Anthropic Releases Claude Opus 5.5 - Input and Output Cut 20%, Most Cybersecurity Work Rerouted to Opus 4.8

Anthropic released Claude Opus 5.5 on September 22, 2026. It is the first entry in the company’s new Claude 5.5 family, and Anthropic positions it as a model that “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5”¹. Sonnet 5.5 and Haiku 5.5 are due in the following weeks.

Two things bear directly on an adoption decision: what the price cut actually consists of, and the fact that Fable 5.1-class safeguards have reached the top Opus tier for the first time.

Last Time It Was Cache Only; This Time the Per-Token Rates Moved

With Fable 5.1 and Mythos 5.1 on September 1, input and output rates stayed put and only cache reads came down. This time the rates themselves changed.

Per 1M tokens	Opus 5.5	Opus 5
Input	$4	$5
Output	$20	$25
Cache reads	$0.20	$0.50
Cache writes	$5	$6.25

That works out to 20% below Opus 5 on input and output, and 60% below on cache reads¹. Anthropic characterizes cache reads as the bulk of what agentic and coding workloads cost, which is where the weight of this cut sits.

The headline “40% less” does not come from per-token rates alone. In the company’s own tests, the lower rates combine with fewer tokens consumed per task to net out at 40% on typical workloads at default settings¹. Put the other way, the actual saving shifts with effort settings and usage patterns. Claude Code and the Claude Platform also offer a fast mode at up to 2.5x speed, priced at $8 input and $40 output — double the standard rate.

Alongside the price cut, five-hour usage limits went up on the Pro, Max, Team, and seat-based Enterprise plans. Subscription users also receive a rate limit reset that can be banked and spent at a time of their choosing.

Cybersecurity Work Falls Back to Opus 4.8

Anthropic considers Opus 5.5 comparable to Claude Mythos 5.1 in biology and cybersecurity¹. On that basis it applied the same class of safeguards Fable 5.1 carries across three areas — cybersecurity, biology, and distillation. This is the first Opus model to ship with them, and in each area the request silently drops through to a different model.

The cybersecurity handling is where practical effects show up soonest. Locating and repairing bugs in your own code during ordinary development remains possible, but most other cybersecurity work is rerouted to Opus 4.8¹. For defenders, the company says it will soon widen its Cyber Verification Program to cover Opus 5.5, with the new program split into three tiers and including access to Mythos models.

On the biology side a route already exists. Opus 5.5 joins the Life Sciences Verification Program, which opened for applications on September 17. Vetted organizations can use it for drug discovery and research biology under relaxed safeguards.

The fallback also affects how the benchmark numbers should be read. Anthropic notes that the published scores were measured with production safeguards switched on, and that wherever those safeguards fired, Opus 4.8 handled the cybersecurity items while Opus 5 handled biology and frontier LLM development — which the company says likely drags Opus 5.5’s scores down.

The Caveats the Company Attaches Itself

Anthropic’s benchmark table puts Terminal-Bench 4.0 at 66.4% for Opus 5.5, 55.8% for Fable 5.1, and 57.9% for GPT-6 Astra¹. The table notes that the Opus 5.5 figure is at xhigh effort and the Astra figure at high effort, with the Astra number taken from OpenAI’s own reporting — OpenAI’s table at the Astra launch listed 57.7% as a best-of-any-effort score. In the same text Anthropic says that at this level of capability the margins between benchmark scores have become a weaker guide to how models differ in practice, and that its own usage shows a smaller separation between Opus 5.5 and Fable 5.1 than the table implies.

The alignment results carry a similar note. The company reports better scores than any recent Claude model on nearly every measure of misaligned behavior in an automated behavioral audit spanning close to 2,000 scenarios, while also reporting indications that Opus 5.5 frequently suspects an evaluation is underway — something it says makes real-deployment behavior harder to assess¹. It still describes the task of building evaluations that catch every failure ahead of release as an unsolved one. This is the first release since the essay calling for pacing the frontier, and outside evaluators including Frontier Design and METR examined the model before it shipped.

Conditions to Check Before Adopting It

Several delivery conditions are fixed. Opus 5.5 runs on all platforms including AWS, Google Cloud, and Microsoft Azure, and the Claude Platform identifier is claude-opus-5-5¹. Zero data retention is available. On the other hand, outputs carry a watermark under the company’s EU AI Act compliance measures, and the model can no longer be run with “thinking” turned off.

Preserved thinking, the anti-distillation safeguard introduced with Fable 5.1, applies here too. It blocks attempts to pull Claude’s reasoning out by rewriting earlier context through the API, and it covers Fable 5.1 and Opus 5.5 for API accounts created on or after August 31, 2026. The distillation activity behind it is documented in the September threat intelligence report. If you already run an integration, checking when the account was created makes triage faster.

In GitHub Copilot the model reached Pro+, Max, Business, and Enterprise users on launch day, selectable from the model picker in VS Code, Visual Studio, Copilot CLI, the coding agent, JetBrains IDEs, and other clients². Billing is usage-based at provider list pricing. The easy thing for Enterprise and Business administrators to miss is the default behavior: a new model switches on by itself unless an administrator has already turned off the global default, or has explicitly blocked that particular model. Rollout is described as gradual.

Sources

1. Introducing Claude Opus 5.5 - Anthropic official announcement (September 22, 2026)
2. Claude Opus 5.5 is now available in GitHub Copilot - GitHub Changelog (September 22, 2026)