Checked on November 6, 2026. This is a prepurchase exercise readers can carry out themselves, rather than a performance evaluation we have completed across providers. The materials, function and tasks are fictional or sanitized samples. No real accounts, paid model requests or transaction data were used in testing.
Start with the work, rather than the highest product tier. A useful purchase decision answers what you must deliver, where it will run, which account will use it, what permissions it needs, and what evidence will establish that it works. Fifteen minutes is a practice budget for narrowing the question, not a model response time, delivery deadline or verification SLA. Leave insufficient evidence marked “pending”; there is no need to place an order just to finish the exercise.
1. Write a requirements card: turn “more capable” into a checkable result
Choose one representative task from work that recently stalled: turning two pages into a verifiable brief, for example, or reviewing a small function. Do not upload customer lists, a complete private repository, cookies, API keys, payment records or identity verification links. Replace names, domains, paths and numbers while preserving the structure that affects the decision. If you cannot sanitize the material reliably, use the fictional examples below.
The card should say what you need and what you do not need. Someone writing only in a browser should not buy a different entry point because a desktop automation demonstration looks attractive. Someone maintaining local files cannot assume the whole workflow is available because a chat window can answer questions. Specify the input, artifact, reviewer and prohibited actions so that later price comparisons refer to the same job.
Sanitized requirements card
Task: turn the materials I provide into a verifiable deliverable.
Required entry point: browser / desktop / local development tool / API (select only what is actually needed).
Input: fictional text or sanitized files confirmed as suitable for use.
Output: format, length, fields to retain and supporting evidence.
Permissions: read-only; do not send, pay, delete, publish or modify external systems.
Acceptance conditions: list 3 conditions I can check independently.
Existing account and data that must be retained: sanitized description only.
Unresolved: workspace, tool permissions, allowance and delivery type.
Only organize missing information and acceptance steps. Do not invent available permissions or purchase promises.Keep this card in your own notes first, rather than immediately sending it to a seller or model. You can share a summary without sensitive information and ask support to confirm the product type and order conditions. Submit account verification materials only through the channel required by the actual order; do not put them in public comments or these prompts. The card is not a login credential and does not establish eligibility for a top-up.
Live product cards and delivery conditions on this site · Current FAQ and support conditions on this site
2. The 15-minute route: try one representative task and keep evidence
The timings below are an operating sequence designed for this article, not measured product speeds. You can start with an existing account, free entry point or already authorized tool. When you encounter a payment or permission requirement, record the gap. You do not need to buy several providers’ products to complete this exercise.
Use the same route after delivery, but keep separate records for the prepurchase and post-delivery checks. Mark workspace, model entry point and expiry information you have not seen before buying as unknown. After delivery, add evidence against the order one item at a time. Do not record capabilities you hope to receive as capabilities you already possess.
- 0–2 minutes: complete a sanitized requirements card and choose one task
- 2–5 minutes: copy one prompt and try an already authorized entry point
- 5–8 minutes: manually check the artifact, citations and boundaries
- 8–11 minutes: inspect account, workspace, entry point and permissions
- 11–13 minutes: compare delivery categories and conditions in the live catalog
- 13–15 minutes: save observations and write retest and ordering conditions
3. Two copyable tasks: inspect the evidence as well as the wording
Pick the task closest to your work. Writing and research users can try the conflicting-sources exercise; developers can try the function-boundary exercise. These check fidelity to sources, adherence to constraints and diagnosis, rather than produce an overall ranking of intelligence. Use the same input across products and preserve the first result. Do not compare a result corrected over several rounds in one product with a single initial answer in another.
All three sources in the first exercise are fictional. An acceptable answer recognizes that B is newer than A, explains that the stated trial length changed, and leaves C unverified. “Newest” does not establish a fact that can never change, and an unsourced free-forever comment must not become policy. Check the three supplied passages; do not ask the model to find web evidence for these fictional products.
Write a brief using only the fictional materials below. Do not browse or add policy.
