What Claude and ChatGPT Actually Cost a Business
Finance teams spend real energy arguing over whether a seat costs $20 or $25 a month. That argument is worth $60 a year per person. Meanwhile the same company’s engineering team picks a model, and that single choice moves the bill by a factor of ten.
Anthropic’s published API rates make the gap plain. Haiku 4.5 runs $1 per million input tokens and $5 per million output. Fable 5 runs $10 and $50. Same vendor, same billing page, ten times the price.
Almost every article on AI budgeting gets this backwards. The seat price is the visible number, so it gets the scrutiny. The variable that actually determines spend sits three clicks away.
Scrutiny is new here because the spending is. Eurostat reports that 20.0% of EU enterprises with 10 or more employees used AI technologies in 2025, against 13.5% the year before and 7.7% in 2021. The national spread is wide: Denmark sits at 42.0%, Romania at 5.2%. Three years ago this was a line nobody had to defend. Now a fifth of European firms carry it.
The seat prices, in actual numbers
Claude’s ladder starts free. Pro costs $17 a month on an annual subscription, billed as $200 up front, or $20 if you pay monthly. Max starts at $100 a month. Team seats come in two grades: a standard seat at $20 per month billed annually or $25 monthly, and a premium seat at $100 or $125 on the same terms, for teams of 2 to 150. Enterprise runs $20 per seat plus usage charged at API rates.
OpenAI has landed in almost exactly the same place. ChatGPT Business costs $20 per user per month billed annually, or $25 billed monthly, starting at two users. Enterprise pricing goes through sales.
Two competitors, converging on identical numbers for the business tier. Whatever a company decides between them, it will not decide on seat price.
Where the variance actually lives
Per-token pricing is where budgets diverge. Across Anthropic’s current lineup, Haiku 4.5 sits at $1 and $5 per million input and output tokens, Sonnet 5 at $2 and $10, Opus 5 at $5 and $25, and Fable 5 at $10 and $50.
Two levers cut those figures further. Prompt caching drops repeat reads to as little as $0.10 per million tokens on Haiku. Batch processing takes 50% off for work that tolerates a delay.
Stack those together and the spread between a carelessly built pipeline and a carefully built one is not 20% or 30%. A workload routed to the largest model, uncached and unbatched, can cost twenty times the same workload routed to a small model with caching on and batching enabled. That is the number worth an afternoon of somebody’s attention.
Which makes the standard advice — “costs range from trivial to substantial depending on the model” — true and useless. The useful version names the models and the multipliers.
Why checkout is sometimes the hard part
Access to financial services has widened enormously. The World Bank’s Global Findex 2025 puts account ownership at 79% of adults worldwide, with 84% of adults in low- and middle-income economies owning a mobile phone.
An account is not the same instrument as a card that clears a dollar charge from a US merchant, though. Plenty of perfectly ordinary local debit cards fail that specific test: the issuer blocks international e-commerce by default, or the card network declines cross-border card-not-present traffic from certain corridors, or the conversion spread makes a $20 subscription cost $23.
Both vendors narrow the options further. OpenAI’s own pricing FAQ states that Go, Plus, Pro and Business purchases take “any major credit card,” and reserves invoicing for Enterprise customers who contact sales. Anthropic’s plan comparison shows the same shape, with ACH and net terms appearing only as you move up the ladder.
So below the enterprise tier there is one payment rail and no fallback. Buyers whose local cards fail abroad often use a vcc for Claude, funding one card per subscription. The per-tool separation has a side benefit that has nothing to do with acceptance: each statement line maps to exactly one vendor, which makes reconciliation trivial and makes an unused tool obvious.
The line a card cannot cross
This is where the workaround needs a warning label, because the two problems get conflated constantly.
A virtual card solves acceptance. It does not create eligibility. Anthropic publishes a list of supported countries and regions for commercial API access and for Claude.ai separately, and states that it reserves the right to withhold service from entities whose majority ownership traces to nations outside its Supported Regions Policy. The list runs to roughly 190 entries and excludes Ukraine’s Crimea, Donetsk, Kherson, Luhansk and Zaporizhzhia regions explicitly.
Signing up from outside that list breaches the terms whatever card you present. Where sanctions apply, it stops being a contractual matter and becomes a legal one. Anyone weighing a payment workaround should check the supported-countries page first and the card second.
Regulation shapes the cards themselves too. Directive (EU) 2018/843, the Fifth Anti-Money Laundering Directive, cut the threshold for identifying prepaid card holders from EUR 250 to EUR 150, and allows anonymous prepaid cards issued outside the Union only where they meet EU standards. Member states had until 10 January 2020 to transpose it. Expect identity checks, in other words, and treat any issuer advertising genuine anonymity as a red flag rather than a feature.
Four habits that keep the number honest
Do not start on a team plan. Claude Team requires two seats and bills $25 each monthly, while two individual Pro seats cost $20 each. The $10 monthly difference buys central billing and SSO, which a two-person shop rarely needs yet.
Treat annual prepayment as a real commitment. Claude Pro at $200 up front against $240 paid monthly saves about 17%, which is worth having once a tool has proved itself. Before three months of steady use, the flexibility is worth more than the discount.
Set spend limits before launch rather than after the first surprising invoice. Anthropic’s Enterprise tier gives admins user-level and organisation-level spend controls; on lower tiers the discipline has to come from whoever configures the API key.
Review seats monthly and cancel the dormant ones yourself. Do not count on a regulator to make that easy. The FTC’s amended “click to cancel” provisions did not survive court challenge, and the Commission’s Negative Option Rule page now shows a February 2026 revision conforming the rule to those federal court decisions, followed by a March 2026 advance notice seeking comment on fresh amendments. Cancellation friction remains a live rulemaking question rather than a settled protection.
One date deserves a calendar entry. Sonnet 5’s $2 and $10 per-million-token rates are introductory pricing that runs through 31 August 2026, after which the standard $3 and $15 apply. For any team with a Sonnet-based pipeline in production, that is a 50% increase on a line item most forecasts still carry at the old number.
Observer Voice is the one stop site for National, International news, Sports, Editor’s Choice, Art/culture contents, Quotes and much more. We also cover historical contents. Historical contents includes World History, Indian History, and what happened today. The website also covers Entertainment across the India and World.
