At the beginning of October 2026, Anthropic changed three prices in one launch. Its new small model, Claude Haiku 5.5, costs 90% less per token than Haiku 4.5 on requests up to 100,000 tokens (a token is a small chunk of text). Sonnet 5.5 got cheaper on text it rereads, such as instructions sent with every request. And Max and Team subscription plans now get a monthly API credit, capped at $500 per team. Each change reaches a buyer differently, or not at all.
Why It Matters
What reaches you depends on what your contract counts. If you pay per token for Sonnet 5.5, the cut applies with no switch, but only to text your setup already rereads, and the next invoice shows how much. The Haiku cut arrives only if you switch and Haiku 5.5 does your work just as well. The credit lands only on Max and Team plans, and only until it runs out. A seat fee doesn’t fall when tokens get cheaper, though some seats also carry a token meter: Anthropic’s Enterprise seat bills every token at standard API rates.
The Decision
The question is how much of next year’s AI budget to change before an invoice or your own test shows the cut. Change on announcements and you may book savings that never arrive. Wait for every invoice and a line needing a model switch stays at the old cost.
What To Do This Week
- Ask your finance lead to list each AI contract line with what it counts: seats, tokens or credits, sometimes two on one line.
- For each seat plan, ask its owner whether it now includes usage credits and what happens when they run out.
- For the largest per-token line, ask your AI lead to compare, by October 30, 2026, a cheaper model such as Haiku 5.5 with the current one on cost per finished task your team accepts without rework, so that line’s number comes from the test, not the announcement.
What Not To Do
Don’t book an announced cut against a fixed seat fee. Don’t treat a new credit allowance as free room before you know what happens when it runs out. And don’t budget a vendor’s average as your saving: Anthropic’s footnote puts its “around 75% less to run” for Haiku below the 90% because it counts longer requests and slightly more tokens per task, on its request mix, not yours.
Signal Boost
Check What Your AI Contract Counts, A-Eye Level, October 1, 2026, our article on contract meters. Its four checks turn step 1’s list into a record per contract.