Claude Sonnet 5 arrived this month as the new default model on Free and Pro plans, and it is already live in Claude Code and the API. The headline is the price: launch pricing of US$2 per million input tokens and US$10 per million output tokens, holding until 31 August 2026, then stepping up to US$3 and US$15. On paper that undercuts Opus 4.8 by a wide margin. But the sticker price is only half of what decides your bill, and for Australian teams planning their AI spend, the other half matters more than usual.
Two things the sticker price does not show
The first is a new tokenizer. Anthropic's release notes flag that the same input can consume anywhere from 1.0 to 1.35 times as many tokens as it did on Sonnet 4.6. Your prompts have not changed; the meter simply counts differently. A document set that metered at 100,000 tokens last month can meter at 135,000 tokens today, before you alter a single word.
The second is output verbosity. On one third-party benchmark, the Artificial Analysis Intelligence Index, Sonnet 5 generated close to twice the output tokens of Opus 4.8 on the same tasks. That is a single benchmark rather than a universal ranking, so treat it as directional. The direction is clear enough, though: Sonnet 5 tends to write more to reach the same answer, and you pay for every token it writes.
Put the two together and the arithmetic gets interesting. Because the per-token price is much lower, Sonnet 5's total cost on that benchmark still came in below Opus 4.8, and in a later run below Fable 5 as well. Cheaper, yes. As cheap as the headline suggests, no. The token multiplier quietly eats a slice of the advertised savings.
The worked math
Take a mid-sized Sydney professional services firm running document review and client correspondence through the API. Under Sonnet 4.6 assumptions it processed 50 million input tokens and 10 million output tokens a month. A naive Sonnet 5 budget says 50 x US$2 plus 10 x US$10, or US$200 a month. Cheap.
Now apply the multipliers. At a mid-range tokenizer factor of 1.2, that 50 million of input meters as 60 million. If outputs run 1.5 times longer, 10 million becomes 15 million. The launch-price bill is 60 x US$2 plus 15 x US$10, which is US$270 a month, 35 per cent above the naive estimate. From 1 September the same volume costs US$405 a month at the stepped-up US$3 and US$15 rates. In Australian dollars that is roughly $620 a month, or about $7,400 a year, for a workload the sticker-price estimate said would cost well under half that.
The point is not that Sonnet 5 is expensive. It is that the gap between the advertised price and your real bill is decided by your token profile, not the rate card.
What this means for Australian teams
Your usage has already moved. Sonnet 5 is the new default, so most Free and Pro traffic shifts to it whether or not anyone chose it. Check which model your bill is actually metering.
Budget from your token profile, not the per-token price. A verbose model at a low rate can out-cost a terse model at a higher one. Measure tokens per task on your own workloads.
Diarise 31 August 2026. Launch pricing steps up 50 per cent on both input and output the next day. If your budget was set on US$2 and US$10, it is wrong from 1 September.
Treat single benchmarks as directional. The verbosity figure comes from one third-party index; Anthropic's quality and safety numbers are its own testing. Your workloads are the only benchmark that decides your bill.
Keep API pricing in US dollars in your models and convert at your treasury rate. Exchange movement adds a second variable on top of the token multiplier.
How to check what you are actually paying
Start with a baseline. Pull last month's usage by model from the Claude Console and record tokens in, tokens out, and cost per workflow. Then re-run a sample of the same tasks on Sonnet 5 and compare the token counts, not just the dollar totals. The ratio between the two runs is your real multiplier, and it will differ by task: code generation, summarisation and extraction all tokenize differently.
Next, set spend alerts before the September price step rather than after it. Claude's admin tooling now supports per-group cost visibility and alerting, and even a simple weekly export into a spreadsheet catches drift early. The failure mode to avoid is discovering the tokenizer change three invoices late, after the finance team has already flagged the variance.
Finally, re-baseline after 31 August. A budget built on launch pricing has a 50 per cent error baked in from day one of September. Two hours of measurement now saves an awkward conversation with your CFO in October.
Automata AI is a Sydney consultancy that sets up token-level cost governance for Australian businesses, so you know what each Claude workflow really costs before the invoice lands. If you want your Sonnet 5 numbers checked against your actual workloads, book a brainstorm call and we will walk through the math with you.
