A GitHub issue filed by developer stellaraccident (issue #42796) presents an independent analysis of 234,760 Claude Code tool calls, arguing that a sharp reduction in the model's reasoning token allocation has degraded performance on complex engineering tasks. The analysis, drawn from 6,852 session logs and 17,871 thinking blocks collected between late January and April 2026 across four repository projects, is not an official Anthropic report but a community-driven investigation shared as a bug report.
What the Analysis Found
According to the issue, median thinking depth fell roughly 67% between late January and late February 2026, from an estimated ~2,200 characters to ~720 characters. Shortly after, Anthropic introduced thinking content redaction (redact-thinking-2026-02-12), which the poster says went from 0% redaction on March 4 to 100% by March 12 — meaning post-March 12 thinking-length figures are estimates derived from signature field length correlations, not direct token counts.
Other metrics cited in the issue include:
- File reads per edit dropping from 6.6 to 2.0, a 70% decline the poster interprets as reduced pre-edit research.
- Edits made without prior reading rising from 6.2% to 33.7% of modification actions.
- Full-file rewrites more than doubling, from 4.9% to 11.1%, at the expense of surgical edits.
- Reasoning loops (the model reversing a decision five or more times) rising from 8.2 to 26.6 per 1,000 tool calls.
- A custom automated guardrail the poster built catching 173 instances of premature stopping or responsibility-dodging over 17 days, versus zero before March 8.
The Poster's Interpretation
Stellaraccident argues the data shows a mechanism shift: as thinking budgets shrank, Claude Code moved from research-first behavior — grepping repositories, reading headers, checking dependencies — toward immediate edits, which the issue says caused broken function signatures and superficial "simplest fix" patches instead of architectural corrections.
The poster describes running up to 50 concurrent agent sessions on large systems-engineering projects (MLIR compilers, C codebases, GPU drivers) using 5,000+ word CLAUDE.md guideline files, and says these workflows depend on longer reasoning chains to maintain context and verify work.
The issue requests that Anthropic offer guaranteed "max thinking" budget tiers, expose token usage metrics in API responses, and provide regression telemetry — asks that reflect this individual developer's proposed fixes rather than any confirmed Anthropic roadmap.
As of publication, Anthropic has not issued a public response to the issue confirming or disputing the analysis.