Claude Opus 5.5 Costs 20% Less Per Token. Here Is What It Really Saves a Business, and the Four Catches in the Small Print

On 22 September Anthropic released Claude Opus 5.5 and cut the price of its flagship model: $4 per million input tokens and $20 per million output, down from $5 and $25 on Opus 5. Anthropic also says that on typical workloads it will cost 40% less to run. Both numbers are real, and they are not the same number. If you pay for Claude through the API, or you are weighing whether a Claude-powered tool is worth it, this article settles what actually got cheaper, where the 40% comes from, what changed in the small print that can break a working integration, and what we would do this week having run AI in production for our own products.

What Anthropic Shipped on 22 September

Anthropic calls Opus 5.5 "the first model in our new Claude 5.5 family" and says it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5". Fable 5.1 is Anthropic's top model at $10 in and $50 out, so the claim is that you now get close to the top tier at well under half its price. That is Anthropic's own assessment. We cover what the benchmarks do and do not tell you further down.

• Where you can use it. In Claude's own apps, in the API as claude-opus-5-5, and on Amazon Web Services, Google Cloud and Microsoft Azure from day one.
• How much it can read and write. A 1 million token context window and up to 128,000 tokens of output per request, the same as Opus 5.
• How long it will be around. Anthropic's models page says it will not be retired sooner than 22 September 2027, so you can build on it for at least a year.
• What comes next. "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks," according to the announcement. That matters for the choice we discuss below.
• Speed. Anthropic says it "generates output more than 30% faster than Opus 5".

The New Price, Line by Line

Every figure below is from Anthropic's pricing documentation, read on 23 September 2026. Prices are in US dollars per million tokens. A token is roughly three quarters of an English word.

The line that moved most is the cache read. Normally a cache hit costs one tenth of the input price. On Opus 5.5 Anthropic set it at one twentieth, so reused context now costs 20 cents per million tokens. Anthropic's announcement says cache reads "make up the majority of agentic and coding work costs", which is why it singled that line out. For a support bot that re-reads the same product manual on every question, that is where the saving lands.

Two things did not change. Sending data through US-only processing still adds 10% to every line. And fast mode, which Anthropic says runs up to 2.5 times faster, still costs double the standard rate, and only through Anthropic's own API.

Where the 40% Comes From, and Why Your Number Will Differ

A 20% cut in the price per token cannot by itself produce a 40% cut in the bill. Anthropic explains the gap in one sentence: Opus 5.5 "costs less per token than Opus 5 and uses fewer tokens per task, which nets out to a 40% drop in costs." The first half is fixed in the price list. The second half depends on your work.

To see the split, we repriced the worked example on Anthropic's own pricing page: a one-hour agent session that uses 10,000 fresh input tokens, 40,000 tokens read from cache and 15,000 output tokens. The arithmetic is ours; the token counts and prices are Anthropic's.

On identical work, the new prices alone take about 22% off. Output tokens are 84% of that bill, which tells you where to look: the model that writes less, and thinks less, is the one that costs less. That is also where the catch sits.

• The default effort went down. Anthropic's documentation says a request that does not set an effort level now runs at medium, where Opus 5 ran at high. Part of the saving is the model thinking less by default. For most tasks that is fine. For your hardest ones, check the quality before you celebrate the invoice.
• Your tasks are not Anthropic's tasks. The 40% comes from Anthropic's tests. A job that is mostly reading, such as summarising long documents, gains less from a model that writes more briefly than a job that is mostly writing code.
• Measure cost per finished task, not per token. A cheaper model that needs a second attempt is not cheaper.

A cheaper token is only half the saving. The other half depends on how many tokens the model spends on your work, and only your own tasks can tell you that.

Four Catches in the Small Print

Anthropic's migration guide lists four breaking changes for code that already runs on Opus 5. If a developer built your integration, forward them this section. If you only use the Claude app, you can skip to the next one.

• Thinking cannot be switched off. On Opus 5 you could disable thinking for short, simple jobs to save money. On Opus 5.5 that request returns an error. The replacement is a lower effort setting. If you ran a high-volume classification or tagging job with thinking off, reprice it before switching, because it now always thinks a little.
• You can no longer force a tool call. Setting tool_choice to a specific tool, or to "any" tool, returns an error. Pipelines that pull fields out of invoices, receipts or forms often rely on exactly that. The fix is Anthropic's strict tool use or structured outputs, plus an instruction in the prompt. It is a small code change, but it has to be made.
• Some requests are answered by a different model. Opus 5.5 ships with the same class of safeguards as Fable 5.1. In Anthropic's words, "most cybersecurity tasks will be re-routed to Opus 4.8", and the cyber, biology and distillation safeguards "all ... fall back to another model transparently". On the API a declined request comes back marked as a refusal. Automatic retry on another model is an opt-in beta, and when it runs the response names the model that answered. A refusal that arrives before any output is not billed. If you run an IT or security business, test your own prompts first.
• Progress messages move. The short notes the model writes between tool calls now come back inside thinking blocks, hidden by default. Nothing fails, but an app that showed customers "checking your order..." can fall silent. There is a setting to bring those notes back.

