Your team rolled out Claude or ChatGPT across the company. Adoption spiked. Then the invoice arrived, and one engineer's line was forty times another's.
That gap is the whole story of 2026 AI budgeting, and almost no budget written in late 2025 accounts for it. Most were built on chat assumptions: a person types a prompt, tokens go out, an answer comes back. The cost driver now is agents, which plan, retrieve, call tools, verify and retry, many times per task. Australian card data from Weel, covering more than 4,000 SMBs, shows the split plainly: the average Anthropic customer spent USD $1,082 per month in the June 2026 quarter against USD $298 for the average OpenAI customer, a 3.6 times difference driven by consumption rather than seats.
This guide covers what you are actually buying, where the money goes, how to govern it without killing adoption, and how to code it properly in your ledger. If you need to defend the spend upstairs afterwards, the companion piece on how to justify AI spend to your board picks up where this one ends.
Published: September 2026
There are three cost families, and mixing them is the most common budgeting error we see.
Seats and subscriptions. A fixed monthly fee per user. Predictable, forecastable, behaves exactly like the rest of your software stack.
Usage and tokens. Metered consumption billed at API rates. Varies by which model runs, how hard it runs, and how much context it is handed. This is the variance.
AI embedded in software you already buy. Xero's JAX, Dext, Employment Hero, Fathom. You are already paying for these inside existing subscriptions. They behave like SaaS, and they belong in a different conversation from generative tokens. We cover that side separately in how Australian SMEs use AI in HR and finance.
Traditional software gives you one number twelve months out with near-perfect accuracy. Only the first of these three families does that.
Prices below were checked on 1 September 2026 against vendor pricing pages. All figures are USD unless stated. Re-check before you commit anything to a budget, because these move monthly.
Anthropic. Claude Enterprise is now published rather than quote-only: $20 per seat per month, billed annually, plus all usage billed separately at standard API rates. There is no included token allowance in the seat fee. Enterprise adds SCIM, audit logs, custom data retention, role-based access and a 500K context window on the default model. Team seats are $20 standard or $100 premium per month on annual billing ($25 and $125 monthly), and you can mix seat types. Pro is $17 per month annual or $20 monthly.
Current API rates per million tokens: Haiku 4.5 at $1 in / $5 out, Sonnet 5 at $2 / $10, Opus 5 at $5 / $25, and Fable 5 at $10 / $50. Cache reads cost roughly 10 per cent of base input. The Batch API is 50 per cent off both directions. US-only inference adds 10 per cent. Opus 5 fast mode is double standard pricing. Two line items most budgets miss entirely: Managed Agents bill $0.08 per session-hour of active runtime on top of tokens, and web search is $10 per 1,000 searches plus tokens.
One caveat worth flagging. Sonnet 5 launched at an introductory $2 / $10 with a scheduled increase to $3 / $15. Anthropic's documentation currently states the introductory rate has become the standard price and the increase will not proceed. Confirm on the day you publish or budget.
OpenAI. ChatGPT Business is $20 per seat annual or $25 monthly, two-seat minimum, capped at 200 paid seats from 24 August 2026, with a Premium tier at $100 / $125. Enterprise remains custom, with procurement reports clustering near $60 per user per month at around a 150-seat minimum. Treat that as reported, not published. The API bills separately from seats.
Microsoft. Microsoft 365 Copilot in Australia lists at roughly AUD $44.90 per user per month for the full add-on and about AUD $31.40 for Copilot Business under 300 seats on annual billing, ex GST, and requires a qualifying base licence. GitHub Copilot Business sits near USD $19.
The shape difference matters more than the price. Copilot is close to classic SaaS: one predictable number per seat. Claude Enterprise is a seat plus a meter. A board that signed off Copilot at $45 a user has not implicitly approved a token bill, and treating the two as the same category is how finance loses control of the line.
A marketing manager drafting emails on Sonnet is a small, predictable line. A senior engineer running a coding agent on Opus at high effort against a large repository is a different product entirely. Same seat. Wildly different bill.
