What Claude Sonnet 5.5 changes for accounts payable. A cheaper-to-run version of the model your team uses to read an invoice, pull the line items, code them to a gl account, and route them for approval. The price per token did not change. The cost per invoice did, because it finishes the same job in fewer tokens.
This is written for a controller at a 200-person services firm. Not for the people who build AI. If you priced out AP automation a year ago and shelved it because the per-document cost did not pencil out against what your clerk earns, the september 28 release is the reason to re-run that math.
#What did Anthropic actually ship on september 28?
Anthropic released Claude Sonnet 5.5 on september 28, 2026, at the same price per token as Sonnet 5. It kept the 1M-token context window. The point that matters for your invoice pile is Anthropic's claim of a lower cost per task, because the model needs fewer tokens to finish the same work. the release and the pricing are here.
This did not land in a vacuum. Anthropic and OpenAI both shipped lower-cost models the week of september 22, per cnbc. The unit economics your budget assumed six months ago have moved. If the cost case for automating small invoices was marginal in march, it is a different case today.
How much does AI invoice processing cost now?
The model cost per invoice fell with Claude Sonnet 5.5, released september 28, 2026 at the same price per token as Sonnet 5 but with fewer tokens per task. A small invoice that cost next to nothing to read still does, only less, which changes whether automating the long tail pencils out.
#Why does a lower cost per task change which invoices you automate?
Most AP automation business cases only cover the top of the stack. The big vendor bills, the recurring ones, the invoices worth automating because a single one is worth a lot. The long tail gets left to a clerk: the small courier charge, the one-off software licence, the reimbursed parking. Hundreds of them a month, each too small to justify the old per-document spend.
A lower cost per task does not change the big invoices much. They were already worth it. It changes the tail, where the AI cost was close to the value of the document itself. Once reading a small invoice costs next to nothing, it is economical to run extraction across every invoice that arrives, not just the ones over a threshold.
The win was never one expensive invoice. It was the four hundred cheap ones nobody had time to key in.
#What does a full AP build actually do for a services firm?
On our AI implementation invoice-processing work, the flow is the same for a small bill and a large one. The model reads the invoice, pulls the line items, codes them to a gl account, and routes them for approval. It posts to quickbooks or netsuite with an audit trail, and anything it is unsure of lands in a human approval queue.
- Extract: vendor, amount, line items, tax, po number, read off the PDF or the email body
- Code: map each line to the right gl account and cost center
- Route: send to the right approver based on amount and department
- Post: write to quickbooks or netsuite with a full audit trail
- Review: anything below a confidence bar goes to a human approval queue, not straight through
We are
Stennir builds an AP pipeline with a human approval queue and an audit trail on every posting, so a controller stays in control of what hits the ledger.
We aren't
Stennir is not a rip-and-replace ERP that auto-pays invoices with no human in the loop and no record of why a line was coded the way it was.
A lower model cost does not replace the thinking. You still decide which invoices auto-post and which wait for a signature, where the confidence bar sits, and who approves what. That is the AI consulting part of the work, and it is the part that keeps an auditor comfortable. The cheaper model just removes the reason to leave the long tail on a clerk's desk.
If you shelved AP automation on cost, the september 28 release is a clean reason to look again. We will take your real invoice mix, the long tail included, and show you the cost per document at today's pricing against what your team spends keying them by hand. book a 30-min discovery call and bring last month's invoice count. You can read how we think about this kind of build on the /journal.
