Anthropic dropped Claude Opus 5.5 on 22 September 2026, and the number that matters here is the price tag, not the benchmark chart. Input tokens fall from $5 to $4 per million, output from $25 to $20, and cache reads drop 60%, from $0.50 to $0.20. Anthropic says real workloads run about 40% cheaper overall. If you're paying anyone by the token for AI work, that's the line worth reading twice.
What Changed in Claude Opus 5.5 Pricing#
The Claude Opus 5.5 pricing table is the whole story here, per million tokens:
- Input: $4, down from $5
- Output: $20, down from $25
- Cache reads: $0.20, down from $0.50 (the one that matters most for coding agents and document pipelines that re-read the same context over and over)
- Batch: $2 and $10, down from $2.50 and $12.50
Alongside the cut, the model ships a 1 million token context window, adaptive thinking that can no longer be switched off, and real benchmark gains: SWE-bench Pro moved to 89.9% from 79.2%, Terminal-Bench 4.0 to 66.4% from 52.3%, per the numbers Anthropic published and llm-stats.com reported the same day. Full specs sit in the model documentation if you want to check the fine print yourself.
Does This Change Your WhatsApp or Chatbot Costs?#
For most small businesses, no. If you're running a WhatsApp FAQ bot or a basic lead-qualification flow, you're almost certainly already on a smaller, cheaper model, not a frontier tier like Opus. This release doesn't touch that line on your bill.
It matters if you're paying for something that leans on long context or real reasoning: an agentic coding tool, a pipeline that re-reads the same 50-page contract on every turn, or an internal assistant planning multi-step work rather than answering one question. If that kind of workload already runs a few hundred dollars a month, this shifts it by something close to a third.
The Honest Trade-Offs#
The price cut is real and the benchmark gains are genuine. That's the good part.
There's a catch, though. Adaptive thinking is now mandatory. You can't switch it off, and a request that tries gets a 400 error back. Fine for most workloads, but it removes a lever some teams used to hold down cost on simple calls where thinking added nothing but tokens. Default reasoning effort also dropped from "high" to "medium," which Anthropic frames as sensible and I'd mostly agree, though older prompts tuned around the previous default may now behave slightly differently.
Anthropic's own materials note that benchmark margins at this tier can overstate real-world differences. I'd take that seriously. Going from 79% to 89% on SWE-bench Pro looks dramatic on a chart. Whether you feel it depends entirely on what you're actually building.
How We're Already Using Opus 5.5#
At WebEpex we run a managed AI DevOps retainer for ChainBox.ai, a crypto trading platform client, priced around $2,000 a month, built around Claude Code doing codebase context work and automated error detection on their trading engine. That retainer re-reads the same repository on nearly every session, so cache-read pricing is the line we watch closest, not the headline input or output rate.
We moved that retainer's model config to Opus 5.5 within the week and ran it alongside the old setup for a couple of billing cycles before cutting over fully. The cache-read discount alone justified the switch before we'd even weighed the benchmark gains.
We'd assumed the benchmark gains would show up clearly on our own coding workload too. On this specific client's codebase they were smaller than the published numbers suggested, which is worth knowing before anyone sells you the switch purely on the score sheet. We're running the same review on every other build touching a frontier-tier model this month, the same way the Fugu Max routing release and the DeepSeek V4.1-Flash pricing cut made us re-check things a few weeks back.
What Should You Do This Week?#
Running simple WhatsApp or FAQ automation? Nothing. Genuinely, close the tab.
If you or your agency run anything on Claude Opus for coding, document work, or agentic reasoning, ask one plain question: is the model ID still pinned to claude-opus-5, or has it already moved to claude-opus-5-5? For most setups that's a config change, not a migration project, and there's little reason to keep paying the old rate once you've confirmed your workflow behaves the same on the new default.
We had a version of this same conversation with a client about GLM-5.3-Flash's pricing move going the opposite direction this month. The lesson holds either way: know which model you're actually paying for, and check again every time a vendor changes the number.
If you want a second pair of eyes on whether this actually touches your setup, send me what you're running and I'll tell you straight, no pitch attached. cal.com/webepex/growth-review