The mechanics explain the gap. Gartner's March 2026 analysis found agentic workflows make 5 to 30 model calls per user-initiated task rather than one. GitHub research from May 2026 put agentic coding tasks at roughly 1,000 times the tokens of a single-turn query. Nothing about the seat price signals that.
The consequences are already public. Gartner reported in June 2026 that $200 to $500 per developer per month has become common among its clients, with bills above $2,000 appearing and outliers reaching $20,000 and $32,000 in a single month. Gartner projects AI coding token costs will match global-average developer monthly pay by 2028.
Practical conclusion: budget chat and agents as two separate sub-codes. If they sit in one line, you can see the total move but never which half moved.
If your company has not bought a business plan, your people are on personal subscriptions with the company card. Weel's index is literally reading those transactions: 30.8 per cent of Australian SMBs on its cards were paying for AI tools in June 2026, up from 22.1 per cent in January.
Before you set a single cap, do this:
Skip this and you will cap the official invoice while the real spend keeps sitting in Software, Other. Our Australian SME technology spend benchmarks give you a reference point for what the total software line should look like at your size.
Phase 1, weeks 1 to 6. Role-based soft caps plus one organisation-level hard cap as the safety net. Do not set per-user hard caps yet, because you do not have the data to know where the thresholds belong. Tier engineering, data and research highest, operations and finance in the middle, general business users lowest. Set the default model to Haiku or Sonnet 5 and put the expensive models behind an allow-list.
Phase 2, months 2 to 4. With four to six weeks of usage data, set per-user caps at roughly 1.5 times the 80th percentile for each tier. Alert at 70 and 90 per cent of cap, by team rather than only at organisation level. Require a short written justification for any increase, which keeps the culture open without writing blank cheques.
Phase 3, month 5 onwards. Usage normalises as people stop experimenting and start building repeatable workflows. Cost per unit of output should now be falling even if total usage is flat. Governance shifts from controlling spend to improving the ratio.
Expect a 2 to 3 times spike in the first 60 to 90 days after any enablement workshop. That is people learning the tool, not a budget failure. Setting your annual budget from month one will overshoot; panicking at month one will cut too early.
Model routing. The largest lever for most companies. Sonnet 5 at $2 / $10 is now much closer to Haiku than the old Sonnet was, which changes the arithmetic. The new risk is the opposite direction: teams defaulting agent sessions to Fable 5 or Opus fast mode. A workable 2026 split for a mixed Australian team is roughly 60 per cent Haiku, 30 per cent Sonnet 5, 8 per cent Opus 5, 2 per cent Fable 5, with Fable reserved for long-running agents.
Effort and thinking level. This is now the number two lever and almost nobody uses it. Datadog published first-party numbers in August 2026: it saved over $1 million per month across its engineering organisation, $687,000 from defaulting Opus to Sonnet and $288,000 from dropping coding-agent effort from high to medium. A one-week A/B across more than 1,000 engineers moved cost per user from $156.70 to $114.40, a 27 per cent drop.
Prompt caching. Cache reads cost around a tenth of base input. Cache writes pay for themselves after one subsequent read. The highest-value targets are not chat prompts but agent system prompts, repository maps and policy documents, which repeat on every call.
Batch processing. Overnight reports, bulk analysis and non-urgent data work take a flat 50 per cent discount through the Batch API. The trade is asynchronous delivery.
Context discipline. Bloated context is why agent sessions explode. Pasting an entire codebase is a finance control problem, not just an engineering habit.
Stop-the-line on retries. An agent that fails and loops can burn four figures while nobody is watching. Per-user caps do not help when a single session can sprint. The organisation-level hard cap is the real backstop.
Seat hygiene. Multi-model is becoming the default, and 32 per cent of Weel's AI-paying customers use both major vendors. Paying for a chat tool, a coding agent, a code-host assistant and an editor for the same twelve engineers gets you to several hundred dollars per developer per month in seats alone, before a single token. Standardise on one chat surface and one coding agent.