A: A draft dated 2026-11-01 says practice product X has a 7-day trial.
B: A revision dated 2026-11-04 says practice product X now has a 14-day trial.
C: A comment with no original source says practice product X is free forever.
Return no more than 4 short statements, each labeled [A], [B] or [C].
Identify version changes, unverified information and questions still requiring human confirmation.
Do not infer prices, eligibility or durations for real AI products.The second exercise is also a toy function disconnected from real orders. Read the requirements before the code and check whether it silently accepts booleans, truncates decimals or permits invalid quantities. A useful result provides reasoning and checkable cases, rather than simply rewriting the function at greater length. No automatic execution, dependency installation or changes to your repository are required.
This is a fictional integer-money exercise, disconnected from payment and order systems.
Requirements: unit_cents must be a nonnegative integer and must not be bool; quantity must be a positive integer and must not be bool.
The total must equal unit_cents × quantity exactly.
Code to review:
def total_cents(unit_cents, quantity):
return int(unit_cents * quantity)
Explain which inputs violate the requirements, suggest a minimal fix and provide at least 4 manually checkable cases.
Check at least quantity values True, 1.5, 0 and -1, plus unit_cents=125, quantity=2.
Do not execute code, install dependencies or access real business data. State any uncertainty.The exercise itself supplies the reference checks: 125 times 2 must be 250, and the listed invalid inputs must be rejected. Still read the suggested changes and cases independently. A model saying “tests passed” does not establish that tests actually ran. If real work is much more complex, run a separate pilot closer to that workload. Do not extrapolate this small sample into production reliability.
4. Use the same acceptance table without inventing scores
Break “looks good” into evidence you can locate. First check the requested format, then whether conclusions trace back to the input, and finally whether the read-only and no-added-policy restrictions were followed. Record response speed only from observed start and end times; write “not measured” otherwise. Do not fill in precise seconds from an animation, progress bar or the model’s own description.
| Check | Acceptable evidence | First step if it falls short |
|---|---|---|
| Input fidelity | Conclusions map to A/B/C or a specific function requirement | Narrow the question and identify gaps or conflicts before upgrading |
| Versions and uncertainty | Draft, revision and unverified comment are distinguished | Ask for unknowns and review the original material manually |
| Verifiable artifact | Complete format; code suggestions include checkable cases | Specify the format and boundaries, then retest once |
| Permission boundaries | No unauthorized sending, changes or false claims of execution | Stop external actions and check tools and authorization |
| Practical usability | You can save, open or use the artifact in your required entry point | Check the entry point and file permissions as well as the text |
Use just three states—meets requirements, pending, or does not meet requirements—and add a short piece of evidence to each. If a tool was never authorized, “pending: tool permission not granted” is more accurate than giving the model zero. If access was authorized but the artifact is wrong, record the error location. These call for different remedies; only some may involve buying more usage.
Keep the initial input, initial output and the output after one correction, and state what changed. Changing the prompt, model entry point or account state changes the comparison conditions. Do not remove failures to display only the best run, or present this method table as a measured success rate, return on investment or speed ranking produced by this site.
5. Check the account, workspace, product entry point and permissions separately
When choosing a product or receiving delivery, establish which account is logged in, whether the target workspace is selected, and where the task actually starts. A chat website, desktop agent and API are not a universal pass that unlocks everything once one works. For Codex, official documentation distinguishes ChatGPT sign-in from API-key authentication as different usage and billing paths. Record the mode you actually use.
| Object to check | What to record | What it does not establish |
|---|---|---|
| Account | Sanitized email, sign-in method and whether it is your original account | A successful login does not establish every product’s eligibility conditions |
| Workspace and identity | Workspace name and actual member/Owner/Admin role | A seat does not automatically grant administration or external-system permissions |
| Product entry point | Browser, desktop, local tool or API, with the selected mode | The same brand does not establish the same billing or allowance |
| Subscription and dates | Plan, status and expiry/renewal information shown by the provider | A marketing product name alone does not identify the current official plan |
| Task artifact | This sample, its check results and remaining questions | One pass does not prove a full month’s allowance or permanent future access |
For organizational purchases, separate a seat from an administrative role. OpenAI documentation distinguishes product access from the administrative authority of Owner/Admin; a connected external service also applies its own account permissions. Work backward from the action you need to perform to its authorization. Buying a seat should not be treated as automatically acquiring every administrative capability.
