Demonstrable regression: Claude Code vs Codex 5.5 — same prompt, incomparable results (Feb–Apr 2026)

Context

Jean Duarte — Riaworks Brazil
Senior engineer, 25 years in software. Paying Max 20 subscriber.
A separate billing/refund request has been submitted through official support channels.

The regression

Between February and April 2026, Claude Code regressed massively. Not slightly. Massively. I felt it. Hundreds of developers in my professional network felt it. Then your own team admitted it.

Stella Laurenzo, Senior Director of AI at AMD, published forensic analysis on April 2: 6,852 sessions, 17,871 thinking blocks, 234,760 tool calls. Her conclusion: Claude Code's median visible thinking collapsed 73% — from 2,200 characters in January to 600 in March. Files read before editing dropped from 6.6 to 2.0. API retries per task increased up to 80x.

Source: https://github.com/anthropics/claude-code/issues/46099

Your own engineering post on April 23 admitted three separate engineering missteps caused the degradation. Boris Cherny acknowledged the changes publicly. Thariq Shihipar acknowledged session limit changes during peak hours.

Coverage:

• https://fortune.com/2026/04/24/anthropic-engineering-missteps-claude-code-performance-decline-user-backlash/
• https://www.axios.com/2026/04/16/anthropic-claude-power-user-complaints
• https://venturebeat.com/technology/is-anthropic-nerfing-claude-users-increasingly-report-performance
• https://www.theregister.com/2026/04/13/claude_outage_quality_complaints/

My own evidence — side-by-side comparison

Today I ran an identical multi-file audit prompt on Claude Code and on Codex 5.5. Same project. Same prompt. Same hour.

Claude Code: sample-based. Admitted partial coverage only when pressed. Cited line numbers and file paths it had not actually read. Direct quote from session: "auditoria por amostragem factual, não linha-a-linha. Vou ser honesto."

Codex 5.5: full coverage. 971 files audited. 115,744 lines scanned. 4,737 CSS variable occurrences. 16,793 Tailwind classes captured. Reproducible scanner script. Line-by-line CSVs.

Same prompt. Different planet.

What this cost me

• Subscription fees for Max 20 access during months of degraded service
• Compute tokens burned on retries and partial outputs
• Engineering hours rebuilding work Claude marked "done" without completing
• Project schedule slippage on a Design System refactor I had to roll back entirely
• Cost of switching tools mid-project, evaluating alternatives, retraining workflow

This is not a token issue. This is operational damage caused by a product that regressed without warning, was defended publicly while users complained, and was only acknowledged after a competitor launched and forced a postmortem.

What I expect

1. Transparent changelog of what changed in Claude Code between January and April 2026 — not a blog post, a technical diff
2. Concrete metrics showing current state vs January baseline, so subscribers can make informed renewal decisions
3. Acknowledgment that users who reported degradation during Feb–Apr were correct, not confused

I have logs, screenshots, and the full side-by-side audit outputs. Available on request.

---

Jean Duarte — Riaworks Brazil