Committed spend. Both major vendors sell it. Useful only once you have eight to twelve weeks of real data. Do not commit in week two of a pilot.
This is the part almost nobody writes, and it is where a finance function earns its keep.
Classification. SaaS and API access where you do not control the underlying software is a service contract, expensed as incurred, the same as your cloud bill. The IFRIC agenda decision on cloud configuration costs points the same way. The ATO treats periodic payments to use software in your business as deductible in the year incurred; the in-house software rules apply to software you hold, not to a chat subscription. If you are buying a fixed-term coding subscription solely to build an identified internal system, that becomes a facts-and-circumstances question under AASB 138. Default to expensing and raise it with your accountant if that specific fact pattern applies.
Chart of accounts. Split the parent code into four children:
Seats are forecastable. Tokens are the variance. If they share one code, the board can see the total move but cannot see what caused it. If your account structure needs work before you can do this, start with our guide to a chart of accounts.
GST. Where a non-resident supplier such as OpenAI or Anthropic does not charge GST, this is generally a business-to-business imported service rather than an inbound intangible consumer supply. The reverse charge under Division 84 can apply where the buyer cannot claim a full input tax credit, which matters for financial services, insurance, residential property and mixed entities. For a normal fully taxable trading company it usually nets to zero. If the supplier does charge 10 per cent GST because you did not provide your ABN and GST-registered status, claim the input tax credit as usual. Confirm your position with your BAS agent rather than assuming.
Prepayments and accruals. An annual seat invoice goes to prepayments and releases monthly. Usage is accrued monthly from the vendor dashboard, not from the credit card date. Otherwise August looks cheap and September looks like a crisis.
Foreign exchange. API rates are USD. Book at the payment-date rate and let a small FX gain or loss sit where it belongs. Never forecast twelve months of token spend in AUD without stating an FX assumption, and never add USD API rates to AUD salary figures in the same total.
Group structures. If your engineers sit in an offshore entity and the Australian company holds the vendor contract, recharge on an arm's-length basis. Token spend attributes cleanly by seat, so this is easier than most intercompany allocations.
Illustrative only, using published seat prices as at September 2026. Seats are USD $20 per month, so 50 seats is USD $12,000 per year. At an assumed AUDUSD of 0.65, that is roughly AUD $18,500. State your own rate.
Rather than invent an average usage figure, model three scenarios:
Chat only, Haiku and Sonnet defaults, caps on. Usage USD $15,000 to $25,000. Total USD $27,000 to $37,000. This behaves like SaaS.
Mixed, no routing, some coding agents. Usage USD $50,000 to $90,000. Total USD $62,000 to $102,000. This is where most Australian technology SMEs land in year one.
Uncapped agents defaulting to the top models. Usage USD $150,000 to $400,000 or more. Total USD $162,000 to $412,000 plus. This is the slide that gets your cap approved.
Applying routing, caching, batch and an effort default of medium to the middle scenario should reduce usage costs materially. Datadog's published 27 per cent per-user reduction from two settings changes is a large-company existence proof, not a promise for your file.
For the cash-flow consequences of a variable line this size, run it through the cash flow forecast calculator rather than treating it as a fixed monthly cost.
Four jobs, none of which take long:
That is the difference between a software line and a managed cost. If nobody in the business has time to own it, that is exactly the kind of work our bookkeeping support for budget ai tools picks up, and it is the same discipline described in our month-end close process for Australian SMEs.
Does Claude Enterprise include AI usage in the seat price?
No. Claude Enterprise is $20 per seat per month billed annually, and all usage is billed separately at standard API rates with no included token allowance. Administrators can set spend limits at both organisation and individual user level, which is the only real control you have over the variable half of the bill.
Why is one engineer's AI bill forty times another's?
Model choice, effort setting, context size and whether they are running agents. An agent makes many model calls per task rather than one, and Gartner has measured 5 to 30 calls per user-initiated task, with GitHub research putting agentic coding at roughly 1,000 times the tokens of a single query. It is rarely about hours spent in the product.