There are two more changes, but they rarely matter to a small business. Computer-use integrations on Anthropic's API and Google Cloud must move to a new toolset. And for API accounts created on or after 31 August 2026, editing earlier turns of a conversation and replaying the model's thinking now returns an error.

Opus 5.5, Sonnet 5 or Fable 5.1: Which One to Pay For

Anthropic's own documentation now says: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads." Here is the lineup as Anthropic lists it on 23 September.

On benchmarks Anthropic reports Opus 5.5 ahead of Opus 5 across the board. On GDPval-AA, a test of real professional work across 44 occupations, it reports 1846 against 1708 for Opus 5 and 1735 for Fable 5.1. These are the vendor's numbers, run at maximum effort. Anthropic also writes that "benchmark margins have become a less reliable guide to real-world differences" and that the gap to Fable 5.1 is "narrower than these scores suggest". We agree with that caution, and we would not choose a model on a leaderboard.

Our view is simple.

• Moving up from Opus 5. Switch, after testing. It is cheaper per token and newer, and you keep the same context size.
• Paying for Fable 5.1 on everyday work. Try Opus 5.5 on the same tasks. If quality holds, that is 60% off the per-token price.
• Running Sonnet 5 at high volume. Stay put for now. Sonnet 5 is still half the price, cache reads now cost the same on both, and Sonnet 5.5 is due within weeks. Test that before you move a busy workload up a price tier.
• Moving from an older model such as Sonnet 4.6. The newer tokenizer produces about 30% more tokens for the same text, according to Anthropic. The same document costs more tokens, so compare bills, not price lists.

What Changes If You Pay $20 a Month for Claude

If you use Claude through a subscription rather than the API, you do not pay per token. Two things changed for you.

• Higher five-hour limits. Anthropic says it is "increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans". It has not published the new numbers. Some reports described the five-hour caps as scrapped. That is not what Anthropic's announcement says.
• A reset you can save. Subscribers get "a rate limit reset, which you can now save and use whenever you choose". It is useful on the day a month-end close or a proposal runs long.
• Clearer writing. Anthropic says Opus 5.5 "puts the most important information up front" and follows the writing rules you give it. For a business owner who reads its output rather than its code, this may be the change you notice first.

What We Would Do This Week

We run AI inside our own products every day, and we have changed models underneath them more than once. Each time the price list looked like good news. What the change actually did showed up in the invoice a month later. This is the order we would work in.

• Take 30 to 50 of your real tasks and run them on both models. Record the cost per finished task and how often a person had to fix the result. That is your real saving, not ours or Anthropic's.
• Try low effort before medium. Anthropic's own guidance is to step down where quality holds. Many routine jobs, such as drafting replies, tagging tickets and summarising calls, never needed high effort.
• Search your code for the two errors. Look for thinking set to disabled and for forced tool_choice. Both now return an error on Opus 5.5, and it is better to find them in testing than in production.
• Turn prompt caching on. If your instructions and reference material stay the same between calls, reused context now costs 5% of the input price.
• Handle the refusal. Decide what your app does when a request is declined, whether that is retry on another model, hand it to a person or explain it to the customer, before a customer finds out for you.
• Know which model you are on. Claude model names are fixed snapshots, so nothing moves you to Opus 5.5 until someone changes claude-opus-5 to claude-opus-5-5. Opus 5 remains available as a legacy model while you test.

If You Would Rather Not Track Model Prices at All

Model prices now move every few weeks. We tracked the next scheduled change in our post on what doubles on 1 January 2027. If you are using Claude directly, our guide to Claude for Small Business shows what the ready-made workflows do and what they cost. If you are deciding whether to build something yourself on a model like this, our build-vs-buy post covers when that works and when it does not.

For many businesses the simpler answer is a flat monthly price, where changing the model underneath is someone else's job. Our AI Chat Assistant answers website visitors around the clock for a fixed monthly fee, runs on four AI engines with built-in redundancy, and hands conversations to your team on live chat or WhatsApp. Your price does not depend on what any single model costs this month. If you want a second opinion on an AI bill you are already paying, talk to us.