Claude Opus 5.5 Is a Pricing Story About Agents

Anthropic released Claude Opus 5.5 on Tuesday, with the usual wall of benchmark charts. I'm going to skip them. I use Claude Code every day, and what I care about is what a model costs to run in a loop, and what breaks when I switch to it. On both, this release says more than the charts do.

The price card, read as an agent

An agent session isn't one prompt and one answer. It's dozens of turns, and on every turn the model re-reads the same large prefix: system prompt, tool definitions, the files it has looked at, the conversation so far. With prompt caching, that re-reading is billed at the cache read rate, not the full input rate. For an agent, the cache read price is often the line that matters most.

Anthropic's docs list Opus 5.5 at:

	Opus 5	Opus 5.5
Input (per million tokens)	$5	$4
Output	$25	$20
5-minute cache write	$6.25	$5
Cache read	$0.50	$0.20

Input and output dropped 20%. Cache reads dropped 60%, to 0.05x the input price (on Opus 5 they were 0.1x). There's also a 1-hour cache write at $8, batch processing at half price, and a fast mode at $8 / $40 per million.

Anthropic's headline is that "at default settings it will cost 40% less than Opus 5 on typical workloads." Note the "at default settings." That's the next part.

Thinking is always on, and the default changed

Two changes affect the bill without touching the price card:

• Thinking can't be disabled. On Opus 5 you could turn thinking off at lower effort. On Opus 5.5, sending thinking: {"type": "disabled"}, or a manual thinking budget, returns a 400 error. The control is now the effort parameter.
• The default effort dropped from high to medium. And Anthropic's docs say that at the same effort level, Opus 5.5 "tends to think more per turn" than Opus 5, most of all at xhigh and max.

Thinking tokens are billed as output, at $20 per million. So here's an illustrative example (my arithmetic on list prices, not a measurement). Take an agent session of 40 turns, each re-reading a 60,000-token cached prefix:

• Cache reads on Opus 5: 2.4M tokens × $0.50 = $1.20
• Cache reads on Opus 5.5: 2.4M tokens × $0.20 = $0.48
• But if each turn also thinks for 2,000 tokens, that's 80,000 output tokens × $20 = $1.60

The cache savings are real, but in a long agent loop the thinking can outweigh them. And since the "40% less" is measured at the new, lower default effort, I wouldn't assume the whole 40% if you pin effort to high or above.

Your conversation has to be append-only

The change I'd most want to know about before migrating anything: thinking blocks are now tied to the conversation that produced them. The API checks whether anything before a thinking block (the system prompt, the tools, an earlier message) has changed since it was produced. For accounts created on or after August 31, 2026, replaying a block after such a change returns a 400.

Anthropic's advice is to keep the conversation append-only: change instructions or tools with mid-conversation system messages rather than by editing the prefix. A new beta even lets you add or change a tool definition mid-conversation "without losing the prompt cache."

With my systems hat on, the cheapest, most correct way to run an agent now looks a lot like event sourcing: never edit history, only append to it. A stable prefix makes caching cheap, and it now also keeps the API from rejecting your request. Cost and correctness point the same way.

Two smaller breaking changes: forced tool use is gone (tool_choice of any or tool returns a 400; use auto plus strict tool use), and text the model writes between tool calls now arrives in thinking blocks, so a UI that streams those progress notes can go quiet.

Cyber questions go to an older model

Anthropic says Opus 5.5 has "extremely strong cyber capabilities," so it ships with safeguards similar to Fable 5.1's. Routine work is fine: you can "identify and fix bugs in their code as part of the routine software development lifecycle." But "most cybersecurity tasks will be re-routed to Opus 4.8," and a separate verification program for defenders is being expanded.

For anyone building on the API, that means a request can be answered by a different model than the one you asked for, or declined with stop_reason: "refusal". The docs describe a server-side fallback option (in beta) and say a refusal still counts against your rate limits. If your app touches security in any way, handle the refusal path explicitly and log which model answered. Don't assume one model ID means one model's behaviour.

What I want to check

I haven't measured any of this, so this is a try-list:

1. Cost per session before and after, on the same task, at the same explicit effort level, not the default.
2. How much my cached prefix actually changes between turns. Anything that edits it is now a cost and a 400 risk.
3. Whether medium is enough for multi-file changes, or whether I end up at high anyway.

Quick hits

• OpenAI ships GPT-6 Sol and Luna (Sep 22). About 90 minutes after Opus 5.5, OpenAI released two cheaper GPT-6 models for ChatGPT Work, Codex and the API, at about half the price of the GPT-5.6 versions. OpenAI credits caching and inference improvements for the cut. Both launches led with price on the same day.
• An AI memory library shipped a worm (Sep 23). Malicious versions of MemTensor's MemoryOS on PyPI and its OpenClaw plugin on npm ran a credential stealer that hunted for npm and PyPI publish tokens and Hugging Face tokens. The Python payload ran on import, not on install, so --ignore-scripts-style defences didn't help. Semgrep says to pin to the last clean versions and rotate credentials.
• Claude subscribers get higher five-hour limits (Sep 22). Alongside the launch, Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and added a rate-limit reset you can save for later. That's separate from the weekly-limit change The Decode #009 covered.

What I'm watching

Whether Anthropic's smaller 5.5 models, which it says are coming "in the coming weeks," get the same cache pricing.

Sources

Claude Opus 5.5

• Anthropic, Claude Opus 5.5 announcement (Sep 22, 2026)
• Claude Platform docs, What's new in Claude Opus 5.5
• Claude Platform docs, Opus 5.5 migration guide
• Silicon UK, coverage of the Opus 5.5 release (Sep 25, 2026)

Quick hits

• OpenAI, GPT-6 Sol and Luna announcement (Sep 22, 2026)
• TechCrunch, OpenAI launches GPT-6 Sol and Luna (Sep 22, 2026)
• Semgrep, analysis of the MemTensor compromise (Sep 23, 2026)
• Anthropic, Claude Opus 5.5 announcement (usage limits, Sep 22, 2026)