Analyzing AI Coding Tools: Dissecting 3,177 API Calls

The recent analysis conducted on four AI coding tools—Claude Code Opus 4.6, Claude Code Sonnet 4.5, Codex GPT-5.3, and Gemini 2.5 Pro—highlights substantial differences in managing API call context windows. Using the Context Lens tracer, the study intercepted 3,177 API calls to evaluate the tools’ efficiency and strategy in handling the context window when tasked with bug fixes in an Express.js environment.
Each coding tool tackled a specific bug—an incorrectly reordered null check in res.send(). Opus, Sonnet, Codex, and Gemini were tasked with identifying and fixing the bug, followed by running the test suite to verify the fix. They all succeeded, albeit with varying approaches and resources.
Claude Code Opus 4.6 consistently used around 23K to 27K tokens, mainly consisting of tool definitions (69% of the context). This indicates a reliance on re-sending these definitions due to the architecture, causing significant caching overhead. Codex (GPT-5.3) presented a wider range from 29.3K to 47.2K tokens, mostly tool results (72%), providing more variability depending on test command specificity. Sonnet, with similar variance, mixed definitions and results more evenly.
Gemini stands out due to its disproportionate use of tokens, peaking at 350.5K, utilizing almost exclusively tool results (96%), exploiting its large 1M context window. Despite a lower cost per token, Gemini’s inconsistent and expansive usage pattern without convergence across runs indicates a unique, albeit less efficient strategy.
These findings illustrate considerable disparities in how AI coding tools manage context windows, impacting both performance and cost efficiency. Developers should weigh token usage strategies when choosing the appropriate tool for their needs, particularly for tasks involving iterative changes or extensive project histories.
📖 Read the full source: HN LLM Tools
👀 See Also

monk: A skill that silences agent narration to save context and tokens
A Reddit user published 'monk', a skill that strips narration, preambles, and postambles from Claude agent responses, claiming ~54% output token reduction per turn and 29-39% context capacity gain at 100 rounds.

DeepMind DiscoRL Meta Learning Update Rule Ported from JAX to PyTorch
A developer has ported DeepMind's DiscoRL meta learning update rule from the 2025 Nature article from JAX to PyTorch. The implementation includes a GitHub repository with a Colab notebook, API, and weights hosted on Hugging Face.

OpenCortex: A Self-Improving Memory System for OpenClaw
OpenCortex replaces OpenClaw's flat MEMORY.md file with structured memory files organized by projects, contacts, workflows, preferences, runbooks, tools, and infrastructure. It includes nightly distillation with principle enforcement audits and weekly synthesis with pattern detection and auto-runbook creation.
UI and Server for Anthropic's Natural Language Autoencoders on llama.cpp
A custom llama.cpp server and Mikupad UI for Anthropic's open-weight Natural Language Autoencoders, supporting activation extraction, explanation, reconstruction, and steering via explanation editing.