/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

Claude Opus 4.8 and Sonnet 5 seem worse at tool calling than older models, likely as post-training optimized them primarily for Claude Code-like environments

A very strange Pi issue sent me down a rabbit hole over the last two days.  The short version is that newer Claude models sometimes call Pi's edit tool …

Armin Ronacher's Thoughts and Writings Armin Ronacher

Context & Ripple Effects

Claude’s recent releases have been positioned around coding, computer use, instruction following and agentic work, building on earlier support for alternating reasoning with tool use. Sonnet 5 was specifically presented as approaching Opus 4.8 performance at lower prices.

This report challenges the portability of those gains: behavior optimized for a Claude Code-like harness may not translate cleanly to Pi’s editing tool or other tool interfaces. It also lands amid user complaints about perceived performance changes in earlier Claude and Claude Code versions.

First-order effects

  • Pi users and developers may see less reliable edit-tool behavior from Claude Opus 4.8 and Sonnet 5 than from older models, requiring model-specific fallbacks or prompt and tool-interface adjustments.
  • Anthropic faces a concrete compatibility signal: models marketed for agentic work can perform unevenly when tool semantics differ from the environment emphasized in post-training.

Second-order effects

  • Tooling vendors will have stronger reason to benchmark models against their own schemas and execution loops rather than treating a model’s coding or agentic score as portable across harnesses.
  • Teams deploying Claude across multiple agents may keep older models in production for particular tool workflows, reducing the simplicity of upgrading to the newest, lower-cost release.

Third-order effects

  • If similar reports recur, agentic-model competition will shift from broad claims of tool-use capability toward compatibility with specific harnesses, tool contracts and evaluation suites.
  • Post-training may increasingly create ecosystem-specific strengths: model providers and agent-platform builders could gain leverage by controlling the environments used to train and measure tool use.

The trend: This is one data point in the shift from general-purpose model releases toward agent systems whose real-world performance depends on the fit between model post-training and the surrounding tool harness.

Discussion

  • @mitsuhiko Armin Ronacher on x
    I had some vibes that Opus 4.8 was performing worse than older ones for some of uses that are off distribution and now I have the receipts. Latest Opus/Sonnet are causing tool invocation failures on Pi's edit tool when older ones did not! I wrote about it. https://lucumr.pocoo.or…
  • @xeophon Florian Brand on x
    Ants models surprisingly performed as well/better in other harnesses on benches, Armins blog hints that this might change in the near future
  • @completeskeptic Diogo Almeida on x
    this is a smell of too much on-policy RL - I was considering a longer write up on this, but it might be too technical for most audiences! this is also why it's been so hard to transfer good engineering practices to LLMs (decomposition, abstraction, state management, etc.)