1. Home
2. News
3. Anthropic

Opus 5.5 sends most cybersecurity tasks to Opus 4.8, and costs 20% less per token

Anthropic's first model since the pacing essay ships with Fable 5.1's biology and cyber safeguards, and it says the model often suspects it's being tested.

By Think Facility · September 22, 2026 · Anthropic AI products

Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in what it calls the Claude 5.5 family. Input and output tokens cost $4 and $20 per million, 20% under Opus 5.

It's also the company's first model since Dario Amodei asked the labs to slow down, and the announcement says so in its second paragraph: Opus 5.5 is "our first release since we called for pacing the frontier". Frontier Design and METR tested it before release.

Price

$4 and $20 per million, 20% less than Opus 5

Cache reads

$0.20 per million tokens, 60% less than Opus 5

Typical workloads

40% less than Opus 5 at default settings

Cybersecurity

most tasks re-routed to Opus 4.8

Next

Claude Sonnet 5.5 and Claude Haiku 5.5 in the coming weeks

Where the 40% comes from

The token price is only part of it. Anthropic says the model also gets through a task on fewer tokens, and the two together come out at 40% less than Opus 5 on typical workloads at default settings. Cache reads fall from $0.50 to $0.20 per million (most of what agentic and coding work costs is cache reads, by Anthropic's own account).

On Anthropic's own table, Opus 5.5 takes Terminal-Bench 4.0 with 66.4%, against 55.8% for Claude Fable 5.1 and 57.9% for GPT-6 Astra at the score OpenAI reported. The paragraph above that table tells you not to lean on it too hard.

Safeguards that hand the job to an older model

This is the first Opus model to launch with the class of safeguards Anthropic built for Fable 5.1, covering cybersecurity and biology, plus the anti-distillation measure it calls preserved thinking. They fall back to another model rather than refusing. Most cybersecurity tasks get re-routed to Opus 4.8, and biology work the safeguards get in the way of means applying to Anthropic's new Life Sciences Verification Program.

What I didn't expect is that those safeguards were running while Anthropic ran the benchmarks in the table above. The footnote says what happened when they fired.

Anthropic adds that this likely pulled Opus 5.5's scores down. So 66.4% is what the model got with an older Claude doing some of the work, and printing that beside your own benchmark win is, I think, the more interesting half of the table.

What the alignment section says it can't do

Anthropic runs an automated behavioral audit across nearly 2,000 scenarios (simulated ones, graded by Anthropic), and Opus 5.5 beat every recent Claude model on nearly every measure of misaligned behavior in it. On a new test of whether a model will step over a containment boundary, it tried around 85% less often than Opus 5 or Claude Mythos 5.1. Every attempt it did make was low severity, and the model flagged it itself.

Then the same section gives the limit on all of that. A model that can tell it's in a test is hard to test.

Opus 5.5 is no longer available with thinking mode switched off. Preserved thinking, which stops API users editing Claude's prior context to pull its reasoning out, applies to API accounts created on or after August 31, 2026. Claude Sonnet 5.5 and Claude Haiku 5.5 follow in the coming weeks.