Claude Opus 5.5 ships: Opus-tier quality at a 20% lower price

Context 1M · max output 128K · text, image, pdf · Vendor-documented

Diff vs Claude Opus 5:

Availability matrix

Field history

• 2026-09-23: claude-opus-5-5/bedrock/ap-south-1 stage — → preview (AWS What's New (demo))

Related Pulse items

• Claude Opus 5.5 reaches AWS Bedrock in Mumbai (ap-south-1), preview 23 Sept 2026

Sources: Anthropic announcement · Models overview (docs) · Pricing

Migrate one endpoint from Opus 5 to Opus 5.5 in 20 minutes

1. Search your code for thinking: {type: "disabled"} and budget_tokens; remove them.
2. Set output_config.effort explicitly on every request ("high" to match Opus 5 behavior).
3. Replace forced tool_choice with auto + strict tools + an instruction naming the tool.
4. Run your eval set on both models; compare cost per task, not cost per request.

Run it with your own API key. We never see or store keys.

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();

const msg = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 16000,
  // Thinking can't be disabled; effort is the only control.
  // Default effort drops to "medium" — set it explicitly.
  output_config: { effort: "high" },
  // Forced tool_choice returns 400: use auto + strict tools + a prompt instruction.
  tool_choice: { type: "auto" },
  tools: tools.map((t) => ({ ...t, strict: true })),
  messages,
});

Migration diff from Claude Opus 5

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();

const msg = await client.messages.create({
  model: "claude-opus-5",
  max_tokens: 16000,
  thinking: { type: "disabled" },          // allowed on Opus 5 at effort <= high
  tool_choice: { type: "tool", name: "search" }, // forced tool use
  tools,
  messages,
});

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();

const msg = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 16000,
  // Thinking can't be disabled; effort is the only control.
  // Default effort drops to "medium" — set it explicitly.
  output_config: { effort: "high" },
  // Forced tool_choice returns 400: use auto + strict tools + a prompt instruction.
  tool_choice: { type: "auto" },
  tools: tools.map((t) => ({ ...t, strict: true })),
  messages,
});

What breaks when you switch

• Requests that disable thinking fail with 400.
• Forced tool calls fail with 400.
• Latency and quality shift if you rely on the old default effort.
• Harnesses that edit earlier turns can invalidate replayed thinking blocks.

Every run publishes the exact model version and endpoint, prompts, dataset, harness code, raw outputs, cost, date and what was NOT tested.

Cost calculator

Submit a POC for Claude Opus 5.5Discussion thread for this version