Artificial intelligence
Anthropic's New Flagship Beats Its Biggest Model for Less. For Cyber Work It Hands You an Older One
Claude Opus 5.5 outscores the Mythos-class Fable 5.1 on most of Anthropic's published benchmarks while costing 20% less per token than Opus 5. The detail further down the announcement matters more: most cybersecurity requests are re-routed to Opus 4.8, so the model string you call is no longer a promise about which model answers.
MAI
Anthropic released Claude Opus 5.5 on Tuesday morning, and the release has an unusual shape. On most of the benchmarks the company published, the model beats Fable 5.1 — Anthropic's largest and most expensive model, sitting a tier above it — while costing 20% less per token than the Opus 5 it replaces. That is the headline, and it holds up. The more consequential detail sits further down the announcement, in the safeguards section: if you send Opus 5.5 a cybersecurity task, Opus 5.5 is not what answers it.
A smaller model eating a bigger one
Anthropic's published comparison puts Opus 5.5 ahead of Fable 5.1 across agentic coding, computer use and knowledge work.
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — |
| GDPval-AA v2.1 | 1846 Elo | 1735 | 1708 | 1542 |
| Humanity's Last Exam | 67.7% | 65.6% | 63.6% | 57.2% |
| OSWorld 2.0 | 81.8% | 80.7% | 74.0% | — |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% |
Two of those rows are losses to OpenAI's Astra, and Anthropic printed them anyway — a six-point deficit on Terminal-Bench-Science is not the kind of number a vendor includes by accident. That restraint is worth noting, because it makes the rest of the table easier to credit. The eleven-point jump on Terminal-Bench over Fable 5.1 is the result that matters commercially: the tier boundary between Opus and the Mythos-class models has stopped meaning what it meant four months ago.
The price is the argument
| Per million tokens | Opus 5.5 | Opus 5 |
|---|---|---|
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache reads | $0.20 | $0.50 |
| Cache writes | $5 | $6.25 |
| Fast mode input / output | $8 / $40 | — |
The sticker cut is 20%, but Anthropic's claim is larger: "Claude Opus 5.5 leads in agentic coding and knowledge work, and costs 40% less to run than Opus 5 on typical workloads," on the argument that the model spends fewer tokens reaching an answer. Box, quoted in coverage of the release, reported roughly a third of Opus 5's token consumption in its own evaluations with output 40% less verbose. The New Stack's write-up puts Opus 5.5 ahead of GPT-6 Astra at around 20% of the per-task cost.
For anyone running agents at volume, verbosity is the whole bill. A 20% rate cut is a line item; a two-thirds reduction in tokens spent per completed task changes which workloads are viable at all.
The model you call is not always the model that runs
Here is the part that deserves more attention than it will get. Anthropic's announcement states plainly:
Users will be able to identify and fix bugs in their code as part of the routine software development lifecycle, but most cybersecurity tasks will be re-routed to Opus 4.8.
Not refused. Re-routed — to a model two releases old. The same mechanism arrived with Fable 5.1 earlier this month, which redirected penetration testing, exploit generation and binary-based vulnerability scanning to Opus models rather than declining them. Opus 5.5 inherits it, and Anthropic says the Cyber Verification Program will be expanded to cover the new model for defenders who apply and are approved.
The New Stack draws out the implication: a request sent to Opus 5.5 may be handled by Opus 4.8 or Opus 5 instead, depending on whether the safeguards fire. For a chat session that is a curiosity. For a multi-step agent, it means a single run can silently span three model versions with different capabilities, different failure modes and different costs, and the developer has no explicit signal about where the seam fell. Anyone building evaluation harnesses on top of the API now has a variable they cannot see.
Capability is being unbundled from access
Read the safeguards together and a pattern emerges. Biology capability is gated behind the Life Sciences Verification Program. Cyber capability is gated behind the Cyber Verification Program. Distillation is blocked by "preserved thinking," the anti-distillation measure introduced with Fable 5.1, applied to API accounts created after 31 August 2026. And Opus 5.5 "is also no longer available with 'thinking' mode switched off" — a configuration option removed rather than deprecated.
What a model can do and what your account can reach are no longer the same question. The frontier is being gated per customer, not per model, and the API string has become a request rather than a specification. That is a defensible response to the misuse Anthropic has publicly dealt with this year, and it is also a real change in what buying access to a model means. The benchmarks will be reproduced by third parties within the week. The routing behaviour is harder to measure from outside, and it is the thing that will shape how these models are actually used.
Sources: Introducing Claude Opus 5.5 (Anthropic) · Anthropic releases Opus 5.5 with lower prices and Fable-level performance (TechCrunch) · Anthropic releases Opus 5.5 and cuts pricing by 20% (The New Stack) · Google, Anthropic, and OpenAI Unveil Cyber AI Models, Safeguards, and Access Programs (The Hacker News)