Rereading the Claude Opus 5 Announcement While Using Opus 5.5
Anthropic announced Claude Opus 5 on July 24, describing it as a model that comes close to Fable 5's intelligence at half the price. The original covers performance and cost, working with the model, alignment and safety, safeguards, and getting started. Three months later, now using Opus 5.5, I follow four of those threads here: self-verification, judgment, efficiency, and fallbacks.
Before we start — key terms
The effort setting controls how much a model thinks and how many tokens it spends on a request. Higher means deeper reasoning; lower means faster and cheaper. The original's performance charts show results at each setting.
A classifier is a separate model that checks whether a request or response falls into a category such as risky cyber work. Opus 5's cyber classifiers allow finding vulnerabilities in source code but block binary-based vulnerability scanning, penetration testing, and exploit generation.
A fallback is the route taken when the original one isn't available. In the original, a request flagged by a classifier goes to another model instead of being blocked. In Claude.ai, Claude Code, and Cowork, Opus 5 falls back to Opus 4.8 by default.
A test harness is a setup that feeds inputs to code and checks the results to confirm it works. In one of the original's examples, Opus 5 built one itself when it had nothing to validate against.
First, self-verification.
Claude Opus 5 is much stronger at verifying its work and iterating carefully until it succeeds.
One example is an engineer at a trading firm who built a market data feed for a new exchange in a single session. With no live feed to validate against, Opus 5 built its own test harness to check that its code parsed the exchange's data correctly. Another customer describes it opening pages in a browser at desktop and phone widths, then catching and fixing a product hidden below the mobile fold and an off-screen checkout button.
This part has definitely gotten better since the Opus 4 versions.
There are reviews about judgment, too. This one is from an engineer:
During a rearchitecting session, Claude Opus 5 pushed back on a design I proposed, and it didn't fold when I insisted. Instead, it explained exactly what was valuable in my idea, narrowed its objection to a single design question, and proposed a compromise that kept the good part while fixing the flaw.
I had moments like this with Opus 5 too. It reminds me of the times we ended up debating. In 5.5, that side seems to have eased a bit, and I don't miss it.
On efficiency, the original says:
Claude Opus 5 provides greatly improved performance for the same cost as its predecessor, Opus 4.8.
On Frontier-Bench v0.1 it more than doubled Opus 4.8's performance at a lower cost per task, and several customers said quality held up at lower effort settings. Pricing stayed the same as Opus 4.8: $5 per million input tokens and $25 per million output tokens.
On efficiency, the 5.5 I use now feels better.
Last, the safeguards. According to the original, Anthropic intentionally avoided training Opus 5 on cyber tasks, yet as the model became more generally capable, it came close to Mythos 5 at finding vulnerabilities. So it shipped with classifiers and fallbacks.
In Claude.ai, Claude Code, and Claude Cowork, any flagged requests will fall back to Opus 4.8 by default. Fallbacks to Opus 4.8 can also be enabled on the API.
The original expected the classifiers to intervene around 85% less often than they do for Fable 5. I don't remember hitting a fallback with Opus 5, but with 5.5 I run into it often.
Looking back three months later, I think it turned out the way the announcement said.
