Anthropic has published a detailed prompting guide for Claude Opus 5.5 that reads, on the surface, as standard developer documentation. Dig into it and you find something more revealing: a catalog of behavioral regressions, harness fragilities, and workarounds that expose how much invisible labor sits between a frontier model and a working product. The headline capability numbers are real. Opus 5.5 generates tokens 30 percent faster than Opus 5 and finishes tasks with fewer tokens. At medium effort it matches or beats Opus 5 at high effort on coding and knowledge work. Visual comprehension improved enough that low-effort Opus 5.5 reads dense charts more accurately than high-effort Opus 5. These are genuine generativity gains — real capability expanding the frontier of what automated systems can do. But the guide's structure tells a different story about resilience. The entire document is organized around failure modes: agents that stop mid-task because they issue progress reports instead of tool calls, thinking behavior that can't be disabled anymore, effort levels that don't map across model versions, prompt caches that invalidate when you change effort settings, and safeguard refusals that block legitimate requests. Each section is a patch for something that broke or changed between versions. Developers who built working systems on Opus 5 face a migration tax. The most telling detail is the thinking-always-on change. Opus 5 let developers disable thinking; Opus 5.5 doesn't. This forces every integration to restructure how it reads responses — checking block types instead of assuming text comes first, removing instructions that substituted for thinking, re-testing mitigations that only existed because thinking was off. It's a unilateral API contract change that shifts engineering cost from Anthropic to every downstream developer. The agentic reliability section is equally revealing. Opus 5.5 voluntarily stops mid-task to report progress, which breaks unattended agent loops. The fix requires developers to build checklist systems, continuation logic, stuck-detection, and sometimes a second smaller model to check whether the primary model actually finished. This is not simplification — it's complexity transfer. The model got smarter but less predictable, and the developer absorbs the difference. Effort calibration tells the same story at the cost level. The guide warns that keeping Opus 5 effort settings will produce longer turns and more output tokens — meaning higher bills — and that effort level names don't correspond to the same amount of thinking across models. Developers must re-benchmark every integration. The maxtokens setting now includes hidden thinking tokens even when thinking content isn't returned, so limits sized for the previous model silently truncate responses. What Anthropic has published is genuinely useful documentation — better than most vendors provide. But read structurally, it's a map of how much friction exists in the AI supply chain between capability improvement and reliable deployment. Every model version change is a tax on the ecosystem, and that tax is paid entirely by developers and their users, not by the model provider.
