Armin Ronacher, creator of Flask and Jinja, has reported that Anthropic's latest AI models, Opus 4.8 and Sonnet 5, exhibit a regression in tool usage, fabricating non-existent parameters in approximately 20% of tool calls during extended coding sessions. This issue was not present in older Anthropic models or OpenAI's Codex models. Ronacher suggests that Anthropic's training environment, which is forgiving of malformed tool calls, may be the root cause, leading the models to invent fields when interacting with stricter schemas. Implementing a 'Strict mode' and removing conversational history significantly reduce these failures. AI
IMPACT Potential issues with tool use in advanced AI models could impact the reliability of AI agents in complex tasks.
RANK_REASON This is a commentary on a reported issue with Anthropic's models, not a direct release or announcement from Anthropic.
Read on dev.to — Anthropic tag →
- Anthropic
- Armin Ronacher
- codex
- Flask
- Haiku
- Jinja Template Engine
- OpenAI
- Opus 4.5
- Opus 4.8
- SENTRY
- Sonnet 5
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