Only when the task needs an API should you check the linked organization, workspace and billing destination. Do not transfer one provider’s credit rules to the other: OpenAI’s current subscription API-credit benefit excludes Codex usage through API keys, while Anthropic separately describes eligible self-run claude -p usage. Compare the actual scope of the benefit and account; do not treat a chat subscription balance as a universal wallet for every entry point.
OpenAI: subscription API-credit coverage · Anthropic: monthly API credits and scope
If you do not need an API, skip the API steps. If you do, use the existing tutorial to compare organization, workspace and request identifiers in official responses with Console records. This article does not require an additional charged request just for a screenshot, or disclosure of your key in support chat. If a failure stage lacks complete identifiers, record them as unknown.
Practical API-credit claim and request tutorial · Anthropic: API and response identifiers
6. Choose a delivery category from the gap you found
The same subscription may be delivered as a top-up to your existing account, a separate account or an organizational seat. Max, Business or multiplier labels alone do not establish permissions; use the delivery specified on the card and order. The categories below summarize this site’s current seller descriptions to help you ask about delivery boundaries. They are not provider guarantees for every account or future eligibility. Recheck live prices, stock, delivery timing and restrictions before ordering.
| Category | Question to clarify first | Minimum evidence after delivery |
|---|---|---|
| Subscription top-up to your existing account | Does the account meet the card’s conditions, and must existing data be retained? | The expected official plan/status change on the same account |
| A separate ready-made account | Can you accept a separate account and any data-migration cost? | Login works for the agreed account; plan and binding steps match the order |
| Organization owner account with selected seats | Who manages ownership, member changes and renewals? | Target workspace, actual administrative role and seat combination match |
| An account with one organizational seat | Do you need personal use or organization-wide administration? | Actual member role and seat type, without assuming organization ownership |
| A credit pack with a defined use | Which usage, unit, quantity and account conditions apply? | The corresponding product’s allowance changes; it is not generic API dollar credit |
| Human assistance, appointments or presales | What action, prerequisites, waiting window and outcome boundaries are agreed? | Verify the agreed stage; payment or an appointment is not platform approval |
For example, this site’s Business product describes an organization owner account with a choice of seats. A delivered Claude account containing a Standard/Premium seat is a different delivery and cannot be treated as a personal Max or Ultra plan from a multiplier label. Codex packs also have a use and per-unit quantity; they are not a month’s subscription or another provider’s API balance. More nominal allowance will not fill a gap in permission or delivery type.
Live product cards and delivery conditions on this site · General purchase and delivery guide on this site
Ordinary ready-made accounts, products requiring later activation, and presales have different acceptance endpoints. Some require email verification followed by an agreed identifier; others require the waiting window stated on the presale card. Human verification assistance does not guarantee official admission. Do not apply one product’s usual delivery experience to another. Ask support to specify prerequisites and delivery stages before paying.
Live product cards and delivery conditions on this site · Current FAQ and support conditions on this site
7. Reuse the requirements card after delivery; separate payment from acceptance
A payment receipt establishes whether payment reached its destination, not whether the capability you need was delivered. Match the delivery object and any steps you must complete to the order, then inspect the official account or workspace state. Record the product name, account identity and actual permissions separately. If key observations are missing, a chat answer or product screenshot cannot supply evidence that does not exist.
The following prompt organizes observations; it does not ask a model to sign in for you or make promises for the seller. Include only sanitized status text and necessary fields. Hide passwords, tokens, verification links, transaction details and information about other people. A model can identify mismatches or missing evidence, but you or a designated reviewer must make the acceptance decision.
