected third-party service follows the same policy. Individual privacy toggles also do not supply a business contract or centralized administration.
↩︎ 1Turn off training for ChatGPT and Codex
In a personal ChatGPT account, open Settings → Data Controls and switch off Improve the model for everyone. OpenAI says this applies to new ChatGPT conversations and Codex tasks across that account.
Codex also has a separate setting for training on full environments. Open Codex Settings and turn that off as well; changing the ChatGPT toggle does not change it. Check both when setting up the account for work.
Opting out leaves saved history in place. Temporary Chat is excluded from training and is scheduled for deletion within 30 days, but it is a separate mode from an ordinary saved conversation.
↩︎ 1Turn off training for Claude and Claude Code
In the personal Claude account, open Settings → Privacy and turn off Help Improve our AI models. Anthropic says the control covers chats and coding sessions on Free, Pro, and Max.
After opting out, new and previously stored chats or coding sessions will not be used for future model training under that setting. It does not reverse training already completed or in progress. Safety-flagged conversations can still be used to improve trust-and-safety systems.
↩︎ 1Training off does not mean zero retention
OpenAI: saved Codex chats remain until deleted. After deletion, they disappear from the account and are scheduled for system deletion within 30 days. Exceptions include data already de-identified and disassociated from the account, or data retained for safety, security, or legal obligations. Archiving is not deletion.
Anthropic: deleted chats leave history immediately and are removed from back-end systems within 30 days. If training is enabled, applicable de-identified data can remain in training systems for up to five years. Turning the setting off does not undo completed or ongoing training.
Anthropic lists longer exceptions: usage-policy-flagged inputs and outputs up to two years; related classification scores up to seven years; feedback data up to five years. Legal obligations and disputes can also extend retention. A 30-day deletion policy does not mean every saved conversation expires 30 days after creation.
↩︎ 1Specific requirements for HIPAA-covered work
OpenAI lists ChatGPT for Healthcare and Enterprise with Regulated Workspace among its HIPAA-eligible products. Coverage requires the applicable Business Associate Agreement and eligible features. Its list excludes Codex cloud, even though specified local Codex access can be covered.
Anthropic requires a HIPAA-ready Enterprise organization. Its Primary Owner activates HIPAA compliance and accepts the BAA under Organization settings → Data and privacy. An ordinary Enterprise label alone is insufficient.
Claude Code coverage is limited to specified CLI and local desktop usage with zero data retention configured. Claude Code on the web and remote sessions are excluded. Claude Cowork is excluded. A personal Pro or Max subscription cannot substitute for these arrangements.
↩︎ 1Monthly flexibility and annual seat commitments
At $25 per month, 20 Standard company seats cost $500. At $20 per month billed annually, they cost $4,800 for a year. The annual saving is $1,200 compared with twelve monthly payments. Our $100 monthly flexibility cost is that difference divided by twelve.
Removing an employee and reducing the subscription's paid seats are different actions. Reductions can take effect only at renewal, depending on the plan and billing cycle. Read the actual seat terms before assuming unused seats will be refunded.
The $10,000 credit and $30,000 spending commitment in the article are a hypothetical contract example, not a current offer from either provider.
↩︎ 1
September 17, 2026. Editorial images generated with GPT Image 2.5.