Claude Opus 5.5 for Business: The $20 Model Changes the Math

Claude Opus 5.5 is Anthropic's new daily-driver model, released September 22, 2026, at $4 per million input tokens and $20 per million output tokens. That is 20 percent below Opus 5 and 60 percent below Fable 5.1 on list price, and Anthropic says it performs at Fable 5.1's level on most work. For most businesses, it turns July's question of which Opus tier to run into a simpler one: run Opus 5.5 unless a specific task proves it needs Fable.

This refreshes our July post on which Opus tier to actually run. The answer moved because the price and the lineup moved. What did not move is how we decide, so most of this post is about the four lines in the announcement that a business will skim past and should not.

How does Opus 5.5 compare with the rest of the Claude lineup?

Here is the current list pricing from Anthropic's pricing docs, read on September 28, 2026. All figures are per million tokens.

Two notes on reading that table honestly.

First, the headline savings. Anthropic says Opus 5.5 costs 40 percent less to run than Opus 5. The per-token price cut is 20 percent. The rest of that 40 percent is token efficiency that Anthropic reports from its own tests, meaning the model uses fewer tokens to get the same work done. That may well hold for you, but you cannot read it off a price list. You find out by running your own work through it.

Second, Mythos 5.1 is on the pricing page, but it is offered by invitation through Anthropic's Project Glasswing, not to everyone. If you are planning a budget, plan around the models you can actually call.

Who actually gets Opus 5.5?

Availability is the part people get wrong most often after a launch, so here it is from the vendor pages:

• API and clouds. Opus 5.5 is available on the Claude API with the model ID claude-opus-5-5, and Anthropic says it is on Amazon Web Services, Google Cloud, and Microsoft Azure.
• Paid Claude plans. Anthropic raised five-hour usage limits on its paid plans with this release. The consumer pricing page lists Opus for Pro, Max, Team, and Enterprise.
• Free plan. The pricing page lists Sonnet and Haiku only. If your team is testing on free accounts, it is not testing Opus.
• Fast mode. Faster output is available as a research preview on the Claude API only, not on partner clouds, at $8 input and $40 output per million tokens for Opus 5.5.

Before you tell a team they have the new model, check which plan and which platform they are actually on.

Why does always-on thinking matter for a business workload?

This is the first line most readers skip. Anthropic's announcement says Opus 5.5 is no longer available with thinking switched off. Every request gets the model's extended reasoning, whether the task needs it or not.

For the work Opus is built for, that is the point. An agent that plans a change across a codebase, reconciles a messy document set, or works through a multi-step support case benefits from thinking before it acts. That is where the quality comes from.

For short, high-volume jobs, it is a tax. Consider the kind of work that runs thousands of times a day:

• classifying an inbound email into one of six queues
• cleaning up a mailing address
• answering yes or no on whether a ticket mentions a refund
• rewriting a product title to a house style

An always-thinking model pays for deliberation nobody asked for on every one of those, in both latency and output tokens. Those jobs still belong on Sonnet 5, now permanently at $2 and $10 (Anthropic dropped the planned September 1 increase to $3 and $15), or on Haiku 4.5 at $1 and $5. Our model org chart walks through how we assign work across tiers, and nothing about this release changes the bottom of that chart.

What do the cyber and biology safeguards mean for your team?

This is the second line to read twice. Opus 5.5 launched with safeguards similar to Fable 5.1 on cybersecurity, biology, and distillation. In practice:

• Most cybersecurity tasks are rerouted to Opus 4.8. If your security team benchmarks Opus 5.5 on its own work, such as reviewing code for vulnerabilities or analyzing an exploit, it may be measuring a different model than it thinks. Anthropic says it will expand its Cyber Verification Program to include Opus 5.5 for verified cyberdefenders.
• Biology research that runs into the safeguards goes through Anthropic's Life Sciences Verification Program.
• Preserved thinking carries over from Fable 5.1 as an anti-distillation safeguard, and the release includes watermarking measures Anthropic describes as complying with the EU AI Act.

None of this matters for a law firm drafting letters or a retailer writing product copy. It matters a great deal for a security team, and it is exactly the kind of detail that gets lost between the announcement and the procurement decision.

How would we decide whether to move a workload?

We published our rule for when to switch models earlier this month in how often to switch AI models. The trigger is never a launch post. It is one of three things: the pricing page changes, a deprecation date appears, or your own test set says the new model does the job better. This time the pricing page moved, so the next step is to run the test set.

The line that decides the bill for us is the one most people ignore: cache reads. In our own agent work, which we wrote up in two years of building with AI, by the usage logs, cache reads are 97 to 98 percent of token volume. Agents reread the same context, instructions, and files over and over, and that reread is cached.

Look at the cache-read column again with that in mind. Opus 5.5 charges $0.20 per million cached tokens. Opus 5 charges $0.50. Fable 5.1 charges $0.25. For a workload shaped like ours, the cache-read price moves the monthly total more than the headline input and output prices do. A 20 percent cut on output looks nice. A 60 percent cut on the line that carries almost all the volume is the story.

Here is how we would sort a typical set of workloads:

• Move first: agent and coding workloads with heavy cache reuse that currently run on Opus 5. Same tier, lower price on every line, and the biggest cut lands where the volume is. Run the test set, then switch.
• Test against Fable: the few tasks where you pay for Fable today because nothing else cleared the bar. If Opus 5.5 clears it on your test set, that is 60 percent off the per-token price for the same outcome.
• Leave alone: short, high-volume jobs on Sonnet 5 or Haiku 4.5. Always-on thinking makes Opus 5.5 the wrong tool there.
• Do not touch without verification: security tooling. Confirm which model is actually answering before you trust a benchmark.

If you want help working through which of your workloads fall where, that is the kind of AI integration for business work we do. And if you are an individual trying to get more out of Claude day to day, this change mostly means the Opus you reach for on a paid plan got better value.

FAQ: Claude Opus 5.5 for business

Is Opus 5.5 cheaper than Fable 5.1? Yes, on list price: $4 and $20 per million input and output tokens against $10 and $50, which is 60 percent lower. Anthropic says it performs at Fable 5.1's level on most work. Test the tasks you currently run on Fable before you move them.

Can I turn thinking off? No. Anthropic says Opus 5.5 is no longer available with thinking switched off. For short, high-volume tasks, use Sonnet 5 or Haiku 4.5 instead.

Is Opus 5.5 on the free Claude plan? The consumer pricing page lists Sonnet and Haiku for the Free plan. Opus starts at Pro.

Should we switch today? Run your own test set first. The pricing page changed, which is a reason to test, not a reason to switch blind. Our switching rule lays out the steps.

The short version

Opus 5.5 makes the middle of the Claude lineup cheaper and better, and for agent work the cache-read price is the number that matters most. Move the cache-heavy Opus work after a test run, check whether your Fable work still needs Fable, leave the small fast jobs where they are, and verify anything security-related before you trust it.

If you want a second set of eyes on which workloads to move and which to leave, talk to us.