Opus 5.5: what changes if you code with an AI agent

You’re deep into a long session with your coding agent. It’s on turn thirty, every reply takes almost a minute, and out of the corner of your eye you watch the spend counter climb. Then you hear that Claude Opus 5.5 is out, and that according to Anthropic it performs like Fable 5.1, their largest model, costs 40% less than Opus 5, and writes 30% faster. The obvious question is whether to switch today.

The short answer is yes, except in three specific cases. Below are the numbers for a real session at the old and new prices, when fast mode is worth paying for, what those three cases are, and how to check it on your own repository. Switching models is one line; testing it properly takes an afternoon.

What exactly came out

Opus 5.5 was released on September 22 and is the first of the 5.5 family: Sonnet 5.5 and Haiku 5.5 arrive “in the coming weeks.” It’s available on Anthropic’s API and on the three big clouds (AWS, Google Cloud, and Azure).

It helps to keep two layers apart. Opus 5.5 is the language model: the thing that reads and writes. Claude Code, Cursor, or Copilot are coding agents: the program around it, which gives it access to your files and decides which commands it can run. Changing the model doesn’t change the agent. Your permissions, your project instructions, and your shortcuts stay the same; what changes is who’s doing the thinking. In Claude Code you pick it with the /model command, and in other agents it usually shows up in the model picker within days.

Price: what you use most got the biggest cut

A model charges by tokens, chunks of text about four characters long, and it charges separately for what goes in and what comes out. The new numbers, in dollars per million tokens:

	Opus 5	Opus 5.5	Change
Input	5	4	−20%
Output	25	20	−20%
Cache write	6.25	5	−20%
Cache read	0.50	0.20	−60%

The row that matters is the last one. The cache is the provider’s memory: when you send the same block of text on every turn (the project instructions, the files it already read), it stores it pre-processed and charges you much less to reuse it. In the article on overspending, a cache read cost a tenth of the normal price. With Opus 5.5 it costs a twentieth.

That hits a coding agent squarely, because an agent re-reads almost everything on every turn. An example with round numbers: a forty-turn session where each turn resends 55,000 tokens that were already cached, adds 5,000 new ones (the file it just opened, the output of a test), and replies with 1,500.

Per session	Opus 5	Opus 5.5
Cache reads (2.2 million)	1.10	0.44
New input (200,000)	1.25	1.00
Output (60,000)	1.50	1.20
Total	3.85	2.64

The same session costs 31% less from the price change alone. The 40% Anthropic advertises comes from adding something else: the model finishing the task with less back and forth. That’s no longer a price, it’s a promise, and you measure it on your own repository (how is further down).

Also notice how the new bill breaks down. With Opus 5, re-reading the cache was almost a third of the bill. With Opus 5.5 it drops to a sixth, and output becomes the most expensive line: nearly half. The advice that pays off most now is to ask the agent for just the lines that change, not the whole file rewritten, because every extra line it writes is billed at the highest price in the table.

If you use Claude through a monthly subscription instead of paying per token, this table doesn’t touch you directly. Your news is different: you can now save a usage-limit reset and spend it whenever you want. For example, keep it for the day of a big migration instead of running out of quota at four in the afternoon.

Speed: fast mode and when to pay for it

Opus 5.5 generates text more than 30% faster than Opus 5 out of the box. On top of that there’s fast mode, available in Claude Code and the API, which reaches up to 2.5 times that speed. It costs double: 8 dollars for input and 40 for output.

There’s a simple way to decide: are you watching the screen while it answers?

• Yes: you’re debugging, the agent proposes a change and you wait to approve or correct it. Every reply that drops from fifty seconds to twenty is your time, and your hour is worth more than the price difference. Fast mode pays for itself here.
• No: you asked it to document a module, write tests for twenty files, or run a long task you’ll review later. Nobody is waiting for that answer, so paying double to get it sooner makes no sense.

One comparison worth keeping in mind: fast-mode output (40 dollars) costs more than regular Opus 5 output (25). Leave it on all day and the savings from the previous section are gone.

Three changes that might break something

For most people, switching models breaks nothing. These are the three cases where it can, and they’re worth checking first.

1. Reasoning can no longer be turned off

Current models can reason before answering: they write intermediate steps you don’t see directly, which improve the answer and are billed as output tokens. Up to Opus 5 you could disable it for mechanical tasks, where thinking adds nothing. Opus 5.5 no longer offers that option.

It only affects you if you built something of your own on the API with reasoning turned off: a script that classifies tickets, one that translates configuration files. What you can do is lower the effort, a parameter that goes from low to max and controls how much it reasons. For mechanical work, low is usually enough. If the numbers still don’t work, that job probably belongs to Haiku 5.5 once it’s out, not to an Opus.

2. Conversation history can no longer be rewritten freely

Opus 5.5 ships with a safeguard Anthropic calls preserved thinking: it prevents editing earlier turns of the conversation to extract the model’s reasoning. It applies to API accounts created on or after August 31, 2026.

If you use Claude Code, Cursor, or any off-the-shelf agent, you won’t notice. If you wrote your own agent and it modifies old messages in the history (for instance, to trim or summarize them), test it with the new model before switching in production, because that kind of history manipulation is exactly what the safeguard restricts.

3. Security tasks are answered by a different model

Opus 5.5 is very capable at computer security, and Anthropic gave it the same controls as Fable 5.1. Finding and fixing bugs in your code, as part of normal work, still works the same. But most cybersecurity tasks are rerouted to Opus 4.8, an earlier model. For example, asking it to set up a penetration test (deliberately attacking your own system to find a way in) or to explain how to exploit a vulnerability in a dependency.

If that’s your line of work, the answers won’t improve with the switch. Anthropic announced it will add Opus 5.5 to its verification program for defense teams, but that isn’t available yet.

How to test it on your repository

Published benchmarks are useful for getting your bearings. Terminal-Bench 4.0, for instance, measures how many real terminal tasks an agent completes end to end: Opus 5.5 scores 66.4%, and according to Anthropic it matches GPT-6 Astra at about 40% of the cost. But those tasks aren’t yours. Your repository has its conventions, its slow tests, and its odd corners, and that’s what decides whether the model works for you.

The test I recommend takes an afternoon:

1. Pick four or five tasks you already solved, where you have the commit with the fix. Mix sizes: a one-line bug, a new function, a change that touches several files.
2. Go back to the commit before each one and give the task to the agent, with the same instruction, once with Opus 5 and once with Opus 5.5.
3. Write down three things per attempt: whether the tests pass, how many times you had to correct it, and what it cost. In Claude Code, the /usage command shows what the session has used.
4. Compare per finished task, not per request. A model that’s cheaper per token but needs two more attempts isn’t cheaper.

While you’re at it, pay attention to something no table shows: how it writes to you. A common complaint about Opus 5 was that its replies were hard to follow. Anthropic says Opus 5.5 puts the important part first and uses less jargon, and early testers agree. If that holds, you’ll notice it in the time you spend reviewing what it did, which is usually the biggest cost of all and never shows up on the bill.

Switch today or wait for Sonnet 5.5

If you use Opus 5 today, switch. It’s cheaper per token, much cheaper over long sessions thanks to the cache, faster, and, by everything published so far, as good or better in quality. The only reasons to wait are the three cases above, and it’s easy to tell whether you’re in one.

If you use Sonnet 5 because Opus was too expensive, the math has changed but it hasn’t flipped. Opus 5.5 still costs twice as much per token as Sonnet 5, and Sonnet 5.5 arrives in a few weeks with its own price cut. That’s when to run the five-task test on both and keep the one that finishes the job cheapest, not the one with the lowest rate.