A developer has raised a pointed question about Anthropic's latest AI models: are the newest versions actually a step backward for one specific job?
According to a post by Armin Ronacher on his blog "Armin Ronacher's Thoughts and Writings," surfaced by the tech-news aggregator Techmeme, Claude Opus 4.8 and Sonnet 5 appear to be worse at "tool calls" than older Claude models.
Tool calls are how an AI model reaches outside its own text output to do things — running code, querying a database, or calling an external service. For AI that acts as an autonomous "agent," reliable tool calling is foundational. If a model fumbles those calls, the tasks built on top of it can break in subtle ways.
Ronacher writes that the issue caught his attention through what he describes as "a very strange Pi issue" that sent him "down a rabbit hole over the last two days."
His proposed explanation is notable: he suggests the weaker performance likely stems from post-training that assumes Claude Code-like harnesses as the target. In plain terms, post-training is the fine-tuning step that shapes a model after its initial training, and a "harness" is the software scaffolding an AI runs inside. Ronacher's theory is that optimizing the models for one particular environment — Anthropic's own Claude Code tooling — may have made them less dependable when used in other setups.
The claims come from a single developer's investigation, not from Anthropic or independent benchmarks, so they should be read as one practitioner's finding rather than a confirmed regression.
Why it matters: as companies race to build AI agents that take real actions, how a model behaves outside its maker's preferred toolset — not just on official benchmarks — is what will determine whether those agents work in the wild.