The landscape

Subscriptions hide the meter

Team and licence tiers mostly show adoption (active users, prompt counts, which apps), not what any single task cost. Good for "is anyone using this?", weak for "where is spend going?"

Enterprise adds the dashboard

Through 2026, all three vendors shipped admin consoles with per-user, per-model breakdowns and spend caps. ChatGPT bills Enterprise in credits; Claude Enterprise reports actual dollars per user. Both expose an API.

APIs expose everything

Direct API usage is the only place with true per-token, per-model, per-call cost, down to input, output and cache. Route calls through one gateway and tag by team, or the "one line called OpenAI" hides four different things.

The shadow-AI gap

None of this sees personal accounts. Analysts estimate most AI use in large firms happens on unsanctioned tools. A dashboard only covers what's on the corporate licence. The rest is invisible.

1. Flat

A fixed price per person per month: ChatGPT Business, Claude Team, Microsoft 365 Copilot, Grok Business, Gemini inside Google Workspace. Within the allowance, heavy and light users cost exactly the same. The vendor carries the risk, so it also sets the limits, and it usually doesn't tell you exactly what they are. "More usage than Pro", "5x Standard" and "expanded" describe the allowance relative to another plan, not in units you can budget.

2. Flat, then metered

Once you exhaust the bundle you can wait, upgrade, or switch on paid extra usage. That extra usage is billed in credits or at API rates. Nearly every major vendor now offers this: Claude's paid plans ("turn on usage credits to keep working at standard API rates"), ChatGPT Business (pooled workspace credits), Google AI Pro and Ultra (AI credits for Flow and Antigravity), GitHub Copilot (AI Credits at $0.01 each), and Cursor and Grok (prepaid top-ups).

3. Pure usage

API access, and increasingly enterprise plans, bill every token. Claude Enterprise is now priced at "$20 per seat per month plus usage billed at API rates". ChatGPT Enterprise offers credit-based or token-based billing. Microsoft Copilot Studio agents bill per action in Copilot Credits.

What the table tells you

No major business plan is truly unlimited for the expensive features any more. Most say "unlimited" only for the lightest model, for example ChatGPT's "Instant" or GitHub's code completions. Everything else is a bundle with a meter behind it.

The default limit is often off

In several of the most widely used products the default spending limit is off, so the decision falls to whoever configures the account. ChatGPT Business: "By default, all seats and users have no limits specified". Claude offers "Set to unlimited". Microsoft auto-adds new metered services to spending policies by default.

Output costs about five times input

Across almost every flagship model: $10 against $50, $4 against $20, $2 against $10. Reasoning models bill their hidden "thinking" as output. OpenAI's pricing page says output prices "include… reasoning tokens, even though reasoning tokens are not visible."

The gap between tiers is about 100×

That's wider than the 50–60× this site's calculators assume. GPT-6 Luna to GPT-6 Astra is $0.10 to $10 for input, and Claude Haiku 5.5 to Fable 5.1 is the same.

Some current prices are introductory

Gemini 4 Argon doubles on 1 December 2026. OpenAI says GPT-5.6 Sol's promotional pricing runs "at least through November 21, 2026". Budget on the post-promotion price.

1. The cliff edge: flat becomes metered

Inside a bundle, your marginal cost is zero. One message past it, you pay full rate. That's a change in kind, not degree. It's why heavy users look cheap right up until the moment they don't.

Shape: a step, then a straight line.

2. Model choice multiplies everything

The same work on a flagship model costs about 100× what it costs on a light one. Ramp, the corporate-card provider, calls "model tier migration" (teams quietly moving from a lightweight model to a frontier one) the single biggest driver of unexpected AI cost increases.

Shape: a constant multiplier, but a large one.

3. Long chats grow with the square of their length

Every message resends the whole conversation. The cost of turn n therefore grows with n, and the total cost of a thread grows with n². In the worked example below, doubling a thread from 20 to 40 turns raises its cost about 3.3 times. Going from 40 to 80 turns raises it about 3.5 times.

Shape: quadratic. This is "compounding" in the everyday sense, not exponential in the mathematical sense.

4. Long-context surcharges switch on above a threshold

OpenAI charges 2× input and 1.5× output above 272K tokens. Grok charges 2× on every token in the request once the prompt reaches 200K. Claude Haiku 5.5 charges 5× above 100K.

Shape: a step that applies to the whole request, not only to the tokens over the line.