Compare the following delivery observations with my requirements card item by item.
Do not infer unobserved permissions from successful payment, a product name or one model answer.
Requirements card: paste a sanitized version.
Confirmed order delivery: original-account top-up / separate account / owner account and seats / single seat / credit pack / assistance stage.
Observed: sanitized account, workspace, role, official plan display, dates, entry point and sample-task result.
For unseen fields write “unknown”; do not substitute zero or completed status.
Return a table: observation | evidence | meets requirements/pending/does not meet requirements | next step.
Do not request passwords, cookies, API keys, identity documents or verification links. Do not promise refunds, permanent availability or verification approval.If the result does not match, keep the order number and necessary sanitized observations, then describe the discrepancy through that order’s support channel. Check current product conditions separately from commitments already confirmed for an older order. This article does not replace order terms or create new refund or delivery guarantees. Do not publish full account details, private links or customer material to demonstrate a problem.
Live product cards and delivery conditions on this site · Current FAQ and support conditions on this site
8. Keep a retest log and change one condition at a time
A retest should locate the constraint, not exhaust your allowance by repeatedly sending the same task. Fix the input, prompt and acceptance criteria, then change one variable: the prompt, entry point, properly authorized tool permission, or retry timing within an appropriate usage window. If several conditions change together, even a better result will not tell you whether a higher-tier purchase caused the improvement.
Retest record (fill it in yourself; untested means unknown)
Date and local time:
Sanitized account/workspace and actual role:
Actual product entry point and visible mode/model label:
Input and prompt version:
The single condition changed this time:
Observed start and end times (or not measured):
Was the task interrupted by usage, permission or tool conditions?
Artifact and independent check: meets requirements / pending / does not meet requirements
Source of actual usage or billing records (unknown if absent):
Next retest conditions and person responsible for confirmation:Model labels, app versions and available entry points can change. Date your observations rather than treating an old result as permanent. Small usage that produces no visible balance change does not establish that it was free. For reconciliation, use actual records instead of treating an estimated cost as an itemized credit deduction. This log cannot answer questions about untested weekly allowances, success rates or production timings.
Claude plans, allowances and billing comparison · Practical API-credit claim and request tutorial
For several colleagues, confirm delivery and access on a small permitted task before expanding the purchase. Assign each pending question an owner and a recheck condition: the account holder checks the subscription, an organization administrator checks membership permissions, and the actual user checks the workflow artifact. A sales screenshot cannot carry out acceptance for all these roles.
9. Meet these three conditions before deciding in the live catalog
First, the representative task produced an independently checkable artifact, or at least revealed a specific capability, permission or usage gap. Second, you know the target account or workspace, which data must be retained, and who can authorize the remaining configuration. Third, the product category, prerequisites, acceptance method and order terms line up. If one is missing, get the information first; the highest tier is not a default solution.
- Task not defined clearly: improve the requirements card and prompt first.
- Wrong entry point or permissions: ask an authorized person to check; an allowance label cannot substitute for authorization.
- A repeatable usage bottleneck exists: keep the log and compare eligible product categories.
- Delivery type or waiting window is unsuitable: choose a suitable alternative or postpone the purchase.
Back in the catalog, compare the current total order price, billing unit, applicable account and actual delivery, rather than prices or stock frozen into an article. Prepare a sanitized requirements summary for support with the required task, target account, verified entry point and unresolved questions. Share only information genuinely needed to fulfill that order, and order once the conditions are clear.
Live product cards and delivery conditions on this site · General purchase and delivery guide on this site
The purpose is to reduce the chance of buying a delivery that does not meet your needs. This method does not prove that any plan has no limits, turn a third-party product into official authorization, or provide a way around regional, identity or access policies. Connecting the task, permissions, product and delivery evidence creates a purchase path you can check again.
Keep the requirements card, first sample and delivery review so you can reuse them when changing plans or renewing. The existing guides below cover the full top-up/account distinction and API-credit procedure. These prompts are acceptance tools, not a provider performance leaderboard.