Skip to content
chornous.dev

Type two or more letters. Esc closes.

Index of sheets
Theme

Sheet 03 · Writing

All notes

Claude Opus 5.5 from a front-end desk: price, speed and the agent parts

Anthropic's September model claims Fable-level work at 40% lower cost than Opus 5. Notes on long codebase migrations, prompt caching and agents that read untrusted pages.

Pl. 59 · network drawing generated from the slug “claude-opus-5-5-for-developers”

Anthropic released Claude Opus 5.5 on September 22, 2026. Anthropic says it "performs at the level of Claude Fable 5.1 on most work" and, at default settings, costs 40% less to run than Opus 5 on typical workloads. Fable 5.1 and Mythos 5.1 came out three weeks earlier, on September 1.

cost vs Opus 5, typical workloads at default settings
−40%
faster output generation than Opus 5
>30%
Terminal-Bench 4.0
66.4%
FrontierCode v1.1
54.4%

It's available on the Claude Platform, AWS, Google Cloud and Microsoft Azure under the model id claude-opus-5-5.

Pricing

Opus 5.5 pricing from the announcement
Token typePrice per 1M tokens
Input$4
Output$20
Cache writes$5
Cache reads$0.20

For code work, compare input with cache reads: 20 to 1. An agent working through a migration rereads the same files, types and instructions many times. Put that stable context at the front of the request and mark it for caching, and most of what you pay for is output.

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const message = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 4096,
  // Stable context first, marked for caching; the task changes per call.
  system: [
    { type: "text", text: repoContext, cache_control: { type: "ephemeral" } },
  ],
  messages: [{ role: "user", content: "Migrate Button.tsx to the new API." }],
});

A rough example, with my arithmetic: a session that reads 1.5M tokens from cache, sends 0.5M fresh input tokens and writes 200k output tokens costs about $0.30 + $2.00 + $4.00, a little over six dollars, plus the one-time cache write. Without caching, the same 2M input tokens would cost $8 before any output.

Benchmarks

Anthropic lists agentic coding, codebase-wide migrations, knowledge work and computer use as the model's strengths. Its numbers: 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, an Elo of 1846 on GDPval-AA v2.1, and 81.8% on OSWorld 2.0, a result Anthropic's table marks as partial. I'd use them to decide what to try first. Then I'd run a migration in my own repository, with my lint rules and my review bar.

The agent parts

Two details apply once the model runs as an agent.

Preserved thinking. Opus 5.5 ships with an anti-distillation safeguard that stops API users from editing Claude's earlier context to extract its reasoning. Anthropic says it applies to accounts created on or after August 31, 2026, and to Fable 5.1 as well. If your integration rewrites earlier assistant turns, test it against the new behavior before you switch. Anthropic's September threat intelligence report covers the distillation activity it has detected and disrupted.

If you run Opus 5 for code today, move one migration with a clear before and after to 5.5, then compare the diff quality and the invoice before you move the rest.

Sources

Volodymyr Chornous