5. Agents call the model in a loop

One instruction can become dozens or hundreds of model calls. Each call carries the growing context of the ones before it, so a single agent task is a long chat that the machine has with itself. Sub-agents multiply this again. GitHub's own billing guidance says an agentic session "will consume significantly more usage than a quick question in chat". OpenAI puts typical Codex cost at "$100–$200 per developer per month", while noting that it "varies significantly".

Shape: quadratic per task, multiplied by the number of agents running in parallel.

6. Multipliers stack

Premium speed modes (2× for Fast, 6× for Ultrafast), long-context surcharges (1.5–2×), data residency (1.1×) and model tier (up to about 100×) all multiply together. Ultrafast GPT-6 Astra on a long prompt is listed at $120 input / $450 output per million tokens. GPT-6 Luna on a short prompt is $0.10 / $0.50. That's 1,200× on input and 900× on output from the same vendor, on the same page.

Shape: the product of every multiplier you have switched on.

Seats, then credits

Team buys seats. Enterprise moved to a credit model with a 2026 Global Admin Console: slice consumption by user, group, project and model, set workspace defaults, cap power users, pull it via a Cost API. Office and agent features now meter in tokens too.

A utility meter

No seat limits: usage draws from one org token pool, with spend caps at every level. Enterprise analytics report cost by group and by user, showing artifacts, files and tools next to their cost, filtered by your existing org chart. Two APIs expose it.

Bundled in the licence

Cost lives inside the M365 licence, so admin reporting is about adoption, not spend: active users, prompts, app usage, agent uptake (via the admin centre, Viva Insights and Graph API). Great for rollout tracking, quiet on per-task cost.

Seats with a console

Business and Enterprise tiers arrived in January 2026: Business at $30 per seat, Enterprise on custom pricing. Management runs through the xAI console: seats, consolidated billing, role-based controls and usage analytics, with Enterprise adding SSO, SCIM directory sync, audit tooling and a customer-key data vault. Cost sits inside the seat, so reporting leans toward adoption rather than per-task spend.

Anthropic: margin recovered fast

Gross margin moved from roughly −94% in 2024 to about 60% in 2026, driven by inference efficiency rather than price rises: blended token prices have been falling throughout. Revenue per megawatt of compute rose from about $16m to a projected $60m in under a year. Enterprise API mix is the reason: business customers generate several times more revenue per token than consumer users, and their query patterns are cheaper to serve.

OpenAI: scale without margin

Gross margin was around 33% in 2025, improving to roughly 39% in early 2026, against a $20.9bn operating loss on $13.07bn of revenue. The drag is structural: a very large free consumer tier consuming inference without matching revenue. Internal plans point to cumulative losses near $115bn through 2029 before cash-flow profitability around 2029–30.

Cheaper per query, bigger bills

Inference cost per query has fallen roughly 95% since early 2023. Total spend still rose, because agentic workloads consume many times more tokens per task than a chat message. Whether margins expand depends on whether costs fall faster than demand grows, and so far demand has won.

Compression, not collapse

AI margins are improving (41% to 52% in two years), but the consensus is they settle in the 60–65% band rather than returning to the SaaS 80%. Traditional software adding AI features typically gives up 12–17 points. The old 80% benchmark is unlikely to come back as the industry standard.

What that means for buyers

Vendors facing structurally lower margins have to recover cost somewhere: usage-based pricing, credit systems, tiering and caps. That's precisely the shift described earlier on this page. Expect more metering, not less, and expect the meter to be passed to you.

Why waiting is not a strategy

Per-token prices have fallen sharply and will likely keep falling. Total spend still went up. If cheaper tokens simply invite more use (longer threads, more agent calls, more re-runs), then discipline, not price, is what determines your bill. Efficiency compounds in a way that waiting does not.

Model access is not a moat

Everyone can reach the same frontier models on similar terms, so capability alone stopped being differentiating. Investor language has shifted markedly during 2026 away from thin wrappers and toward workflow ownership, proprietary data and domain depth. One investor went as far as saying the terminal value of software itself is now being questioned.

What defends a position

A Stanford CodeX white paper ranks vertical AI moats in ascending strength: workflow and interface, vertical tooling, built-in compliance, and strongest of all the accumulated data layer. Regulated verticals (legal, healthcare, financial services) score highest, because domain knowledge and compliance are slow to replicate whatever the model can do.

The compounding loop

Use generates data, data improves the product, the better product drives more use. Sectors where the data is both proprietary and directly useful to model performance benefit most. The honest caveat: most software companies discover on inspection that they don't have a data moat. They have a head start, and head starts erode.