Is AI tool spend capital expenditure or operating expenditure?
Operating expenditure in almost every case. You are paying to consume a service, not acquiring or creating an asset you control, so it sits with your cloud costs. The narrow exception involves a fixed-term subscription bought solely to build an identified internal system, which is a facts-and-circumstances judgement under AASB 138. Ask your accountant before capitalising anything.
Do we have to reverse charge GST on OpenAI or Anthropic invoices?
It depends on your entitlement to input tax credits. For a fully taxable trading company the Division 84 reverse charge usually nets to nil, but businesses making input-taxed supplies, including financial services and residential property, need to look at it properly. If the supplier has charged 10 per cent GST, claim the credit as normal. Confirm your position with your BAS agent.
Should we buy both Copilot and Claude?
Usually one chat surface and one coding agent, not five products. Multi-model is becoming common, with 32 per cent of AI-paying Australian SMBs on Weel's cards using both major vendors, but stacking a chat tool, an editor, a code-host assistant and a coding agent for the same engineers reaches several hundred dollars per developer per month in seats before any usage.
How much cheaper is Haiku than the top models?
Haiku 4.5 is $1 per million input tokens and $5 output. Fable 5 is $10 and $50. That is a ten times gap on both sides. For summarisation, classification and first drafts, the output quality difference rarely justifies it, which is why model routing remains the single largest lever available.
Should we budget AI monthly or annually?
Set an annual range with conservative, base and aggressive scenarios, then review actuals monthly and reset the outlook quarterly. A single project sprint can move the monthly bill by 30 to 50 per cent, so a fixed annual number will be wrong by the second quarter.
What is the fastest saving we can make this week?
Change two defaults: route routine work to the cheapest capable model, and drop coding-agent effort from high to medium. Datadog attributed $687,000 and $288,000 per month respectively to exactly those two changes across its engineering organisation.
Scale Suite is a Sydney-based provider of outsourced finance teams and fractional CFO services for Australian SMEs. We deliver weekly bookkeeping, payroll, BAS/IAS lodgement, cashflow reporting, management accounts, and strategic fractional CFO oversight, all as a fully embedded team that works inside your business.
CA-qualified, Xero Certified, and registered BAS Agents, we replace fragmented bookkeepers and once-a-year accountants with one responsive finance function at a fraction of the cost of full-time hires. We serve growing businesses across Sydney, Melbourne, Brisbane, and Perth, with packages starting from $1,500 per month and no lock-in contracts.
See Sydney bookkeeping when you are dealing with budget ai tools, or see what fractional CFO input on budget ai tools cost at your stage.
We review and check this guide periodically. At the time of writing (September 2026), all information was current. Scale Suite is a registered BAS Agent, not a licensed tax advisor or financial advisor. This content is general information only and does not constitute professional tax, financial, or legal advice. Some details may change over time. AI pricing in particular changes monthly, so verify current rates against official provider documentation before making budget decisions.
Scale Suite is a Sydney-based provider of outsourced finance and HR services for Australian SMEs. We deliver bookkeeping, financial reporting, payroll processing, fractional CFO support, recruitment, employee onboarding, people and culture support, and fractional HR oversight, all as a fully embedded team that works inside your business.
Employment Hero Gold Partner, CA-qualified, and Xero Certified, we replace fragmented finance and HR processes with one responsive, senior-level function at a fraction of the cost of full-time hires. We serve growing businesses across Sydney, Melbourne, Brisbane, and Perth, with packages starting from $1,500 per month and no lock-in contracts.
30 minutes with our team.
We will review your current finance setup, compare the full cost of an internal hire against our embedded team, and show you exactly what your finance function should cost at your stage of growth.
You will leave with a clear view of what is working, what is missing and where you would save.
No lock-in contracts. 30-day money-back guarantee.
Prefer to book directly? Grab a time here.
