WED, SEPTEMBER 23, 2026
Independent · In‑Depth · Practitioner‑Tested
Claude Coding

Claude Opus 5.5 Migration Prompts: Find What Breaks Before Production Does

Opus 5.5 introduced four breaking changes from Opus 5: thinking cannot be disabled, forced tool use returns a 400, thinking blocks are bound to the model and conversation so replays fail, and the old computer-use tool is rejected. These seven prompts walk a codebase looking for each one, then help you measure whether the 40% running-cost claim holds on your traffic. Run them against your own repository before you change any model string.

⌨️ 7 prompts 🕐 Updated Sep 23, 2026
💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.
1
Find Disabled Thinking
Breaking change 1. Explicitly disabled thinking must be removed before the model string changes.
Search this codebase for every call to the Anthropic Messages API.<br/><br/>For each one, report whether the thinking parameter is set, and if so to what value.<br/><br/>Output a table with three columns: file path, line number, current thinking configuration.<br/><br/>Flag every call that explicitly disables thinking. Opus 5.5 does not allow it - the parameter must be omitted or set to adaptive mode.<br/><br/>Do not change any code. Report only.
2
Find Forced Tool Use
Breaking change 2. tool_choice set to any returns a 400 on Opus 5.5.
Search this codebase for every occurrence of tool_choice in an Anthropic API call.<br/><br/>Report each one with its file path, line number, and the value it is set to.<br/><br/>Flag every instance set to "any". Opus 5.5 rejects that value with a 400 error.<br/><br/>For each flagged instance, describe in one sentence what the code was trying to guarantee by forcing a tool call, so I can decide on a replacement. Do not propose the replacement yet.
3
Find Replayed Thinking Blocks
Breaking change 3. The one most likely to pass code review and fail in production.
This is the highest-risk check. Read carefully before answering.<br/><br/>Find every place in this codebase where assistant messages are stored, serialised, cached, or read back and resubmitted to the Anthropic API.<br/><br/>For each one, answer: does the stored payload include thinking blocks?<br/><br/>Opus 5.5 binds thinking blocks to a specific model and conversation. Replaying a stored block from another model or another conversation fails.<br/><br/>List every code path where a stored assistant turn could be resubmitted. Include conversation-resume features, retry logic, test fixtures and any message-history table.
4
Find the Old Computer-Use Tool
Breaking change 4. A simple string search, but easy to miss in config files.
Search this codebase and all configuration files for the string computer_20251124.<br/><br/>Report every occurrence with file path and line number.<br/><br/>Opus 5.5 rejects that tool version. The replacement is computer_toolset_20260801.<br/><br/>Also check for any hardcoded tool-version strings that follow the same date-suffix pattern, in case other tools are pinned to versions that will need updating.
5
Build the Migration Checklist
Turns four separate audits into one ordered plan with verification steps.
You now have the results of four audits: disabled thinking, forced tool use, replayed thinking blocks, and the old computer-use tool.<br/><br/>Produce a single ordered migration checklist.<br/><br/>Sort by risk, highest first. For each item give: the file and line, what breaks, the specific change to make, and how to verify the fix.<br/><br/>Put anything involving stored or replayed assistant turns at the top regardless of count, because those failures appear only on specific runtime paths.<br/><br/>End with a list of the tests I should write before changing the model string.
6
Measure the Real Cost Change
The 40% claim is workload-dependent. This produces the number for your pipeline instead.
Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads. The list price only fell 20% - from $5 in and $25 out to $4 and $20. The rest comes from cache reads at $0.20 per million (down 60%) and roughly 30% faster output generation.<br/><br/>I will paste a sample of my production API logs below.<br/><br/>Calculate: total input tokens, total output tokens, and what fraction of input tokens were cache reads.<br/><br/>Then compute my actual spend under Opus 5 pricing and under Opus 5.5 pricing, and give me the real percentage change for this traffic.<br/><br/>Do not use the 40% figure. Derive it.<br/><br/>LOGS:<br/>[paste]
7
Decide Whether to Wait for Sonnet 5.5
Stops a migration that costs more in engineering time than it saves in tokens.
Anthropic said Claude Sonnet 5.5 and Haiku 5.5 arrive in the coming weeks. Opus 5.5 is available now at $4 in and $20 out per million tokens.<br/><br/>Here is my workload profile:<br/>- Requests per day: [fill in]<br/>- Average input tokens per request: [fill in]<br/>- Average output tokens per request: [fill in]<br/>- Share of input that is cacheable and reused: [fill in]<br/>- Tasks are: [bounded single-turn / multi-step agentic / long terminal sessions]<br/><br/>Tell me whether migrating to Opus 5.5 now is worth the engineering time, or whether I should wait for Sonnet 5.5.<br/><br/>Give me a number for the monthly saving and a number for the migration cost in engineering hours. If the payback is longer than three months, say so plainly.